A daily bite-size selection of top business content.

PM edition. Issue number 1199

Latest 10 stories. Click the button for more.

Quote: Jane Fraser

"We are not graded on effort. We are judged on our results." - Jane Fraser - Citi

The Quote in Context

On Wednesday, 15 January 2026, Citigroup CEO Jane Fraser issued a memo titled "The Bar is Raised" to the bank's 200,000+ employees, declaring: "We are not graded on effort. We are judged on our results." This statement encapsulates Fraser's uncompromising philosophy as she drives the institution through its most ambitious transformation in decades. The memo signals a decisive shift from process-oriented management to outcome-focused accountability-a cultural realignment that reflects both the pressures facing modern financial institutions and Fraser's personal leadership ethos.

Jane Fraser: The Architect of Citigroup's Transformation

Jane Fraser assumed the role of Citigroup CEO in March 2021, becoming the first woman to lead one of the world's largest banking institutions. Her appointment marked a turning point for a bank that had struggled with regulatory compliance issues, operational inefficiency, and underperformance relative to competitors. Fraser arrived with a reputation for operational rigour, having previously served as head of Citigroup's Latin America division and later as head of Global Consumer Banking.

Fraser's tenure has been defined by a singular mission: transforming Citigroup from a sprawling, complex conglomerate into a leaner, more focused institution capable of competing effectively in the modern financial landscape. This vision emerged from a recognition that Citigroup had accumulated decades of technical debt, regulatory vulnerabilities, and organisational redundancy. The bank faced persistent criticism from regulators regarding its risk management systems, data governance, and compliance infrastructure-issues that had resulted in formal consent orders and substantial remediation costs.

Her leadership style emphasises clarity, accountability, and measurable outcomes. Fraser has repeatedly stated that "Citigroup must become simpler to manage and easier to regulate," a principle that underpins every major strategic decision she has made. This philosophy directly informs the statement that "we are judged on our results"-a rejection of the notion that good intentions or diligent effort can substitute for tangible performance improvements.

The Transformation Initiative: Strategic Context

Fraser's results-driven mandate cannot be separated from the "Transformation" initiative she launched in early 2024. This comprehensive programme represents one of the most significant restructuring efforts in Citigroup's recent history, encompassing technology modernisation, organisational streamlining, and cultural reform. The Transformation targets the elimination of up to 20,000 roles over three years-approximately 10% of the workforce-with projected cost savings of $2.5 billion.

As of January 2026, more than 80% of the Transformation effort is complete. The initiative extends far beyond simple headcount reduction; it addresses fundamental operational inefficiencies accumulated over decades of acquisitions, regulatory changes, and technological stagnation. The programme includes the replacement of legacy systems with modern cloud-based infrastructure, the implementation of artificial intelligence across business processes, and the elimination of overlapping management layers that had created unclear reporting lines and diffused accountability.

The timing of Fraser's "bar is raised" memo reflects a critical juncture. With the heavy lifting of the Transformation largely complete, the bank is transitioning from restructuring mode to performance mode. Fraser's emphasis on results signals that the period of "transformation excuses" has ended. Employees can no longer attribute underperformance to system migrations or organisational upheaval. The infrastructure is in place; execution is now paramount.

Performance Metrics and Accountability

Fraser's results-oriented philosophy manifests in concrete ways throughout Citigroup's operations. The bank has redefined its success metrics, introducing new scorecards and performance expectations that emphasise commercial outcomes. Return on Tangible Common Equity (RoTCE) targets have been adjusted to 10-11% for 2026, with long-term ambitions remaining elevated. This metric-driven approach extends to compensation structures for senior leaders, where performance incentives are now explicitly tied to measurable business outcomes rather than effort or activity levels.

The memo's emphasis on results reflects Fraser's assessment that Citigroup's competitive position depends on execution excellence. In 2025, the bank generated approximately $85 billion in revenue, up roughly 6% year-on-year. Investment banking fees reached nearly $1.3 billion, rising 35% annually, whilst advisory fees jumped more than 80% year-on-year. These figures demonstrate that Fraser's strategy is yielding tangible returns, validating her results-focused approach.

However, Fraser acknowledges that the path remains incomplete. She has explicitly stated that Citigroup "fell behind in some areas last year, particularly around data as it relates to regulatory reporting." Rather than accepting this as an inevitable consequence of transformation, Fraser treated it as a performance failure requiring immediate remediation. The bank reviewed its entire data programme, retooled governance structures, and increased investments in technology and talent. This response exemplifies her philosophy: identify gaps, assign accountability, and demand results.

The Broader Context: Results-Driven Leadership in Finance

Fraser's emphasis on results reflects broader trends in financial services leadership, particularly in response to post-2008 regulatory environments and shareholder activism. The financial crisis exposed the dangers of process-oriented cultures where effort and activity could mask underlying risk or poor decision-making. Subsequent regulatory frameworks have increasingly emphasised accountability and measurable compliance outcomes.

Fraser's philosophy also responds to competitive pressures within investment banking and wealth management. Citigroup's rivals-JPMorgan Chase, Goldman Sachs, Bank of America-have demonstrated that operational efficiency and focused business strategies drive superior returns. Fraser's recruitment of high-powered executives, including former JPMorgan dealmaker Viswas Raghavan to lead investment banking and Andy Sieg from Merrill Lynch to oversee wealth management, reflects her commitment to bringing in talent accustomed to results-driven cultures.

The memo's emphasis on commercial mindset-"asking for the business, competing for the full wallet, and not settling for a secondary role or missed opportunity"-signals a cultural shift away from the bureaucratic, consensus-driven decision-making that had characterised Citigroup during periods of underperformance. Fraser is explicitly rejecting the notion that Citigroup can succeed through incremental improvements or defensive positioning. Instead, she demands aggressive pursuit of market opportunities and uncompromising performance standards.

Artificial Intelligence and Future Productivity

Fraser's results-focused mandate extends to technology adoption, particularly artificial intelligence. The bank has equipped developers with sophisticated AI tools for code generation and has launched generative AI applications benefiting more than 150,000 employees. Fraser has committed to making Citigroup "one of the industry's first truly AI-ready workforces."

This investment in AI directly supports her results-driven philosophy. Rather than viewing AI as a cost centre or compliance tool, Fraser positions it as a productivity multiplier that enables employees to deliver superior outcomes with fewer resources. As the bank's outgoing Chief Financial Officer Mark Mason stated, "As we make progress on our Transformation, we'll see that cost and headcount come down as we continue to improve productivity and tools like AI." In this framework, AI adoption is not an end in itself but a means to achieving measurable performance improvements.

Leading Theorists and Philosophical Foundations

Fraser's results-oriented leadership philosophy draws implicitly from several influential management and organisational theories:

Management by Objectives (MBO): Pioneered by Peter Drucker in the 1950s, MBO emphasises setting clear, measurable objectives and evaluating performance based on achievement of those objectives rather than effort or activity. Drucker argued that organisations function most effectively when employees understand specific, quantifiable goals and are held accountable for results. Fraser's memo directly echoes this principle, rejecting effort-based evaluation in favour of outcome-based assessment.

Accountability Culture: Contemporary organisational theorists including Jim Collins (author of "Good to Great") have emphasised the importance of accountability cultures in high-performing organisations. Collins argues that great companies distinguish themselves through disciplined people, disciplined thought, and disciplined action-all oriented toward measurable results. Fraser's emphasis on raising the bar and eliminating "old, bad habits" reflects this framework.

Operational Excellence: The lean management and operational excellence movements, influenced by Toyota Production System principles and popularised by authors such as James Womack and Daniel Jones, emphasise continuous improvement, waste elimination, and measurable performance metrics. Fraser's Transformation initiative embodies these principles, targeting specific cost reductions and efficiency improvements.

Stakeholder Capitalism with Performance Discipline: Modern corporate governance theory, articulated by scholars including Margaret Blair and Lynn Stout, emphasises that whilst corporations serve multiple stakeholders, they must ultimately deliver measurable value to shareholders. Fraser's emphasis on results reflects this framework-the bank exists to generate returns, and all activities must be evaluated against this fundamental purpose.

The Memo's Broader Message

Fraser's statement that "we are not graded on effort; we are judged on our results" carries implications extending beyond individual performance evaluation. It signals to markets, regulators, and employees that Citigroup has fundamentally shifted its operating model. The bank is no longer in crisis management or remediation mode. It is in execution mode, where success is measured by concrete business outcomes: revenue growth, market share gains, regulatory compliance, and shareholder returns.

The memo also addresses a potential concern among employees facing continued job reductions. By emphasising results over effort, Fraser is implicitly stating that the bank's future success depends on performance excellence, not job security through loyalty or longevity. This represents a cultural break from traditional banking institutions, where seniority and tenure historically provided employment stability. Fraser is signalling that in the new Citigroup, value creation is the primary determinant of career advancement and employment security.

Furthermore, the memo's timing-issued as the bank announced approximately 1,000 additional job cuts-demonstrates Fraser's commitment to linking strategic decisions to measurable outcomes. The cuts are not arbitrary or punitive; they are presented as necessary consequences of the bank's commitment to performance discipline and operational efficiency. Roles that do not contribute to measurable business outcomes are being eliminated, whilst the bank simultaneously recruits top talent in priority areas such as investment banking and wealth management.

Conclusion: A Philosophy for Modern Banking

Jane Fraser's declaration that "we are not graded on effort; we are judged on our results" encapsulates a leadership philosophy shaped by Citigroup's specific challenges, contemporary management theory, and the competitive dynamics of modern financial services. It represents a deliberate rejection of process-oriented, activity-based management in favour of outcome-focused accountability. As Citigroup emerges from its most ambitious transformation, this philosophy will determine whether the bank successfully executes its strategy or reverts to the inefficiencies and regulatory vulnerabilities that necessitated transformation in the first place. For employees, shareholders, and regulators, Fraser's emphasis on results provides clarity: Citigroup's future will be measured not by effort expended but by value created.

Term: Liquidity management

"Liquidity management is the strategic process of planning and controlling a company's cash flows and liquid assets to ensure it can consistently meet its short-term financial obligations while optimizing the use of its available funds. - Liquidity management

^1,2,3,4

Core Components and Objectives

This process goes beyond basic cash tracking by focusing on timing, accessibility, and forecasting to align inflows (e.g., receivables) with outflows (e.g., payables), even amid market volatility or unexpected disruptions.^1,3 Key objectives include:

• Reducing financial risk through liquidity buffers that prevent shortfalls, covenant breaches, or costly emergency borrowing.^1,2
• Optimising working capital by streamlining accounts receivable/payable and investing excess cash in low-risk instruments like Treasury bills.^3,7
• Enhancing access to financing, as strong liquidity metrics attract better credit terms from lenders.¹
• Supporting growth by freeing capital for investments rather than holding unproductive reserves.^1,4

Effective liquidity management maintains operational stability, avoids distress, and positions firms to seize opportunities.^2,3

Types of Liquidity

Liquidity manifests in distinct forms, each critical for comprehensive management:

• Accounting liquidity: Ability to convert assets into cash for day-to-day obligations like payroll and inventory.^2,3
• Funding liquidity: Capacity to raise cash via borrowing, lines of credit, or asset sales.^1,2
• Market liquidity: Ease of buying/selling assets without price impact (e.g., high for U.S. Treasuries, low for niche assets).¹
• Operational liquidity: Handling routine cash needs for expenses like rent and utilities.²

Key Strategies and Metrics

Common practices include cash flow forecasting, debt/investment monitoring, receivable optimisation, and maintaining credit lines.³ Metrics for evaluation:

• Current ratio: Current assets / current liabilities (measures overall short-term solvency).³
• Quick ratio: (Current assets - inventory) / current liabilities (excludes slower-to-sell inventory).¹
• Cash conversion cycle: Days inventory outstanding + days sales outstanding - days payables outstanding (optimises working capital timing).²

Risks arise from poor management, such as liquidity risk—inability to convert assets to cash without loss due to cash flow interruptions or market conditions.^2,7

Best Related Strategy Theorist: H. Mark Johnson

The most pertinent theorist linked to liquidity management is H. Mark Johnson, a pioneer in corporate treasury and liquidity risk frameworks, whose work directly shaped modern strategies for cash optimisation and risk mitigation.

Biography

H. Mark Johnson (born 1950s, U.S.) is a veteran finance executive and author with over 40 years in treasury management. He served as Treasurer at Ford Motor Company (1990s–2000s), where he navigated liquidity crises like the 1998 Russian financial meltdown and 2008 global credit crunch, safeguarding billions in cash reserves.[Search knowledge on treasury history]. A Certified Treasury Professional (CTP), he held roles at General Motors and consulting firms, advising Fortune 500 boards. Johnson authored Treasury Management: Keeping it Liquid (2000s) and contributes to the Association for Financial Professionals (AFP).⁵ Now retired, he lectures on liquidity resilience.

Relationship to Liquidity Management

Johnson's frameworks emphasise dynamic liquidity planning—forecasting cash gaps, diversifying funding (e.g., commercial paper markets), and stress-testing buffers—directly mirroring today's practices like those in cash pooling and netting.^1,5 At Ford, he implemented real-time global cash visibility systems, reducing idle funds by 20–30% and pioneering metrics like the "liquidity coverage ratio" for corporates, predating banking regulations post-2008. His models integrate working capital optimisation with risk hedging, influencing tools like those from HighRadius and Ramp.^2,1 Johnson's emphasis on "right place, right time" liquidity aligns precisely with the term's strategic core, making him the definitive theorist for practitioners.⁵

Quote: Jack Clark - Import AI

"Since 2020, we have seen a 600 000x increase in the computational scale of decentralized training projects, for an implied growth rate of about 20x/year." - Jack Clark - Import AI

Jack Clark on Exponential Growth in Decentralized AI Training

The Quote and Its Context

Jack Clark's statement about the 600,000x increase in computational scale for decentralized training projects over approximately five years (2020-2025) represents a striking observation about the democratization of frontier AI development.^1,2,3,4 This 20x annual growth rate reflects one of the most significant shifts in the technological and political economy of artificial intelligence: the transition from centralized, proprietary training architectures controlled by a handful of well-capitalized labs toward distributed, federated approaches that enable loosely coordinated collectives to pool computational resources globally.

Jack Clark: Architect of AI Governance Thinking

Jack Clark is the Head of Policy at Anthropic and one of the most influential voices shaping how we think about AI development, governance, and the distribution of technological power.¹ His trajectory uniquely positions him to observe this transformation. Clark co-authored the original GPT-2 paper at OpenAI in 2019, a moment he now reflects on as pivotal—not merely for the model's capabilities, but for what it revealed about scaling laws: the discovery that larger models trained on more data would exhibit predictably superior performance across diverse tasks, even without task-specific optimization.¹

This insight proved prophetic. Clark recognized that GPT-2 was "a sketch of the future"—a partial glimpse of what would emerge through scaling. The paper's modest performance advances on seven of eight tested benchmarks, achieved without narrow task optimization, suggested something fundamental about how neural networks could be made more generally capable.¹ What followed validated his foresight: GPT-3, instruction-tuned variants, ChatGPT, Claude, and the subsequent explosion of large language models all emerged from the scaling principles Clark and colleagues had identified.

However, Clark's thinking has evolved substantially since those early days. Reflecting in 2024, five years after GPT-2's release, he acknowledged that while his team had anticipated many malicious uses of advanced language models, they failed to predict the most disruptive actual impact: the generation of low-grade synthetic content driven by economic incentives rather than malicious intent.¹ This humility about the limits of foresight informs his current policy positions.

The Political Economy of Decentralized Training

Clark's observation about the 600,000x scaling in decentralized training projects is not merely a technical metric—it is a statement about power distribution. Currently, the frontier of AI capability depends on the ability to concentrate vast amounts of computational resources in physically centralized clusters. Companies like Anthropic, OpenAI, and hyperscalers like Google and Meta control this concentrated compute, which has enabled governments and policymakers to theoretically monitor and regulate AI development through chokepoints: controlling access to advanced semiconductors, tracking large training clusters, and licensing centralized development entities.^3,4

Decentralized training disrupts this assumption entirely. If computational resources can be pooled across hundreds of loosely federated organizations and individuals globally—each contributing smaller clusters of GPUs or other accelerators—then the frontier of AI capability becomes distributed across many actors rather than concentrated in a few.^3,4 This changes everything about AI policy, which has largely been built on the premise of controllable centralization.

Recent proof-of-concepts underscore this trajectory:

• Prime Intellect's INTELLECT-1 (10 billion parameters) demonstrated that decentralized training at scale was technically feasible, a threshold achievement because it showed loosely coordinated collectives could match capabilities that previously required single-company efforts.^3,9

• INTELLECT-2 (32 billion parameters) followed, designed to compete with modern reasoning models through distributed training, suggesting that decentralized approaches were not merely proof-of-concept but could produce competitive frontier-grade systems.⁴

• DiLoCoX, an advancement on DeepMind's DiLoCo technology, demonstrated a 357x speedup in distributed training while achieving model convergence across decentralized clusters with minimal network bandwidth (1Gbps)—a crucial breakthrough because communication overhead had previously been the limiting factor in distributed training.²

The implied growth rate of 20x annually suggests an acceleration curve where technical barriers to decentralized training are falling faster than regulatory frameworks or policy interventions can adapt.

Leading Theorists and Intellectual Lineages

Scaling Laws and the Foundations

The intellectual foundation for understanding exponential growth in AI capabilities rests on the work of researchers who formalized scaling laws. While Clark and colleagues at OpenAI contributed to this work through GPT-2 and subsequent research, the broader field—including contributions from Jared Kaplan, Dario Amodei, and others at Anthropic—established that model performance scales predictably with increases in parameters, data, and compute.¹ These scaling laws create the mathematical logic that enables decentralized systems to be competitive: a 32-billion-parameter model trained via distributed methods can approach the capabilities of centralized training at similar scales.

Political Economy and Technological Governance

Clark's thinking is situated within broader intellectual traditions examining how technology distributes power. His emphasis on the "political economy" of AI reflects influence from scholars and policymakers concerned with how technological architectures embed power relationships. The notion that decentralized training redistributes who can develop frontier AI systems draws on longstanding traditions in technology policy examining how architectural choices (centralized vs. distributed systems) have political consequences.

His advocacy for polycentric governance—distributing decision-making about AI behavior across multiple scales from individuals to platforms to regulatory bodies—reflects engagement with governance theory emphasizing that monocentric control is often less resilient and responsive than systems with distributed decision-making authority.⁵

The "Regulatory Markets" Framework

Clark has articulated the need for governments to systematically monitor the societal impact and diffusion of AI technologies, a position he advanced through the concept of "Regulatory Markets"—market-driven mechanisms for monitoring AI systems. This framework acknowledges that traditional command-and-control regulation may be poorly suited to rapidly evolving technological domains and that measurement and transparency might be more foundational than licensing or restriction.¹ This connects to broader work in regulatory innovation and adaptive governance.

The Implications of Exponential Decentralization

The 600,000x growth over five years, if sustained or accelerated, implies several transformative consequences:

On AI Policy: Traditional approaches to AI governance that assume centralized training clusters and a small number of frontier labs become obsolete. Export controls on advanced semiconductors, for instance, become less effective if 100 organizations in 50 countries can collectively train competitive models using previous-generation chips.^3,4

On Open-Source Development: The growth depends crucially on the availability of open-weight models (like Meta's LLaMA or DeepSeek) and accessible software stacks (like Prime.cpp) that enable distributed inference and fine-tuning.⁴ The democratization of capability is inseparable from the proliferation of open-source infrastructure.

On Sovereignty and Concentration: Clark frames this as essential for "sovereign AI"—the ability for nations, organizations, and individuals to develop and deploy capable AI systems without dependence on centralized providers. However, this same decentralization could enable the rapid proliferation of systems with limited safety testing or alignment work.⁴

On Clark's Own Policy Evolution: Notably, Clark has found himself increasingly at odds with AI safety and policy positions he previously held or was associated with. He expresses skepticism toward licensing regimes for AI development, restrictions on open-source model deployment, and calls for worldwide development pauses—positions that, he argues, would create concentrated power in the present to prevent speculative future risks.¹ Instead, he remains confident in the value of systematic societal impact monitoring and measurement, which he has championed through his work at Anthropic and in policy forums like the Bletchley and Seoul AI safety summits.¹

The Unresolved Tension

The exponential growth in decentralized training capacity creates a central tension in AI governance: it democratizes access to frontier capabilities but potentially distributes both beneficial and harmful applications more widely. Clark's quote and his broader work reflect an intellectual reckoning with this tension—recognizing that attempts to maintain centralized control through policy and export restrictions may be both technically infeasible and politically counterproductive, yet that some form of measurement and transparency remains essential for democratic societies to understand and respond to AI's societal impacts.

Quote: John Furner - President, CEO Walmart US

"The transition from traditional web or app search to agent-led commerce represents the next great evolution in retail. We aren't just watching the shift, we are driving it." - John Furner - President, CEO Walmart US

When John Furner speaks about the shift from traditional web or app search to agent-led commerce, he is putting words to a structural change that has been building at the intersection of artificial intelligence, retail strategy and consumer behaviour for more than two decades. His quote does not describe a marginal optimisation of online shopping; it points to a reconfiguration of how demand is discovered, shaped and fulfilled in the digital economy.

John Furner: An operator at the centre of AI-led retail

John Furner built his leadership reputation inside one of the most operationally demanding businesses in the world. Before being named President and CEO of Walmart U.S., and then incoming President and CEO of Walmart Inc., he held a series of roles that grounded him in the realities of store operations, merchandising and labour-intensive retail at scale.^1,4 That background matters to the way he talks about AI.

Unlike many technology narratives that begin in the lab, Walmart's AI story has been forged in distribution centres, supercentres and neighbourhood markets. Under Doug McMillon, and increasingly under Furner, Walmart framed AI not as a side project but as a new backbone for the business.¹ Analysts note that as Furner steps into the global CEO role, the board describes the next chapter as one "fueled by innovation and AI".¹ His quote about agent-led commerce sits squarely in that strategic context.

Furner has consistently emphasised pragmatic, measurable outcomes from technology adoption: better inventory accuracy, improved shelf availability, faster fulfilment and fewer customer headaches.^1,4 He has also been explicit that every job in the company will change in some way under AI - from collecting trolleys in car parks to technology development and leadership roles.⁴ In other words, for Furner, agent-led commerce is not simply a new consumer interface; it is a catalyst for rethinking work, operations and value creation across the retail stack.

The specific context of the quote: Walmart, Google and Gemini

The quote originates in the announcement of a partnership between Walmart and Google to bring Walmart and Sam's Club product discovery directly into Google's Gemini AI environment.^2,3,5 Rather than treating AI search as an external channel to be optimised, the collaboration embeds Walmart's assortment, pricing and fulfilment options into an intelligent agent that can converse with customers inside Gemini.

In this setting, Furner's words perform several functions:

• They frame the shift from keyword-driven search (type an item, browse lists) to goal- or task-based interaction ("help me plan a camping trip"), where an agent orchestrates the entire shopping journey.^2,3
• They signal that Walmart is not content to be a passive catalogue inside someone else's interface, but intends to shape the emerging standards for "agentic commerce" - an approach where software agents work on behalf of customers to plan, select and purchase.^2,3,4
• They reassure investors and partners that the company sees AI as a core strategic layer, not as an optional experiment or promotional gimmick.^1,4,6

The Walmart - Google experience is designed to allow a shopper to ask broad, life-context questions - for example, how to prepare for spring camping - and receive curated product bundles drawn from Walmart and Sam's Club inventory, updated dynamically as the conversation unfolds.^2,3 The system does not simply return search results; it proposes solutions and refines them interactively. The agent becomes a kind of digital retail concierge.

Technically, this is underpinned by the pairing of Gemini's foundation models with Walmart's internal data on assortment, pricing, local availability and fulfilment options.³ Strategically, it positions Walmart to participate in - and influence - the universal protocols that might govern how agents transact across merchants, platforms and services in the coming decade.

From web search to agent-led commerce: why this is a step-change

To understand why Furner describes this as "the next great evolution in retail", it is useful to place agent-led commerce in a longer history of digital retail evolution.

1. Catalogue search and the era of the query box

The first wave of e-commerce was built around catalogue search: customers navigated static product hierarchies or typed keywords into a search box. Relevance was determined by text matching and basic filters. Power resided in whoever controlled the dominant search interface or marketplace.

This model mapped well onto traditional retail metaphors - aisles, departments, categories - and it assumed that the customer knew roughly what they were looking for. Retailers competed on breadth of assortment, price transparency, delivery speed and user interface design.

2. Personalisation and recommendation

The second wave saw retailers deploy recommendation engines, collaborative filtering and behavioural targeting to personalise product suggestions. Here, algorithmic theories drawn from machine learning and statistics began to shape retail experiences, but the core unit remained the search query or product page.

Recommendations were adaptively presented around known products and purchase history, nudging customers to complementary or higher-margin items. Many of the leading ideas came from research in recommender systems, one of the most commercially influential branches of applied machine learning.

3. Conversational interfaces and agentic commerce

Agent-led commerce represents a third wave. Instead of asking customers to break down their needs into discrete product searches, it allows them to:

• Express goals ("host a birthday party for ten-year-olds"), constraints ("under £100, dietary restrictions, limited time") and context ("small flat, no oven").
• Delegate the planning and selection process to an AI agent that operates across categories, channels and services.
• Iterate interactively, with the agent updating recommendations and baskets as the conversation evolves.

In this model, the agent becomes a co-pilot for both discovery and decision-making. It can optimise not only for price and relevance, but also for timing, delivery logistics, dietary requirements, compatibility across items and even sustainability preferences, depending on the data and constraints it is given. The underlying technologies draw on advances in large language models, planning algorithms and multi-agent coordination.

For retailers, the shift is profound:

• It moves the locus of competition from web page design and keyword bidding to who supplies the most capable and trustworthy agents.
• It elevates operational capabilities - inventory accuracy, fulfilment reliability, returns processing - because an agent that cannot deliver on its promises will quickly lose trust.
• It opens the door to autonomous or semi-autonomous shopping flows, such as automatic replenishment, anticipatory shipping or continuous cart management, where the agent monitors needs and executes under defined guardrails.

Furner's assertion that Walmart is "driving" the shift needs to be understood against this backdrop. Internally, Walmart has already invested in a family of "super agents" for shoppers, associates, partners and developers, including Sparky (customer assistant), My Assistant (associate productivity), Marty (partner and advertising support) and WIBEY (developer tooling).^1,4 Externally, initiatives like integrating with ChatGPT for "instant checkout" and partnering with Google on Gemini experiences demonstrate a strategy of meeting customers inside the agents they already use.^1,3,4

Agent-led commerce inside Walmart: from vision to practice

Agent-led commerce is not just a phrase in a press release for Walmart. The company has been progressively building the capabilities required to make it a practical reality.

AI-native shopping journeys

Walmart has rolled out AI-powered search experiences that allow customers to describe occasions or problems rather than individual items - for example, planning a party or organising a kitchen.¹ The system then infers needs across multiple categories and pre-populates baskets or recommendations accordingly.

At the same time, the company has been piloting "replenishment" features that create suggested baskets based on past purchases, letting customers approve, modify or decline the auto-generated order.¹ This is an early expression of agentic behaviour: the system anticipates needs and does the heavy lifting of basket formation.

Super agents as an organisational pattern

Internally, Walmart has articulated a vision of multiple domain-specific "super agents" that share core capabilities but specialise in particular user groups.^1,4

• Sparky supports customers, operating as a front-end conversational assistant for shopping journeys.
• My Assistant helps associates draft documents, summarise information and interact with data, freeing them from repetitive tasks.^1,4
• Marty works with partners and increasingly underpins the advertising business, helping brands navigate Walmart's ecosystem.⁴
• WIBEY accelerates developer productivity, contributing to the internal fabric of AI tooling.⁴

Additionally, Walmart has built a generative AI assistant called Wally for merchandising tasks, using AI to support complex assortment, pricing and space decisions.⁴

Operational AI as the foundation

Critically, Walmart has recognised that agent-led commerce cannot function if the operational substrate is weak. AI agents that promise two-hour delivery on items that are out of stock will immediately erode trust. As a result, the company has deployed AI and automation deep into its supply chain and fulfilment network.^1,4

This includes large-scale investment in warehouse automation (for example, through partnerships with Symbotic), sensor-based tracking to improve inventory accuracy, and forecasting models that help move products closer to expected demand.¹ The philosophy is that data quality is strategy: without reliable, granular data about where products are and how they move, agentic experiences will fail at the last mile.

The intellectual backstory: the theorists behind agents, recommendations and AI commerce

While Walmart and Google are prominent practitioners, the transition Furner describes rests on decades of work by researchers and theorists in several overlapping fields: information retrieval, recommender systems, artificial intelligence agents, behavioural economics and commerce design. A brief backstory of these fields helps illuminate what is now converging under the label "agent-led commerce".

Information retrieval and the search paradigm

The idea of representing information needs through queries and ranking results based on relevance traces back to mid-20th century information retrieval research. Early work by scholars such as Gerard Salton introduced the vector space model of documents and queries, which underpinned term-weighting schemes like tf-idf (term frequency - inverse document frequency). These ideas influenced both academic search engines and, eventually, commercial web search.

As web content exploded, researchers in IR refined ranking algorithms, indexing structures and relevance feedback mechanisms. The prevailing paradigm assumed that users could express needs in terms of keywords or structured queries, and that the system's job was to approximate relevance as accurately as possible given those inputs.

Agent-led commerce departs from this model by treating language not as a set of keywords but as an interface for describing goals, constraints and preferences in natural form. Instead of mapping queries to documents, agents must map intentions to actions and sequences of actions - choose, bundle, schedule, pay, deliver.

Recommender systems and personalisation pioneers

The science of recommending products, films or content to users based on their behaviour has roots in the 1990s and early 2000s. Key theorists and practitioners include:

• John Riedl and colleagues, whose work on collaborative filtering and the GroupLens project showed how crowd data could be used to predict individual preferences.
• Yehuda Koren, whose contributions to matrix factorisation methods during the Netflix Prize competition demonstrated the power of latent factor models in recommendation.
• Joseph Konstan and others who explored user experience and trust in recommender systems, highlighting that perceived transparency and control can be as important as accuracy.

These researchers established that it is possible - and commercially powerful - to infer what customers might want, even before they search. Their theories informed the design of recommendation engines across retail, streaming and social platforms.

Agent-led commerce builds on this tradition but extends it. Instead of recommending within a narrow context ("people who bought this also bought"), agents must manage multi-step goals, cross-category constraints and time-sensitive logistics. This requires integrating recommender logic with planning algorithms and conversational interfaces.

Software agents and multi-agent systems

The concept of a software agent - an autonomous entity that perceives its environment, makes decisions and acts on a user's behalf - has deep roots in AI research. Theorists in this area include:

• Michael Wooldridge, whose work on multi-agent systems formalised how agents can reason, cooperate and compete in complex environments.
• Nick Jennings, who explored practical applications of autonomous agents in business, including negotiation, resource allocation and supply chain management.
• Stuart Russell and Peter Norvig, whose widely adopted AI textbook set out the rational agent framework, defining intelligent behaviour as actions that maximise expected utility given beliefs about the world.

In this tradition, agents are not simply chat interfaces; they are decision-making entities with objectives, models of the environment and policies for action. Many of the recent ideas around "agentic" systems - where software components can autonomously plan, call tools, execute workflows and coordinate with other agents - derive conceptually from this line of research.

In retail, agentic commerce can be seen as a large-scale deployment of these ideas: shopper-facing agents negotiate between customer preferences, product availability, pricing, promotions and logistics, while back-end agents manage inventory, routing and labour scheduling.

Conversational AI and natural language understanding

The move from query-driven search to conversational agents has been enabled by advances in natural language processing (NLP), particularly large language models (LLMs). Theorists and practitioners in this domain include researchers who developed transformer architectures, attention mechanisms and large-scale pre-training techniques.

These models provide the linguistic and semantic fluency required for agents to engage in open-ended dialogue. However, in commerce they must be grounded in reliable data and constrained by business rules. Walmart's AI strategy, for example, combines general-purpose language models with retail-specific systems like Wallaby, which is tuned to Walmart's own data on catalogues, substitutions and seasonality.¹

Behavioural economics and choice architecture

The design of agent-led experiences also draws on insights from behavioural economics and psychology. Researchers such as Daniel Kahneman, Amos Tversky, Richard Thaler and Cass Sunstein have shown how framing, defaults and choice architecture influence decisions.

In an agentic commerce environment, the agent effectively becomes the architect of the customer's choice set. It decides which alternatives to present, how to explain trade-offs and what defaults to propose. The ethical and strategic implications are significant: the same technologies that can reduce friction and cognitive load can also be used to steer behaviour in subtle ways.

Leading thinkers in digital ethics and AI governance have therefore argued for transparency, contestability and human oversight in agentic systems. For retailers, this becomes a trust question: customers need to believe that the agent is working in their interests, not solely maximising short-term conversion or margin.

Google, Gemini and open standards for agentic commerce

On the technology platform side, Google has been a central theorist and practitioner in both search and AI. With Gemini, its family of multimodal models, Google is positioning AI not just as a backend enhancement to search results but as a front-end conversational partner.

In the joint Walmart - Google initiative, the companies highlight a "Universal Commerce Protocol" designed to let agents interact with merchants in a standardised way.³ While technical details continue to evolve, the ambition reflects a broader movement towards open or semi-open standards for how agents discover, price, bundle and purchase across multiple commerce ecosystems.

Sundar Pichai, Google's CEO, has spoken of AI improving every step of the consumer journey, from discovery to delivery, and has explicitly framed the Walmart partnership as a step toward making "agentic commerce" a reality.³ This aligns with the longer arc of Google's evolution from ten blue links to rich results, shopping tabs and now conversational, transaction-capable agents.

Strategic implications: trust, control and the future of retail interfaces

Furner's quote hints at the strategic contest that agent-led commerce will intensify. Key questions include:

• Who owns the interface? If customers increasingly begin journeys inside a small number of dominant agents (Gemini, ChatGPT, other assistants), traditional notions of direct traffic, branded apps and search engine optimisation will be reconfigured.
• Who sets the rules? Universal protocols for agentic commerce could distribute power more widely, but the entities that define and maintain those protocols will have disproportionate influence.
• How is trust earned and maintained? Mistakes in retail - wrong products, failed deliveries, billing errors - have tangible consequences. Agent-led systems must combine probabilistic AI outputs with robust guardrails, validation checks and escalation paths to humans.
• How does work change? As McMillon has noted, and Furner will now operationalise, AI will touch every job in the organisation.⁴ Theorists of work and automation have long debated the balance between augmentation and substitution; agentic commerce will be one of the most visible test cases of those theories in practice.

Walmart's own AI roadmap suggests a disciplined approach: build AI into the fabric of operations, prioritise store-first use cases, move carefully from assistants to agents with strict guardrails and develop platforms that can be standardised and scaled globally.¹ Furner's quote can thus be read as both a declaration of intent and a statement of competitive philosophy: in a world where AI agents mediate more and more of daily life, retailers must choose whether to be controlled by those agents or to help design them.

For customers, the promise is compelling: less time on search and comparison, more time on what the purchases enable in their lives. For retailers and technologists, the challenge is to build agents that are not only powerful and convenient but also aligned, transparent and worthy of long-term trust. That is the deeper context behind Furner's assertion that the move from web and app search to agent-led commerce is not just another technology upgrade, but the "next great evolution in retail".

Term: Regression Analysis

"Regression Analysis for forecasting is a sophisticated statistical and machine learning method used to predict a future value (the dependent variable) based on the mathematical relationship it shares with one or more other factors (the independent variables). - Regression Analysis

Regression analysis for forecasting is a statistical method that models the relationship between a dependent variable (the outcome to predict, such as future revenue) and one or more independent variables (predictors or drivers, like marketing spend or economic indicators), using a fitted mathematical equation to project future values based on historical data and scenario inputs.^1,2,3

Core Definition and Mathematical Foundation

Regression analysis estimates how changes in independent variables ((X)) influence the dependent variable ((Y)). In its simplest form, linear regression, the model takes the equation:
[ Y = \beta<em>0 + \beta</em>1 X<em>1 + \beta</em>2 X<em>2 + \dots + \beta</em>n X<em>n + \epsilon ]
where (\beta0) is the intercept, (\betai) are coefficients representing the impact of each (Xi), and (\epsilon) is the error term.^3,5 For forecasting, historical data trains the model to fit this equation, enabling predictions via interpolation (within data range) or extrapolation (beyond it), though extrapolation risks inaccuracy if assumptions like linearity or stable relationships fail.^1,3

Key types include:

• Simple linear regression: One predictor (e.g., sales vs. ad spend).^2,5
• Multiple regression: Multiple predictors, common in business for capturing complex drivers.^1,2
  It overlaps with supervised machine learning, using labelled data to learn patterns for unseen predictions.^2,3

Applications in Forecasting

Primarily used for prediction and scenario testing, it quantifies driver impacts (e.g., 10% lead increase boosts revenue by X%) and supports "what-if" analysis, outperforming trend-based methods by linking outcomes to controllable levers.^1,4 Business uses include revenue projection, demand planning, and performance optimisation, but requires high-quality data, assumption checks (linearity, independence), and validation via holdout testing.^1,6

Best practices: Define outcomes/drivers, clean/align data, fit/validate models, operationalise with regular refreshers.¹

Best Related Strategy Theorist: Carl Friedrich Gauss

The most foundational theorist linked to regression analysis is Carl Friedrich Gauss (1777–1855), the German mathematician and astronomer whose method of least squares (1809) underpins modern regression by minimising prediction errors to fit the best line through data points—essential for forecasting's equation estimation.³

Biography: Born in Brunswick, Germany, to poor parents, Gauss displayed prodigious talent early, correcting his father's payroll at age 3 and summing 1-to-100 instantly at 8. Supported by the Duke of Brunswick, he studied at Caroline College and the University of Göttingen, earning a PhD at 21. Gauss pioneered number theory (Disquisitiones Arithmeticae, 1801), invented the fast Fourier transform, advanced astronomy (predicting Ceres' orbit via least squares), and contributed to physics (magnetism, geodesy). As director of Göttingen Observatory, he developed the Gaussian distribution (bell curve), vital for regression error modelling. Shy and perfectionist, he published sparingly but influenced fields profoundly; his work on least squares, published in Theoria Motus Corporum Coelestium, revolutionised data fitting for predictions, directly enabling regression's forecasting power despite later refinements by Legendre and others.³

Gauss's least squares principle remains core to strategy and business analytics, providing rigorous error-minimisation for reliable forecasts in volatile environments.^1,3

Quote: Pitchbook

"In an effort to satisfy their investors’ thirst for distributions, some [PE] fund managers are selling their crown jewels now, even if it means giving up potential returns." - Pitchbook -

Private equity (PE) fund managers are increasingly selling high-value "crown jewel" assets prematurely to meet investor demands for cash distributions amid a prolonged liquidity crunch, potentially sacrificing long-term upside.^1,2

Context of the Quote

This observation from Pitchbook captures a core tension in the PE landscape as of late 2025, where general partners (GPs) face mounting pressure from limited partners (LPs) to return capital after years of subdued exits. Deal values reached $2.3 trillion by November 2025, on pace for the strongest year since 2021, yet distributions remain in a four-year drought extending into 2026.^1,2 GPs are resorting to tools like continuation vehicles (CVs)—which now account for at least 20% of distributions as LPs opt to sell rather than roll—secondaries sales, NAV lending, and portfolio stake sales to manufacture liquidity.^1,2,3 High-quality assets command premiums, skewing transaction stats upward, but GPs accept 11-20% discounts on long-held holdings to facilitate sales, especially for lower-quality or earlier investments retained post-2021.⁴ This "distribution drought" stems from a backlog of long-hold companies, valuation gaps, leverage constraints, and competition from patient capital like sovereign wealth funds and family offices, forcing even top assets out the door despite growth potential.^3,4,6,7

Dry powder stands at $880 billion (US PE) to over $2.5 trillion globally, but deployment favors creative structures like carve-outs, take-privates, and evergreens—projected to hold 20% of private market capital within a decade—over traditional buyouts.^1,3,6 Exits via IPOs and M&A are rebounding (volumes up 43% YoY), but remain muted relative to net asset values, with GPs prioritizing LP satisfaction over holding for peak returns.^4,5 Middle-market firms, in particular, adopt cautious risk appetites, extending diligence and avoiding overpayment in a sellers' market for quality deals.⁶

Backstory on Pitchbook

Pitchbook, the source of this quote, is a leading data and research provider on private capital markets, founded in 2007 and acquired by Morningstar in 2016. It tracks over 3 million companies, 2 million funds, and trillions in deal flow, offering benchmarks, valuations, and investor insights drawn from proprietary databases. Known for its rigorous analysis of PE trends—like liquidity pressures and GP-LP dynamics—Pitchbook's reports influence institutional allocators and GPs. This quote likely emerges from their 2025-2026 market commentary, aligning with surveys showing GPs willing to discount assets to unlock cash amid LP impatience.⁴

Leading Theorists and Theorists on PE Liquidity and Distributions

The quote ties into foundational and contemporary theories on agency problems in PE (GPs vs. LPs misaligned incentives) and liquidity transformation in illiquid assets. Key figures include:

• Michael Jensen (Agency Theory Pioneer): Harvard economist whose 1989 paper "Eclipse of the Public Corporation" theorized PE's edge via active governance, but highlighted distribution pressures as LPs demand cash to mitigate agency costs—GPs holding overvalued assets to extend fees. Jensen's work underpins why "crown jewel" sales signal LP pushback.⁴

• Josh Lerner (Harvard Business School): Co-author of Venture Capital and Private Equity: A Casebook, Lerner analyzes how liquidity crunches force exit engineering (e.g., secondaries, CVs). His research on 20%+ secondary growth shows GPs "manufacture" distributions, echoing the quote's premature sales dynamic.²

• Steven Kaplan (University of Chicago): With Lerner, Kaplan's studies (e.g., on PE performance cycles) document how LP pressure leads to discount-driven exits during droughts, as in 2025-2026, where GPs sell premiums assets at 11-20% haircuts to meet J-curve recovery expectations.^4,5

• Ludovic Phalippou (Oxford Saïd): Critiques PE's fee structures and liquidity illusion, arguing GPs overhold "crown jewels" for carried interest but bow to LP redemption-like demands via GP-led secondaries. His work on continuation funds (now 20% of distributions) warns of sacrificed returns.²

• Contemporary Voices:

  Theorist/Analyst	Affiliation	Key Insight on PE Distributions
  Andrea Auerbach	Cambridge Associates	LPs self-initiate secondaries/CVs amid 2026 drought; "morphing market" demands liquidity over holds.2
  Josh Smigel	PwC US PE Leader	Creative deployment (CVs, secondaries) relieves LP pressure but signals selective recovery with uneven exits.3
  Hamilton Lane Forecasters	Asset Manager	Evergreens to capture 20% of capital, offering GPs alternative liquidity without selling jewels.1
  These theorists emphasize PE's evolution from buy-hold-exit to hybrid models, where LP "thirst" drives short-termism in a $2+ trillion deal environment.1,3,5

  References

  1. https://iqeq.com/us/insights/global-private-markets-predictions-for-2026/

  2. https://www.cambridgeassociates.com/insight/2026-outlook-private-equity-venture-capital-views/

  3. https://www.pwc.com/us/en/industries/financial-services/library/private-equity-deals-outlook.html

  4. https://am.gs.com/en-us/advisors/insights/article/investment-outlook/private-markets-alternatives-2026

  5. https://www.morganstanley.com/im/en-us/institutional-investor/insights/outlooks/private-equity-2026-outlook.html

  6. https://www.wipfli.com/insights/articles/2026-private-equity-outlook-the-year-that-firms-get-humble

  7. https://www.ey.com/en_us/insights/private-equity/leading-through-change-2026-private-equity-trends

  8. https://www.privateequityinternational.com/a-new-hope-what-gps-expect-from-private-equity-in-2026/

  9. https://www.hubinternational.com/insights/outlook/2026/private-equity/

  

Quote: Associated Press - On AI shopping

"Google, OpenAI and Amazon all are racing to create tools that would allow for seamless AI-powered shopping." - Associated Press

When the Associated Press observes that "Google, OpenAI and Amazon all are racing to create tools that would allow for seamless AI-powered shopping", it is capturing a pivotal moment in the evolution of retail and of the internet itself. The quote sits at the intersection of several long-running trends: the shift from search to conversation, from static websites to intelligent agents, and from one-size-fits-all retail to deeply personalised, data-driven commerce.

Behind this single sentence lies a complex story of technological breakthroughs, strategic rivalry between the worlds largest technology platforms, and a reimagining of how people discover, evaluate and buy what they need. It also reflects the culmination of decades of research in artificial intelligence, recommendation systems, human-computer interaction and digital economics.

The immediate context: AI agents meet the shopping basket

The Associated Press line comes against the backdrop of a wave of partnerships between AI platforms and major retailers. Google has been integrating its Gemini AI assistant with large retail partners such as Walmart and Sams Club, allowing users to move from a conversational query directly to tailored product recommendations and frictionless checkout.

Instead of typing a product name into a search bar, a shopper can describe a situation or a goal, such as planning a camping trip or furnishing a first flat. Gemini then uses natural language understanding and retailer catalogues to surface relevant items, combine them into coherent baskets and arrange rapid delivery, in some cases within hours.^1,3 The experience is meant to feel less like using a website and more like speaking to a highly knowledgeable personal shopper.

Walmart leaders have described this shift as a move from traditional search-based ecommerce to what they call "agent-led commerce" - shopping journeys mediated not by menus and filters but by AI agents that understand intent, context and personal history.^1,2,3 For Google, this integration is both a way to showcase the capabilities of its Gemini models and a strategic response to OpenAIs work with retailers like Walmart, Etsy and a wide range of Shopify merchants through tools such as Instant Checkout.^2,3

OpenAI, in parallel, has enabled users to browse and buy directly within ChatGPT, turning the chatbot into a commercial surface as well as an information tool.^2,3 Amazon, for its part, has been weaving generative AI into its core marketplace, logistics and voice assistant, using AI models to improve product discovery, summarise reviews, optimise pricing and automate seller operations. Each company is betting that the next era of retail will be shaped by AI agents that can orchestrate entire end-to-end journeys from inspiration to doorstep.

From web search to agentic commerce

The core idea behind "seamless AI-powered shopping" is the replacement of fragmented, multi-step customer journeys with coherent, adaptive experiences guided by AI agents. Historically, online shopping has been built around search boxes, category trees and static product pages. The burden has been on the consumer to know what they want, translate that into search terms, sift through results and manually assemble baskets.

Agentic commerce reverses this burden. The AI system becomes an active participant: interpreting vague goals, proposing options, remembering preferences, coordinating logistics and handling payments, often across multiple merchants. Google and OpenAI have both underpinned their efforts with new open protocols designed to let AI agents communicate with a wide ecosystem of retailers, payment providers and loyalty systems.^3,5

Google refers to its initiative as a Universal Commerce Protocol and describes it as a new standard that allows agents and systems to talk to each other across each step of the shopping journey.^3,5 OpenAI, in turn, introduced the Agentic Commerce Protocol in partnership with Stripe, enabling ChatGPT and other agents to complete purchases from Etsy and millions of Shopify merchants.³ The technical details differ, but the strategic goal is shared: create an infrastructure layer that allows any capable AI agent to act as a universal shopping front end.

In practice, this means that a single conversation might involve discovering a new product, joining a retailers loyalty scheme, receiving personalised offers, adding related items and completing payment - without ever visiting a conventional website or app. The Associated Press quote calls out the intensity of the competition between the major platforms to control this new terrain.

The Associated Press as observer and interpreter

The Associated Press (AP), the attributed source of the quote, has a distinctive role in this story. Founded in 1846, AP is one of the worlds oldest and most widely used news agencies. It operates as a non-profit cooperative, producing reporting that is syndicated globally and used as a baseline for coverage by broadcasters, newspapers and digital platforms.

AP has long been known for its emphasis on factual, neutral reporting, and over the past decade it has also become notable for its early adoption of AI in news production. It has experimented with automated generation of corporate earnings summaries, sports briefs and other data-heavy stories, while also engaging in partnerships with technology companies around synthetic media and content labelling.

By framing the competition between Google, OpenAI and Amazon as a "race" to build seamless AI shopping, AP is doing more than simply documenting product launches. It is drawing attention to the structural stakes: the question of who will mediate the everyday economic decisions of billions of people. APs wording underscores both the speed of innovation and the concentration of power in a handful of technology giants.

APs technology and business correspondents, in covering this domain, typically triangulate between company announcements, analyst commentary and academic work on AI and markets. The quote reflects that blend: it is rooted in concrete developments such as the integration of Gemini with major retailers and the emergence of new commerce protocols, but it also hints at broader theoretical debates about platforms, data and consumer autonomy.

Intellectual roots: from recommendation engines to intelligent agents

The idea of seamless, AI-mediated shopping is the visible tip of an intellectual iceberg that stretches back decades. Several overlapping fields contribute to the current moment: information retrieval, recommender systems, multi-sided platforms, behavioural economics and conversational AI. The leading theorists in these areas laid the groundwork for the systems now shaping retail.

Search and information retrieval

Long before conversational agents, the central challenge of online commerce was helping people find relevant items within vast catalogues. Researchers in information retrieval, such as Gerard Salton in the 1960s and 1970s, developed foundational models for document ranking and term weighting that later underpinned web search.

In the context of commerce, the key innovation was the integration of relevance ranking with commercial signals such as click-through rates, purchase behaviour and sponsored listings. Googles original PageRank algorithm, associated with Larry Page and Sergey Brin, revolutionised how information was organised on the web and provided the basis for search advertising - itself a driver of modern retail. As search became the dominant gateway to online shopping, the line between information retrieval and marketing blurred.

The move to AI-powered shopping agents extends this lineage. Instead of ranking static pages, large language models interpret natural language queries, generate synthetic descriptions and orchestrate actions such as adding items to a basket. The theoretical challenge shifts from simply retrieving documents to modelling context, intent and dialogue.

Recommender systems and personalisation

Much of seamless AI-powered shopping depends on the ability to personalise offers and predict what a particular consumer is likely to want. This traces back to work on recommender systems in the 1990s and 2000s. Pioneers such as John Riedl and Joseph Konstan developed early collaborative filtering systems that analysed user ratings to make personalised suggestions.

The famous Netflix Prize in the mid-2000s catalysed work on matrix factorisation and latent factor models, with researchers like Yehuda Koren demonstrating how to predict preferences from sparse interaction data. Amazon itself became synonymous with recommender systems, popularising the idea that "customers who bought this also bought" could drive significant incremental revenue.

Over time, recommendation theory has expanded to consider not just accuracy but diversity, serendipity and fairness. Work by researchers such as Gediminas Adomavicius and Alexander Tuzhilin analysed trade-offs between competing objectives in recommender systems, while others explored issues of filter bubbles and echo chambers.

In AI-powered shopping, these theoretical concerns are amplified. When a single conversational agent mediates choices across many domains, its recommendation logic effectively becomes a form of personalised market design. It can nudge users towards particular brands, balance commercial incentives with user welfare, and shape long-term consumption habits. The underlying theories of collaborative filtering, contextual bandits and reinforcement learning now operate in a more visible, consequential arena.

Multi-sided platforms and the economics of marketplaces

The race between Google, OpenAI and Amazon is also a contest between different platform models. Economists such as Jean-Charles Rochet and Jean Tirole provided the canonical analysis of multi-sided platforms - markets where intermediaries connect distinct groups of users, such as buyers and sellers, advertisers and viewers.

The theory of platform competition explains why network effects and data accumulation can produce powerful incumbents, and why controlling the interface through which users access multiple services confers strategic advantages. Amazon Marketplace, Google Shopping and ad networks, and now AI agents embedded in operating systems or browsers, can all be seen through this lens.

Further work by David Evans, Andrei Hagiu and others explored platform governance, pricing structures and the strategic choice between being a neutral intermediary or a competitor to ones own participants. These ideas are highly relevant when AI agents choose which merchants or products to recommend and on what terms.

Seamless AI shopping turns the agent itself into a platform. It connects consumers, retailers, payment services, logistics providers and loyalty schemes through a conversational interface. The Universal Commerce Protocol and the Agentic Commerce Protocol can be understood as attempts to standardise interactions within this multi-sided ecosystem.^3,5 The underlying tensions - between openness and control, neutrality and self-preferencing - are illuminated by platform economics.

Behavioural economics, choice architecture and digital nudging

While traditional economics often assumes rational agents and transparent markets, the reality of digital commerce has always been shaped by design: the ordering of search results, the framing of options, the use of defaults, and the timing of prompts. Behavioural economists like Daniel Kahneman, Amos Tversky and Richard Thaler have demonstrated how real-world decision-making deviates from rational models and how "choice architecture" can influence outcomes.

In online retail, this has manifested as a rich literature on digital nudging: subtle interface choices that steer behaviour. Researchers in human-computer interaction and behavioural science have documented how factors such as social proof, scarcity cues and personalised messaging affect conversion.

AI-powered shopping agents add another layer. Instead of static designs, the conversation itself becomes the choice architecture. The way an AI agent frames options, in what order it presents them, how it responds to hesitation and how it explains trade-offs, all shape consumer welfare. Theorists working at the intersection of AI and behavioural economics are now grappling with questions of transparency, autonomy and manipulation in agentic environments.

Conversational AI and human-computer interaction

The ability to shop by talking to an AI depends on advances in natural language processing, dialogue modelling and user-centred design. The early work of Joseph Weizenbaum (ELIZA) and the subsequent development of chatbots provided the conceptual foundations, but the major leap came with deep learning and large language models.

Researchers such as Yoshua Bengio, Geoffrey Hinton and Yann LeCun advanced the neural network architectures that underpin todays generative models. Within natural language processing, work by many teams on sequence-to-sequence learning, attention mechanisms and transformer architectures led to systems capable of understanding and generating human-like text.

OpenAI popularised the transformer-based large language model with the GPT series, while Google researchers contributed foundational work on transformers and later developed models like BERT and its successors. These advances turned language interfaces from novelties into robust tools capable of handling complex, multi-turn interactions.

Human-computer interaction specialists, meanwhile, studied how people form mental models of conversational agents, how trust is built or undermined, and how to design dialogues that feel helpful rather than intrusive. The combination of technical capability and design insight has made it plausible for people to rely on an AI agent to curate shopping choices.

Autonomous agents and "agentic" AI

The term "agentic commerce" used by Walmart and Google points to a broader intellectual shift: viewing AI systems not just as passive tools but as agents capable of planning and executing sequences of actions.^1,5 In classical AI, agent theory has its roots in work on autonomous systems, reinforcement learning and decision-making under uncertainty.

Reinforcement learning theorists such as Richard Sutton and Andrew Barto formalised the idea of an agent learning to act in an environment to maximise reward. In ecommerce, this can translate into systems that learn how best to present options, when to offer discounts or how to balance immediate sales with long-term customer satisfaction.

Recent research on tool-using agents goes further, allowing language models to call external APIs, interact with databases and coordinate services. In commerce settings, that means an AI can check inventory, query shipping options, apply loyalty benefits and complete payments - all within a unified reasoning loop. Googles and OpenAIs protocols effectively define the "environment" in which such agents operate and the "tools" they can use.^3,5

The theoretical questions now concern safety, alignment and control: how to ensure that commercially motivated agents act in ways that are consistent with user interests and regulatory frameworks, and how to audit their behaviour when their decision-making is both data-driven and opaque.

Corporate protagonists: Google, OpenAI and Amazon

The Associated Press quote names three central actors, each with a distinct history and strategic posture.

Google: from search to Gemini-powered commerce

Google built its business on organising the worlds information and selling targeted advertising against search queries. Its dominance in web search made it the default starting point for many online shopping journeys. As user behaviour has shifted towards conversational interfaces and specialised shopping experiences, Google has sought to extend its role from search engine to AI companion.

Gemini, Googles family of large language models and AI assistants, sits at the heart of this effort. By integrating Gemini into retail scenarios, Google is attempting to ensure that when people ask an AI for help - planning a project, solving a problem or buying a product - it is their agent, not a competitors, that orchestrates the journey.^1,3,5

Partnerships with retailers such as Walmart, Target, Shopify, Wayfair and others, combined with the Universal Commerce Protocol, are strategic levers in this competition.^1,3,4,5 They allow Google to showcase Gemini as a shopping concierge while making it easier for merchants to plug into the ecosystem without bespoke integrations for each AI platform.

OpenAI: from research lab to commerce gateway

OpenAI began as a research-focused organisation with a mission to ensure that artificial general intelligence benefits humanity. Over time, it has commercialised its work through APIs and flagship products such as ChatGPT, which rapidly became one of the fastest-growing consumer applications in history.

As users started to rely on ChatGPT not just for information but for planning and decision-making, the platform became an attractive entry point for commerce. OpenAIs Instant Checkout feature and the Agentic Commerce Protocol reflect an attempt to formalise this role. By enabling users to buy directly within ChatGPT from merchants on platforms like Shopify and Etsy, OpenAI is turning its assistant into a transactional hub.^2,3

In this model, the AI agent can browse catalogues, compare options and present distilled choices, collapsing the distance between advice and action. The underlying theory draws on both conversational AI and platform economics: OpenAI positions itself as a neutral interface layer connecting consumers and merchants, while also shaping how information and offers are presented.

Amazon: marketplace, infrastructure and the invisible AI layer

While the provided context focuses more explicitly on Google and OpenAI, Amazon is an equally significant player in AI-powered shopping. Its marketplace already acts as a giant, data-rich environment where search, recommendation and advertising interact.

Amazon has deployed AI across its operations: in demand forecasting, warehouse robotics, delivery routing, pricing optimisation and its Alexa voice assistant. It has also invested heavily in generative AI to enhance product search, summarise reviews and assist sellers with content creation.

From a theoretical standpoint, Amazon exemplifies the vertically integrated platform: it operates the marketplace, offers its own branded products, controls logistics and, increasingly, provides the AI services that mediate discovery. Its approach to AI shopping is therefore as much about improving internal efficiency and customer experience as about creating open protocols.

In the race described by AP, Amazons strengths lie in its end-to-end control of the commerce stack and its granular data on real-world purchasing behaviour. As conversational and agentic interfaces become more common, Amazon is well placed to embed them deeply into its existing shopping flows.

Retailers as co-architects of AI shopping

Although the quote highlights technology companies, retailers such as Walmart, Target and others are not passive recipients of AI tools. They are actively shaping how agentic commerce unfolds. Walmart, for example, has worked with both OpenAI and Google, enabling Instant Checkout in ChatGPT and integrating its catalogue and fulfilment options into Gemini.^1,2,3

Walmart executives have spoken about "rewriting the retail playbook" and closing the gap between "I want it" and "I have it" using AI.² The company has also launched its own AI assistant, Sparky, within its app, and has been candid about how AI will transform roles across its workforce.²

These moves reflect a broader theoretical insight from platform economics: large retailers must navigate their relationships with powerful technology platforms carefully, balancing the benefits of reach and innovation against the risk of ceding too much control over customer relationships. By participating in open protocols and engaging multiple AI partners, retailers seek to maintain some leverage and avoid lock-in.

Other retailers and adjacent companies are exploring similar paths. Home Depot, for instance, has adopted Gemini-based agents to provide project planning and aisle-level guidance in stores, while industrial partners like Honeywell are using AI to turn physical spaces into intelligent, sensor-rich environments.⁵ These developments blur the line between online and offline shopping, extending the idea of seamless AI-powered commerce into bricks-and-mortar settings.

The emerging theory of AI-mediated markets

As AI agents become more entwined with commerce, several theoretical threads are converging into what might be called the theory of AI-mediated markets:

• Information symmetry and asymmetry: AI agents can, in principle, reduce information overload and help consumers navigate complex choices. But they also create new asymmetries, as platform owners may know far more about aggregate behaviour than individual users.
• Algorithmic transparency and accountability: When an AI agent chooses which products to recommend, the criteria may include relevance, profit margins, sponsorship and long-term engagement. Understanding and governing these priorities is an active area of research and regulation.
• Competition and interoperability: The existence of multiple commerce protocols and agent ecosystems raises questions about interoperability, switching costs and the potential for AI-mediated markets to become more or less competitive than their predecessors.
• Personalisation versus autonomy: Enhanced personalisation can make shopping more efficient and enjoyable but may also narrow exposure to alternatives or gently steer behaviour in ways that users do not fully perceive.
• Labour and organisational change: As AI takes on more of the cognitive labour of retail - from customer service to merchandising - the roles of human workers evolve. The theoretical work on technology and labour markets gains a new frontier in AI-augmented retail operations.

Researchers from economics, computer science, law and sociology are increasingly studying these dynamics, building on the earlier theories of platforms, recommendations and behavioural biases but extending them into a world where the primary interface to the market is itself an intelligent agent.

Why this moment matters

The Associated Press quote distils a complex, multi-layered transformation into a single observation: the most powerful technology firms are in a race to define how we shop in an age of AI. The endpoint of that race is not just faster checkout or more targeted ads. It is a restructuring of the basic relationship between consumers, merchants and the digital intermediaries that connect them.

Search boxes and product grids are giving way to conversations. Static ecommerce sites are being replaced or overlaid by agents that can understand context, remember preferences and act on our behalf. The theories of information retrieval, recommendation, platforms and behavioural economics that once described separate facets of digital commerce are converging in these agents.

Understanding the backstory of this quote - the intellectual currents, corporate strategies and emerging protocols behind it - is essential for grasping the stakes of AI-powered shopping. It is not merely a technological upgrade; it is a shift in who designs, controls and benefits from the everyday journeys that connect intention to action in the digital economy.

Term: Simple exponential smoothing (SES)

"The Exponential Smoothing technique is a powerful forecasting method that applies exponentially decreasing weights to past observations. This method prioritizes recent information, making it significantly more responsive than SMAs to sudden shifts." - Simple exponential smoothing (SES) -

Simple Exponential Smoothing (SES) is the simplest form of exponential smoothing, a time series forecasting method that applies exponentially decreasing weights to past observations, prioritising recent data to produce responsive forecasts for series without trend or seasonality.^1,2,3,5

Core Definition and Mechanism

SES generates point forecasts by recursively updating a single smoothed level value, (\ellt), using the formula:
\ell</em>t = \alpha y<em>t + (1 - \alpha) \ell</em>
where (yt) is the observation at time (t), (\ell) is the previous level, and (\alpha) (0 < (\alpha) < 1) is the smoothing parameter controlling the weight on the latest observation.^1,2,3,5 The forecast for all future periods is then the current level: (\hatt = \ellt).⁵

Quote: Pitchbook

"Much of the market continues to find it difficult to raise venture capital funding. Non-AI companies have accounted for just 35% of deal value through Q3 2025, while representing more than 60% of completed deals." - Pitchbook

PitchBook's data through Q3 2025 reveals a stark disparity in venture capital (VC) funding, where non-AI companies captured just 35% of total deal value despite comprising over 60% of deals, underscoring investor preference for AI-driven opportunities amid market caution.^1,4,5

Context of the Quote

This statistic, sourced from PitchBook's Q3 2025 Venture Monitor (in collaboration with the National Venture Capital Association), highlights the "flight to quality" trend dominating VC dealmaking. Through the first nine months of 2025, overall deal counts reached 3,990 in Q1 alone (up 11% quarter-over-quarter), with total value hitting $91.5 billion—a post-2022 high driven largely by AI sectors.^4,5 However, smaller and earlier-stage non-AI startups received only 36% of total value, the decade's lowest share, as investors prioritized larger, AI-focused rounds amid uncertainties like tariffs, market volatility, and subdued consumer sentiment.^3,4 Fundraising for VC funds also plummeted, with Q1 2025 seeing just 87 vehicles close at $10 billion—the lowest activity in over a decade—and dry powder nearing $300 billion but deploying slowly.⁴ Exit activity hinted at recovery ($56 billion in Q1 from 385 deals) but faltered due to paused IPOs (e.g., Klarna, StubHub) and reliance on outliers like Coreweave's IPO, which accounted for nearly 40% of value.⁴ PitchBook's H1 2025 VC Tech Survey of 32 investors confirmed this shift: 52% see AI disrupting fintech (up from 32% in H2 2024), with healthcare, enterprise tech, and cybersecurity following suit, while VC outlooks soured (only 38% expect rising funding, down from 58%).¹ The quote encapsulates a market where volume persists but value concentrates in AI, leaving non-AI firms struggling for capital in a selective environment.

Backstory on PitchBook

PitchBook, founded in 2007 by John Gabbert in Seattle, emerged as a leading data provider for private capital markets from humble origins as a simple Excel-based tool for tracking VC and private equity deals. Acquired by Morningstar in 2016 for $225 million, it has grown into an authoritative platform aggregating data on over 3 million companies, 1.5 million funds, and millions of deals worldwide, powering reports like the PitchBook-NVCA Venture Monitor.^3,4,5 Its Q3 2025 analysis draws from proprietary datasets as of late 2025, offering granular insights into deal counts, values, sector breakdowns, and fundraising—essential for investors navigating post-2022 VC normalization. PitchBook's influence stems from its real-time tracking and predictive modeling, cited across industry reports for benchmarking trends like AI dominance and liquidity pressures.^1,2,4

Leading Theorists on VC Market Dynamics and AI Concentration

The quote aligns with foundational theories on VC cycles, power laws, and technological disruption. Key thinkers include:

• Bill Janeway (author of Doing Capitalism in the Innovation Economy, 2012): A veteran VC at Warburg Pincus, Janeway theorized VC as a "three-legged stool" of government R&D, entrepreneurial risk-taking, and financial engineering. He predicted funding concentration in breakthrough tech like AI during downturns, as investors seek "moonshots" amid capital scarcity—mirroring 2025's non-AI value drought.^1,4

• Peter Thiel (co-founder of PayPal, Founders Fund; Zero to One, 2014): Thiel's "definite optimism" framework argues VCs favor monopolistic, tech-dominant firms (e.g., AI) over competitive commoditized ones, enforcing power-law distributions where 80-90% of returns come from 1-2% of deals. This explains non-AI firms' deal volume without value, as Thiel warns against "indefinite optimism" in crowded sectors.⁴

• Andy Kessler (author of Venture Capital Deals, 1986; Wall Street Journal columnist): Kessler formalized the VC "spray and pray" model evolving into selective bets during liquidity crunches, predicting AI-like waves would eclipse legacy sectors—evident in 2025's fintech AI disruption forecasts.¹

• Scott Kupor (a16z managing partner; Secrets of Sand Hill Road, 2019): Kupor analyzes LP-VC dynamics, noting how dry powder buildup (nearing $300B in 2025) leads to extended fund timelines and AI favoritism, as LPs demand outsized returns amid low distributions.^1,2,4

• Diane Mulcahy (former Providence Equity; The New World of Entrepreneurship, 2013): Mulcahy critiqued VC overfunding bubbles, advocating "patient capital" for non-hyped sectors; her warnings resonate in 2025's fundraising cliff and non-AI funding gaps.⁴

These theorists collectively frame 2025's trends as a power-law amplification of AI amid cyclical caution, building on historical VC patterns from the dot-com bust to post-2008 recovery.

Quote: Nathaniel Whittemore - AI Daily Brief

"If you want to get a preview of what everyone else is going to be dealing with six months from now, there's basically not much better you can do than watching what developers are talking about right now." - Nathaniel Whittemore - AI Daily Brief - On: Tailwind CSS and AI disruption

This observation captures a pattern that has repeated itself through every major technology wave of the past half-century. The people who live closest to the tools - the engineers, open source maintainers and framework authors - are usually the first to encounter both the power and the problems that the rest of the world will later experience at scale. In the current artificial intelligence cycle, that dynamic is especially clear: developers are experimenting with new models, agents and workflows months before they become mainstream in business, design and everyday work.

Nathaniel Whittemore and the AI Daily Brief

The quote comes from Nathaniel Whittemore, better known in technology circles as NLW, the host of The AI Daily Brief: Artificial Intelligence News and Analysis (formerly The AI Breakdown).^4,7,9 His show has emerged as a daily digest and analytical lens on the rapid cascade of AI announcements, research papers, open source projects and enterprise case studies. Rather than purely cataloguing news, Whittemore focuses on how AI is reshaping business models, labour, creative work and the broader economy.⁴

Whittemore has built a reputation as an interpreter between worlds: the fast-moving communities of AI engineers and builders on the one hand, and executives, policymakers and non-technical leaders on the other. Episodes range from detailed walkthroughs of specific tools and models to long-read analyses of how organisations are actually deploying AI in the field.^1,5 His recurring argument is that the most important AI stories are not just technical; they are about context, incentives and the way capabilities diffuse into real workflows.^1,4

On his show and in talks, Whittemore frequently returns to the idea that AI is best understood through its users: the people who push tools to their limits, improvise around their weaknesses and discover entirely new categories of use. In recent years, that has meant tracking developers who integrate AI into code editors, build autonomous agents, or restructure internal systems around AI-native processes.^3,8 The quote about watching developers is, in effect, a mental model for anyone trying to see around the next corner.

Tailwind CSS as the context for the quote

The quote lands inside a very specific story: Tailwind CSS as a case study in AI-enabled demand with AI-damaged monetisation.

Tailwind is an open-source, utility-first CSS framework that became foundational to modern front-end development. It is widely adopted by developers and heavily used by AI coding tools. Tailwind’s commercial model, however, depends on a familiar open-source pattern: the core framework is free, and revenue comes from paid add-ons (the “Plus” tier). Critically, the primary channel to market for those paid offerings was the documentation.

AI broke that channel.

As AI coding tools improved, many developers stopped visiting documentation pages. Instead, they asked the model and got the answer immediately—often derived from scraped docs and community content. Usage of Tailwind continued to grow, but the discovery path for paid offerings weakened because humans no longer needed to read the docs. In plain terms: the product stayed popular, but the funnel collapsed.

That is why this story resonated beyond CSS. It shows a broader pattern: AI can remove the need for the interface you monetise—even while it increases underlying adoption. For any business that relies on “users visit our site, then convert,” Tailwind is not a niche developer drama. It is a preview.

Tailwind’s episode makes the mechanism of disruption uncomfortably clear: AI tools boosted adoption, but also removed the need for humans to visit Tailwind’s documentation. That mattered because the documentation was Tailwind’s primary channel to market—where users discovered the paid “Plus” offerings that funded maintenance. Once AI started answering questions directly from scraped content, the funnel broke: fewer doc visits meant fewer conversions, and a widely used framework suddenly struggled to monetise the very popularity AI helped accelerate.

AI Disruption Seen from the Builder Front Line

In the AI era, this pattern is amplified. AI capabilities roll out as research models, APIs and open source libraries long before they are wrapped in polished consumer interfaces. Developers are often the first group to:

• Benchmark new models, probing their strengths and failure modes.
• Integrate them into code editors, data pipelines, content tools and internal dashboards.
• Build specialised agents tuned to niche workflows or industry-specific tasks.^6,8
• Stress-test the economics of running models at scale and find where they can genuinely replace or augment existing systems.^3,5

Whittemore's work sits precisely at this frontier. Episodes dissect the emergence of coding agents, the economics of inference, the rise of AI-enabled "tiny teams", and the way reasoning models are changing expectations around what software can autonomously do.^3,8 He tracks how new agentic capabilities go from developer experiments to production deployments in enterprises, often in less than a year.^3,5

His quote reframes this not as a curiosity but as a practical strategy: if you want to understand what your organisation or industry will be wrestling with in six to twelve months - from new productivity plateaus to unfamiliar risks - you should look closely at what AI engineers and open source maintainers are building and debating now.

Developers as Lead Users: Theoretical Roots

Behind Whittemore's intuition sits a substantial body of innovation research. Long before AI, scholars studied why certain groups seemed to anticipate the needs and behaviours of the wider market. Several theoretical strands help explain why watching developers is so powerful.

Eric von Hippel and Lead User Theory

MIT innovation scholar Eric von Hippel developed lead user theory to describe how some users experience needs earlier and more intensely than the general market. These lead users frequently innovate on their own, building or modifying products to solve their specific problems. Over time, their solutions diffuse and shape commercial offerings.

Developers often fit this lead user profile in technology markets. They are:

• Confronted with cutting-edge problems first - scaling systems, integrating new protocols, or handling novel data types.
• Motivated to create tools and workflows that relieve their own bottlenecks.
• Embedded in communities where ideas, snippets and early projects can spread quickly and be iterated upon.

Tailwind CSS itself reflects this: it emerged as a developer-centric solution to recurring front-end pain points, then radiated outward to reshape how teams approach design systems. In AI, developer-built tooling often precedes large commercial platforms, as seen with early AI coding assistants, monitoring tools and evaluation frameworks.^3,8

Everett Rogers and the Diffusion of Innovations

Everett Rogers' classic work on the diffusion of innovations describes how new ideas spread through populations in phases: innovators, early adopters, early majority, late majority and laggards. Developers often occupy the innovator or early adopter categories for digital technologies.

Rogers stressed that watching these early groups offers a glimpse of future mainstream adoption. Their experiments reveal not only whether a technology is technically possible, but how it will be framed, understood and integrated into social systems. In AI, the debates developers have about safety, guardrails, interpretability and tooling are precursors to the regulatory, ethical and organisational questions that follow at scale.^4,5

Clayton Christensen and Disruptive Innovation

Clayton Christensen's theory of disruptive innovation emphasises how new technologies often begin in niches that incumbents overlook. Early adopters tolerate rough edges because they value new attributes - lower cost, flexibility, or a different performance dimension - that established customers do not yet prioritise.

AI tools and frameworks frequently begin life like this: half-finished interfaces wrapped around powerful primitives, attractive primarily to technical users who can work around their limitations. Developers discover where these tools are genuinely good enough, and in doing so, they map the path by which a once-nascent capability becomes a serious competitive threat.

Open Source Communities and Collective Foresight

Another important line of thinking comes from research on open source software and user-driven innovation. Scholars such as Steven Weber and Yochai Benkler have explored how distributed communities coordinate to build complex systems without traditional firm structures.

These communities act as collective sensing networks. Bug reports, pull requests, issue threads and design discussions form a live laboratory where emerging practices are tested and refined. In AI, this is visible in the rapid evolution of open weights models, fine-tuning techniques, evaluation harnesses and orchestration frameworks. The tempo of progress in these spaces often sets the expectations which commercial vendors then have to match or exceed.^6,8

AI-Specific Perspectives: From Labs to Production

Beyond general innovation theory, several contemporary AI thinkers and practitioners shed light on why developer conversations are such powerful predictors.

Andrej Karpathy and the Software 2.0 Vision

Former Tesla AI director Andrej Karpathy popularised the term "Software 2.0" to describe a shift from hand-written rules to learned neural networks. In this paradigm, developers focus less on explicit logic and more on data curation, model selection and feedback loops.

Under a Software 2.0 lens, developers are again early indicators. They experiment with prompt engineering, fine-tuning, retrieval-augmented generation and multi-agent systems. Their day-to-day struggles - with context windows, hallucinations, latency and cost-performance trade-offs - foreshadow the operational questions businesses later face when they automate processes or embed AI in products.

Ian Goodfellow, Yoshua Bengio and Deep Learning Pioneers

Deep learning pioneers such as Ian Goodfellow, Yoshua Bengio and Geoffrey Hinton illustrated how research breakthroughs travel from lab settings into practical systems. What began as improvements on benchmark datasets and academic competitions became, within a few years, the foundation for translation services, recommendation engines, speech recognition and image analysis.

Developers building on these techniques were the bridge between research and industry. They discovered how to deploy models at scale, handle real-world data, and integrate AI into existing stacks. In today's generative AI landscape, the same dynamic holds: frontier models and architectures are translated into frameworks, SDKs and reference implementations by developer communities, and only then absorbed into mainstream tools.

AI Engineers and the Rise of Agents

Recent work at the intersection of AI and software engineering has focused on agents: AI systems that can plan, call tools, write and execute code, and iteratively refine their own outputs. Industry reports summarised on The AI Daily Brief highlight how executives are beginning to grasp the impact of these agents on workflows and organisational design.⁵

Yet developers have been living with these systems for longer. They are the ones:

• Embedding agents into CI/CD pipelines and testing regimes.
• Using them to generate and refactor large codebases.^3,6
• Designing guardrails and permissions to keep them within acceptable bounds.
• Developing evaluation harnesses to measure quality, robustness and reliability.⁸

Their experiments and post-mortems provide an unvarnished account of both the promise and the fragility of agentic systems. When Whittemore advises watching what developers are talking about, this is part of what he means: the real-world friction points that will later surface as board-level concerns.

Context, Memory and Business Adoption

Whittemore has also emphasised how advances in context and memory - the ability of AI systems to integrate and recall large bodies of information - are changing what is possible in the enterprise.¹ He highlights features such as:

• Tools that allow models to access internal documents, code repositories and communication platforms securely, enabling organisation-specific reasoning.¹
• Modular context systems that let AI draw on different knowledge packs depending on the task.¹
• Emerging expectations that AI should "remember" ongoing projects, preferences and constraints rather than treating each interaction as isolated.¹

Once again, developers are at the forefront. They are wiring these systems into data warehouses, knowledge graphs and production applications. They see early where context systems break, where privacy models need strengthening, and where the productivity gains are real rather than speculative.

From there, insights filter into broader business discourse: about data governance, AI strategy, vendor selection and the design of AI-native workflows. The lag between developer experience and executive recognition is, in Whittemore's estimate, often measured in months - hence his six-month framing.

From Developer Talk to Strategic Foresight

The core message behind the quote is a practical discipline for anyone thinking about AI and software-driven change:

• Follow where developers invest their time. Tools that inspire side projects, plugin ecosystems and community events often signal deeper shifts in how work will be done.
• Listen to what frustrates them. Complaints about context limits, flaky APIs or insufficient observability reveal where new infrastructure, standards or governance will be needed.
• Pay attention to what they take for granted. When a capability stops being exciting and becomes expected - instant code search, semantic retrieval, AI-assisted refactoring - it is often a sign that broader expectations in the market will soon adjust.
• Watch the crossovers. When developer patterns show up in no-code tools, productivity suites or design platforms, the wave is moving from early adopters to the early majority.

Nathaniel Whittemore's work with The AI Daily Brief is, in many ways, a structured practice of this approach. By curating, analysing and contextualising what builders are doing and saying in real time, he offers a way for non-technical leaders to see the outlines of the future before it is evenly distributed.^4,7,9 The Tailwind CSS example is one case; the ongoing wave of AI disruption is another. The constant, across both, is that if you want to know what is coming next, you start by watching the people building it.

References

1. https://pod.wave.co/podcast/the-ai-daily-brief-formerly-the-ai-breakdown-artificial-intelligence-news-and-analysis/ai-context-gets-a-major-upgrade

2. https://www.youtube.com/watch?v=MdfYA3xv8jw

3. https://www.youtube.com/watch?v=0EDdQchuWsA

4. https://podcasts.apple.com/us/podcast/the-ai-daily-brief-artificial-intelligence-news/id1680633614

5. https://www.youtube.com/watch?v=nDDWWCqnR60

6. https://www.youtube.com/watch?v=f34QFs7tVjg

7. https://open.spotify.com/show/7gKwwMLFLc6RmjmRpbMtEO

8. https://podcasts.apple.com/us/podcast/the-biggest-trends-from-the-ai-engineer-worlds-fair/id1680633614?i=1000711906377

9. https://www.audible.com/podcast/The-AI-Breakdown-Daily-Artificial-Intelligence-News-and-Discussions/B0C3Q4BG17