!["We’re seeing 20, 30, 40% productivity gains in coding, in some cases, much more than that. I think the more we can develop that [AI] capability, the better we'll be… we'll be able to push more output, you know, with a relatively similar number of coders." - John Waldron - Goldman Sachs](https://i0.wp.com/globaladvisors.biz/wp-content/uploads/2026/05/20260525_09h30_GlobalAdvisors_Marketing_Quote_JohnWaldron_GAQ.png?fit=1200%2C1200&ssl=1 "20260525_09h30_GlobalAdvisors_Marketing_Quote_JohnWaldron_GAQ")
“We’re seeing 20, 30, 40% productivity gains in coding, in some cases, much more than that. I think the more we can develop that [AI] capability, the better we’ll be… we’ll be able to push more output, you know, with a relatively similar number of coders.” – John Waldron – Goldman Sachs
Productivity claims on the scale now associated with generative AI would, if even half accurate, amount to one of the largest step-changes in white-collar work since the spreadsheet entered finance in the 1980s 1. For a bank like Goldman Sachs, whose economics are deeply leveraged to the output of knowledge workers rather than machines or factories, the idea that software development teams could reliably produce 20 to 40 percent more code, features, or shipped products without a corresponding increase in headcount is strategically transformative. It reshapes the cost base, alters the calculus of technology investment, and changes what counts as a competitive moat in financial services.
To understand the weight of this shift, it is important to see coding productivity not as a niche concern of the IT department, but as the engine room behind trading platforms, risk systems, client portals, and regulatory infrastructure. Every trading desk, risk committee, and operations team inside a modern investment bank ultimately sits atop layers of software. When senior leadership starts publicly anchoring on double-digit productivity gains from AI systems 1, they are signalling more than enthusiasm for a new tool; they are outlining a potential re-rating of how fast the firm can build, adapt, and respond to markets.
From incremental tools to multiplicative leverage
Historically, software productivity improvements in large banks have been incremental. Better integrated development environments, automated testing pipelines, and cloud infrastructure each chipped away at bottlenecks. But none promised to fundamentally alter the ratio between engineering headcount and shipped functionality. Generative AI coding tools, and increasingly AI agents that can manage multi-step tasks, are qualitatively different because they do not just automate infrastructure; they intervene directly in the intellectual work of design and implementation.
When leaders talk about 20, 30 or 40 percent productivity gains, they are implicitly referring to a bundle of effects: boilerplate generation, test creation, documentation drafting, code review assistance, and sometimes design suggestions. Each individually appears modest. Together, they turn into a form of multiplicative leverage: every engineer can attempt more ideas, refactor more aggressively, and support more systems without burning out. In a bank running thousands of applications, even a small reduction in friction compounds across portfolios and years.
In that sense, AI assistance functions less like a discrete tool and more like an ambient capability woven into the entire software lifecycle. Instead of one-off automation projects, you get a steady ratcheting-up of throughput across planning, coding, testing, and maintenance. That is what makes senior executives willing to talk publicly about a future in which output rises while headcount remains roughly flat: the promise of ambient leverage rather than isolated productivity hacks.
Goldman Sachs as a technology-driven bank
Goldman Sachs has, for years, framed itself as a technology company as much as an investment bank. The development of platforms like Marquee for institutional clients and Marcus in consumer banking depended on large-scale internal software capability. That positioning matters because it reveals why leadership is particularly sensitive to changes in the economics of code. When an institution already sees engineering as a revenue driver rather than a pure cost centre, a potential 40 percent improvement in software output is not just an efficiency story; it is a growth narrative.
John Waldron, as president and chief operating officer, sits at the intersection of business lines and operational capabilities. His comments about AI-enabled productivity gains are therefore not trivia from a technical specialist but guidance from someone responsible for balancing risk, cost, and strategic opportunity. When someone in that role speaks about being able to deliver more output with a relatively stable cohort of coders, he is hinting at a coming shift in the scale and ambition of internal projects as much as a possible evolution in staffing philosophy 1.
In a bank that spends billions annually on technology, modest percentage changes accumulate into hundreds of millions of dollars in effective capacity. The firm can either bank those gains as lower marginal cost or redeploy them into new products, better risk analytics, and differentiated client services. Historically, Goldman has often preferred to reinvest in capability rather than simply shrink, which suggests that AI productivity improvements are more likely to manifest as faster feature velocity, more experimentation, and broader platform coverage than as crude headcount reduction.
The mechanics of AI-enabled coding productivity
When executives cite specific percentages, they are usually drawing from early pilots, internal metrics, and vendor studies. In practice, productivity in software development is notoriously difficult to measure. Lines of code are a poor metric; story points and sprint velocity are context-dependent. Yet some mechanisms of improvement are sufficiently concrete to be credible.
First, AI tools reduce the time spent on repetitive or boilerplate tasks. Generating data access layers, configuration scaffolding, logging, and simple integration code can often be done in moments. Second, they accelerate developers through unfamiliar libraries or languages, by synthesising examples and suggested usage patterns quickly. Third, they assist in debugging and refactoring, offering alternative implementations or pointing to likely sources of error. Finally, they can draft tests and documentation, activities that are vital but often under-resourced in deadline-driven environments.
These are not speculative benefits. Engineers report that, when properly integrated, AI coding assistants cut the friction associated with starting work, reduce context-switching time, and make it easier to maintain focus on higher-level design. The gains are uneven across tasks and teams, but they skew heavily toward routine coding and glue work that, in aggregate, absorbs a substantial fraction of engineering hours in a complex enterprise environment.
Why stable headcount is part of the message
The claim that institutions will be able to push more output with a similar number of coders carries an implicit reassurance to employees and regulators: this is a story about augmentation, not immediate mass displacement. For a highly regulated bank, signalling that AI is a force for capability rather than indiscriminate job cuts helps maintain trust internally and externally. It also reflects a practical reality: complex financial systems cannot simply be rebuilt overnight by machines, nor can firms dispense with human oversight in areas touching client money, market integrity, or regulatory reporting.
In the near to medium term, AI-native development still depends on human engineers to frame problems, validate outputs, manage integration risks, and interpret ambiguous requirements. What changes is the ratio between time spent wrestling with syntax and time spent reasoning about system behaviour and business impact. In this environment, holding headcount roughly constant while raising expectations on throughput is a rational management posture: it keeps institutional knowledge in place while stretching teams toward higher-value work.
There is also a capital markets angle. Public statements from a top executive that output can grow without proportional increases in staff serve as a signal to investors about operating leverage. If technology spending can be kept under control while the firm expands its platform capabilities, then margins can improve without sacrificing growth. That narrative plays well with analysts accustomed to viewing banks as cost-heavy, regulation-burdened institutions struggling to differentiate themselves.
Strategic tension: efficiency vs innovation
Yet the same capability that allows a bank to deliver more with the same team can, if mismanaged, produce a culture of relentless efficiency at the expense of thoughtful innovation. When every developer can suddenly be assigned more work, there is a risk that AI productivity gains simply feed into higher utilisation targets and tighter deadlines, rather than curiosity-driven exploration. The strategic question for leadership becomes whether to frame AI as a way to strip out cost or as a means of investing in better products and resilience.
For Goldman Sachs, which competes both with large universal banks and nimble fintech firms, the temptation to use AI purely to reduce technology budgets will be balanced by the need to maintain a reputation for sophisticated, high-touch services. Pushing more output with the same headcount could mean more risk models, better scenario analysis, richer client analytics, or more real-time insights for traders. Alternatively, it could deteriorate into feature bloat and technical debt if the organisational incentives reward shipping volume over quality.
This tension is amplified by the fact that AI-generated code is not free from defects or biases. If teams move faster but do not invest proportionally in testing, observability, and governance, the bank could accumulate invisible vulnerabilities inside trading systems, pricing engines, or compliance tools. The costs of such vulnerabilities in regulated markets can be severe, including fines, reputational damage, and forced remediation programmes.
Risk, control, and regulatory scrutiny
Large financial institutions operate under intense regulatory oversight. Supervisors are already paying attention to how critical models are built, validated, and monitored. The introduction of AI into the software factory raises new questions: how is training data governed, who is accountable for AI-suggested changes that later prove faulty, and what happens when model-driven tools shape systems that themselves influence markets?
Regulators will not accept a defence that blames AI tools for coding mistakes. For all practical purposes, responsibility still rests with the bank and its human staff. As a result, the push toward higher productivity must be accompanied by equally disciplined controls: robust code review processes, traceability of changes, clear documentation of where and how AI tools are used, and internal policies on acceptable use. If AI systems are allowed to generate substantial portions of risk or pricing engines, the validation burden will rise accordingly.
There is also an emerging concern about systemic risk. If many large institutions adopt similar AI tools, trained on overlapping corpora of code and best practices, there is a possibility of correlated failure modes. Subtle bugs or design patterns recommended by these tools might propagate across institutions, creating common vulnerabilities. From a systemic perspective, productivity gains at the micro level could translate into new forms of macro fragility if they reduce diversity in implementation approaches.
Debates and objections
Not everyone accepts the headline productivity statistics at face value. Critics argue that current measurement approaches often capture short-term speed-ups in micro-tasks but underweight the costs of mis-specification, integration, long-term maintenance, and security hardening. A developer who codes faster using AI might inadvertently accept suggestions that introduce subtle performance problems or security gaps, which only surface months later.
There is also concern that productivity gains are unevenly distributed. Senior engineers who already understand system architecture may gain moderately from autocomplete and boilerplate generation; junior developers might gain more but also face a risk of becoming dependent on tools instead of learning fundamentals. Over time, this could erode the depth of expertise in the organisation, making it harder to tackle novel problems that fall outside the distribution of patterns seen during AI training.
Another objection centres on cultural dynamics. When managers hear enthusiastic numbers about 30 or 40 percent gains, some may treat them as a new baseline, rather than as an upside scenario. That can erode trust if frontline engineers feel that executive optimism is not grounded in the realities of complex legacy systems, regulatory constraints, or the sheer difficulty of coordinating large programmes of work in a bank with many stakeholders.
Why AI coding productivity matters for finance
Despite these debates, the broader significance of AI-driven productivity in software is hard to overstate for financial services. The structure of modern markets is increasingly defined by code: algorithmic execution, electronic market-making, collateral management, risk aggregation, and regulatory reporting all depend on intricate systems. The speed at which a bank can adapt those systems to new market conditions, products, or rules directly affects its competitiveness.
When senior leaders commit to scaling AI capability across development, they are effectively betting that the institution can absorb change faster than rivals. Faster deployment of risk controls after a market event, quicker rollout of client-facing tools that exploit new data sources, and more rapid iteration on trading algorithms all translate into economic advantage. Productivity in coding thus becomes a lever on the bank’s agility under uncertainty, not just its cost structure.
Moreover, clients increasingly expect digital experiences on par with the best consumer technology platforms. Meeting those expectations demands continuous enhancement of interfaces, analytics, and integrations with client systems. AI-augmented development offers a way to keep pace with that escalating bar without constantly increasing headcount in an already competitive hiring market for software engineers and data scientists.
Medium-term implications for talent and organisation
One of the less discussed consequences of AI-enabled productivity is its impact on the skill profile of engineering teams. As low-level coding becomes easier, the premium shifts further toward system design, domain understanding, and the ability to translate fuzzy business needs into robust technical specifications. In a bank, that means developers who understand derivatives, collateral flows, risk methodologies, or regulatory regimes become even more valuable.
Over time, roles may polarise. Some engineers will specialise in orchestrating AI assistants across complex build pipelines; others will deepen as domain specialists who ensure that what the machines generate makes sense in the context of trading strategies, risk policies, or legal constraints. For leadership, the challenge is to redesign career paths, training programmes, and incentive structures to reflect this new division of labour, rather than treating AI tools as a simple plug-in to existing workflows.
Organisationally, the ability to produce more with the same number of coders may encourage a shift toward smaller, cross-functional teams owning end-to-end services. If each team can deliver more features per unit time, the overhead of coordinating large monolithic programmes may become less attractive. Instead, modular architectures with well-defined interfaces, owned by empowered teams, could flourish. AI assistance then becomes a force multiplier in an already agile organisational design.
Looking ahead: from pilots to structural change
For now, much of the reported 20 to 40 percent productivity improvement exists in the realm of pilots, early adopters, and selected teams. The hard work lies in translating scattered successes into structural change: standardising tools, integrating them into secure enterprise environments, training thousands of developers, and establishing governance frameworks that satisfy internal risk functions and external regulators.
That journey will not be linear. Some teams will experience impressive leaps; others will find that legacy constraints, regulatory requirements, or local culture blunt the impact of AI tools. The aggregate productivity number will likely fluctuate as the organisation learns where AI adds real value and where it introduces more risk than benefit. Metrics will need to evolve beyond simplistic measures of speed to include stability, incident rates, and client satisfaction.
Yet the direction of travel is clear. When a senior executive at a major bank speaks publicly about the ability to produce significantly more software output with stable headcount, it signals an institutional commitment to weaving AI into the core of technology operations 1. The specific percentages may be debated, but the strategic bet is that software, already central to modern finance, will be increasingly co-written by machines, and that the firms that learn to harness that collaboration safely and effectively will shape the competitive landscape.
In that future, productivity gains are not just about doing the same things faster. They are about enabling new kinds of systems, richer analytics, and more responsive products that would have been prohibitively expensive to build with traditional tooling. For Goldman Sachs and its peers, the real test will be whether AI-accelerated coding translates into better-managed risks, deeper client relationships, and resilient infrastructure, rather than simply a busier release calendar. The stakes extend beyond internal efficiency to the functioning of the financial system itself.
References
1 “Goldman’s AI Expectations”, LinkedIn post by Sonali Basak.
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
