“We’re not replacing all of our people with digital agents, but we’re going expect you to understand how to work with a digital agent… Or if you’re in the front office part of the firm talking to a client, we’re going expect you to understand how to use that tooling to make your relationships stronger and to deliver a better value proposition.” – John Waldron – Goldman Sachs
The immediate pressure inside large financial institutions is not a cinematic wave of robot bankers, but a subtler requirement: every human professional is being nudged to become a hybrid operator who can orchestrate digital agents alongside traditional judgement 1. In investment banking, markets, and wealth management, the constraint is no longer computing power; it is how quickly relationship managers, product specialists, and risk teams can retool their daily workflows so that artificial intelligence becomes a multiplier rather than a parallel system they quietly ignore.
This shift reflects a hard commercial edge. If a rival bank can prepare pitches, price scenarios, and risk views in a fraction of the time because its staff are fluent in agentic tooling, the competitive gap compounds quickly. In primary issuance, the team that can synthesise cross-asset signals, simulate client outcomes, and draft tailored communication overnight will win mandates that slower teams never see. Waldron’s message, framed as an expectation rather than an option, signals that the cultural battle is no longer about whether AI will be used, but about who learns to drive it well enough to matter in front of clients 1.
The meaning behind “not replacing all of our people”
On its surface, the reassurance that people are not being wholesale replaced sounds like standard corporate calming language. In context, it is closer to a restructuring of what counts as baseline competence. For a long time, elite finance rewarded those who could manually process information faster than peers: building models at speed, scanning research, remembering client constraints. As digital agents become competent at these tasks, the advantage migrates toward those who can specify problems, structure prompts, check outputs, and integrate them into client-facing decisions.
In practice, that means the benchmark analyst or associate no longer differentiates themselves by staying later to scrub spreadsheets. The organisation is explicitly saying that survival requires learning how to shape, interrogate, and supervise systems that perform the mechanical work. The phrase “we’re not replacing all of our people” is therefore less a promise of job security and more a statement that replacement will be selective and correlated with an individual’s ability to collaborate with the new tools rather than compete against them.
For senior leaders, this reframes talent strategy. Instead of assessing whether someone can master an existing process, they now need to ask whether this person can translate messy, client-specific questions into forms that a digital agent can process reliably, and then explain the augmented answer back to a client in plain language. Those who cannot bridge that gap risk becoming the human bottleneck in an otherwise modernised workflow.
Why front-office AI fluency matters more than back-office automation
Back-office automation has long been an easier sell: it cuts costs, reduces errors, and rarely touches the firm’s image with clients. Front-office AI is far more sensitive. When a president and chief operating officer signals that client-facing staff must understand how to work with digital agents, he is betting that competitive advantage now lies in visibly augmented advice and responsiveness, not only in silent efficiency gains behind the scenes 1.
Consider how a traditional relationship manager prepares for a meeting with a corporate treasurer. Historically, they might pull recent trading history, read internal research, and ask a product team for structured ideas. With agentic tools, the same banker could have a digital assistant that ingests the client’s historical exposure, peer behaviour, current market conditions, and regulatory constraints, then generates scenario-specific talking points, risk flags, and tailored documentation drafts. The human still leads the relationship, but the quality and breadth of information they bring to the table are transformed.
The strategic tension is that such augmentation subtly changes what clients expect. Once a few banks routinely show up with richer insights and faster follow-ups thanks to AI assistance, the old level of preparation begins to look negligent. Front-office AI fluency therefore becomes not just an internal efficiency play, but an external signalling device: a proof that the institution is at the frontier of information use on the client’s behalf.
From “using a tool” to working with an agent
Language matters here. A tool is something you operate directly; an agent is something you delegate to. The move from simple interfaces to digital agents implies workflows where staff specify goals, context, and constraints, then monitor an autonomous or semi-autonomous system as it executes tasks across multiple data sources and applications.
Technically, these agents may chain calls to large language models, internal APIs, pricing engines, and document stores, deciding step by step which resource to invoke next. Conceptually, the employee’s role shifts toward being a supervisor and editor. Rather than manually compiling a client memo, they might instruct an agent: assemble current exposure, relevant research, regulatory developments, and three hedging strategies with quantified trade-offs. The agent generates a draft; the banker then checks, corrects, and contextualises it before sharing.
This shift raises questions of accountability. If an agent suggests a structure that later underperforms or creates unforeseen risks, how will the firm trace the reasoning? In markets, where internal models already embody layers of opacity, adding another AI-driven layer intensifies the need for auditability. Institutions like Goldman Sachs must therefore build governance frameworks in which every agent interaction is logged, reproducible, and attributable, so that human supervisors remain genuinely accountable rather than nominally in charge of a black box.
The quantitative core: productivity, risk, and value
Although Waldron’s statement is framed in everyday language, the underlying trade-offs can be expressed in simple quantitative terms. Suppose a banker’s effective output without AI is represented by Y_0, and the introduction of a well-used digital agent yields a proportional productivity gain \t\th\eta, where \t\th\eta > 0. The new output is Y_1 = (1 + \t\th\eta) Y_0. If a firm targets, for example, \t\th\eta = 0,3, each banker must deliver 30\% more insight, coverage, or revenue at similar or lower risk just to meet expectations.
From a risk perspective, let baseline decision quality (probability of making a correct or acceptable recommendation) be p_0. Introducing an AI agent with independent error probability e does not automatically improve outcomes. If the banker blindly follows the agent, the new error probability may be worse than 1 - p_0. In contrast, if the banker uses the agent as a second opinion and only acts when human judgement and agent output agree, a simplified model of agreement-based acceptance could be written as: error probability P_{err} = e (1 - p_0) + (1 - e)(1 - p_0) q, where q represents the chance of misinterpreting or mis-implementing a correct suggestion. The firm’s training and process design effectively aim to minimise P_{err} while maximising output Y_1.
Over a portfolio of relationships, expected value can be thought of as V = \sum_i R_i - C_i - L_i, where R_i is client revenue, C_i is servicing cost, and L_i is expected loss from errors or misconduct for relationship i. Digital agents are introduced to increase R_i by enabling more tailored solutions, reduce C_i by automating labour, and contain or reduce L_i through better risk detection. The danger, and the core of the internal debate, is that poorly governed AI may reduce C_i but increase L_i by enabling faster, more systematic mistakes.
Strategic context inside Goldman Sachs
Goldman Sachs is not approaching AI from a standing start. The firm has long invested in quantitative research, systematic strategies, and technology-heavy businesses. What is changing is the expectation that AI literacy will spread beyond quant desks and engineering teams into the heart of client coverage. Waldron’s role as president and chief operating officer gives his words operational force: they hint at performance assessments, promotion criteria, and training programmes being redesigned around AI fluency as a core competency 1.
In a firm whose brand rests on high-touch advisory relationships, this introduces a delicate balancing act. On one hand, digital agents promise to scale the reach of top talent: ideas from a leading sector banker or risk specialist can be embedded into prompts and templates used across hundreds of colleagues. On the other hand, there is a reputational risk if clients perceive interactions as scripted, generic, or driven more by machine output than by genuine understanding of their unique situation.
To manage this, Goldman must distinguish between invisible augmentation and visible automation. Invisible augmentation quietly improves the quality, timeliness, and depth of what human bankers bring to meetings. Visible automation, by contrast, risks making conversations feel standardised or mechanised. Waldron’s emphasis on using tooling to make relationships stronger, rather than simply more efficient, implicitly recognises this tension and anchors the strategy in client experience, not just internal cost metrics.
Internal debates and cultural resistance
Within any large bank, reactions to such expectations fall along a spectrum. Younger staff may be eager adopters, viewing AI agents as a way to compress years of manual apprenticeship into a faster learning curve. Mid-career professionals may worry that their hard-won pattern recognition is being devalued or that they will be judged on unfamiliar technical skills. Senior rainmakers may insist that relationships are fundamentally about trust, implying that technology is peripheral.
One internal debate concerns whether mandating AI use risks creating over-reliance. If every pitch deck, market commentary, or client note flows through a digital agent, the danger is that staff gradually lose the habit of constructing narratives and checking numbers themselves. That could hollow out the talent pipeline, leaving the firm with a generation of bankers who are adept at editing but less capable of building original analysis from first principles.
Another debate centres on fairness and performance measurement. If some teams gain early access to powerful agents while others are limited to basic tools, comparisons of productivity or revenue generation become distorted. The firm must therefore decide whether to standardise agent capabilities across divisions or tolerate a period of uneven adoption while experimenting. In a competitive, bonus-driven culture, these choices carry real political weight.
Objections from clients, regulators, and staff
Clients may voice concerns that their confidential data is being fed into opaque systems, potentially used to train models or inform strategies for competitors. To reassure them, institutions need clear boundaries around data use, on-premise or private-cloud deployments, and the ability to explain where and how AI is applied in servicing a particular account. Without that, assurances about stronger relationships may ring hollow.
Regulators, for their part, are already alert to the risks of model-driven decision-making. When client-facing staff rely on digital agents for suitability assessments, product recommendations, or pricing, questions arise about explainability, bias, and control. Supervisory authorities may require firms to demonstrate that AI-assisted processes comply with existing conduct, suitability, and market abuse rules, even if the rules were written long before such systems existed.
Staff themselves may object on ethical grounds, worrying that AI-enabled surveillance of their work will intensify. Because digital agents often sit within monitored platforms, every prompt, draft, and correction can be logged and analysed. This creates opportunities for training and quality control, but also a sense that the margin for informal experimentation is shrinking. Waldron’s expectation that employees learn to use such tools sits against this backdrop of expanded visibility.
Capability building: from training to institutional memory
For a mandate like this to take hold, the firm must invest heavily in capability building. Generic AI literacy courses will not suffice. Relationship managers need domain-specific playbooks: how to brief an agent about a mid-market industrial client versus a sovereign wealth fund; what prompts yield useful early-stage M&A idea screens; how to constrain an agent to only use pre-approved language when drafting sensitive communications.
Over time, the knowledge embedded in these playbooks may become an asset in its own right. As staff experiment, refine prompts, and correct outputs, the firm can capture these interactions, distil best practices, and encode them into institutional templates. In effect, the organisation starts to build a second layer of institutional memory residing not only in human experience and documentation, but in the configuration of its digital agents.
This, however, raises governance questions similar to those that arose around spreadsheets and internal models decades earlier. Who owns a particularly effective prompt structure: the individual banker, their team, or the firm? How are changes to shared agent templates tested and approved? If a flawed prompt leads to a systematic error in client materials, at what point does responsibility shift from the individual user to the central team that maintains the agent framework?
Why this expectation matters beyond Goldman Sachs
When a leading global bank tells its people that working with digital agents is no longer optional, it sends a signal across the industry. Competitors must either match the expectation, risking cultural backlash of their own, or accept that their staff may be less augmented in critical client interactions. Smaller institutions and buy-side firms will watch closely, weighing whether they can emulate the approach without the same scale of technology investment.
Labour markets will also respond. Candidates entering finance increasingly face an implicit test: not just whether they understand accounting, valuation, and markets, but whether they can demonstrate practical fluency with AI-assisted workflows. Business schools and professional training programmes are likely to adapt their curricula accordingly, blending traditional financial theory with hands-on experience in specifying, supervising, and critiquing digital agents.
In the longer term, the expectation that every client-facing professional can work with a digital agent accelerates a broader transformation in how expertise is distributed. Some forms of know-how that were once scarce and localised may become widely accessible through well-designed agents. What remains scarce is the ability to orchestrate that know-how in complex, high-stakes human situations: to choose when to lean on the agent, when to override it, and how to translate its suggestions into actions that preserve trust.
A quiet redefinition of professional competence
Underneath the reassuring language about not replacing people lies a redefinition of what it means to be a high-performing professional in finance. Competence now includes an ability to collaborate with systems that are probabilistic, opaque in places, and evolving. Relationship strength is no longer measured only by how well a banker remembers a client’s history or instincts, but by how effectively they can harness institutional data and AI tooling to anticipate needs and deliver tailored solutions.
In that sense, Waldron’s statement is less about technology and more about identity. The banker, trader, or advisor who sees themselves solely as an individual expert risks obsolescence; the one who sees themselves as the conductor of a small orchestra of digital agents, data sources, and human colleagues is closer to the emerging norm. As these expectations spread, the boundary between human judgement and machine assistance becomes less a line of defence and more a design problem: how to structure work so that each does what it is uniquely good at, in service of relationships that are, paradoxically, meant to feel more human than ever.
References
1 Sonali Basak, “Goldman’s AI Expectations” (LinkedIn analysis of John Waldron’s remarks).
References
1. “Goldman’s AI Expectations” – https://www.linkedin.com/pulse/goldmans-ai-expectations-sonali-basak-ixg8e
2. Investment strategist outlines the next phase of AI trade – Fox Business – 2025-10-02 – https://www.foxbusiness.com/video/6380886913112
3. John E. Waldron – Goldman Sachs – https://www.goldmansachs.com/our-firm/our-people-and-leadership/leadership/executive-officers/john-waldron
