JPMorgan Chase’s AI Coding Scale: 60,000+ Engineers and Counting

JPMorgan Chase’s Global Technology team operates at a scale most enterprises can barely imagine: approximately 60,000–65,000 engineers and technologists as of March 2026, according to Let’s Data Science and NewsBytesApp reporting. This workforce isn’t a passive headcount — it’s the execution engine behind a $17 billion (2024) technology budget projected to climb to roughly $20 billion by 2026. When a firm this size moves on AI coding, the numbers become a case study every engineering leader should dissect. By early 2026, around 40,000 of those engineers had access to AI coding assistants including GitHub Copilot and JPMC’s internal tooling. That’s not a pilot; that’s a platform-level deployment. The mandate became explicit in March 2026 when JPMorgan formalized a dashboard tracking individual GitHub Copilot usage — classifying each engineer as a “light user,” “heavy user,” or “non-user” — and linked those categories to career outcomes. Engineers who lag in AI adoption now face negative performance review impact. The message is unmistakable: AI coding isn’t optional at JPMorgan Chase.

The firm didn’t arrive at this position overnight. JPMorgan began experimenting with GitHub Copilot in 2022, expanded access systematically, and built internal tooling (LLM Suite, PRBuddy) to fill gaps the off-the-shelf tools left open. The Claude Code pilot was preparing to roll out in April 2026, signaling the firm’s continued appetite for agentic coding capabilities beyond autocomplete. By any measurement — headcount reached, tools deployed, budget committed, or governance rigor — JPMorgan’s AI coding program is the de facto benchmark for regulated-industry enterprises.

The Real Velocity Numbers: 10–20% Per Engineer, 30–40% Portfolio-Wide YoY

The headline “30% velocity gain” circulating in press coverage requires careful unpacking because it conflates two separate measurements. At the individual engineer level, JPMorgan’s AI coding assistants have boosted developer efficiency by 10–20% for the tens of thousands of engineers using the tools, according to Emerj and the AIX AI Expert Network. That number is consistent with GitHub’s own research showing similar productivity improvements in controlled experiments. However, the 30–40% figure cited in Tearsheet’s reporting refers to something different and more interesting: AI-attributed benefits growing 30–40% year-over-year across JPMorgan’s entire AI program portfolio, not just coding.

The compounding effect matters enormously. When 40,000+ engineers each gain 10–20% coding efficiency, and those gains compound annually across an expanding base of tools and use cases, the portfolio-wide effect grows substantially faster than any single-tool metric suggests. Code deployments at JPMC increased more than 70% over two years — a number that reflects infrastructure improvements alongside AI coding gains, but which wouldn’t be achievable without the developer velocity gains enabling more frequent, confident releases. Work replanning dropped 20%, meaning engineers spend less time discovering mid-sprint that their initial plans were wrong. The AI coding tools function as a planning accelerant, not just a typing accelerator. For enterprise engineering leaders benchmarking their own programs, the lesson is to measure at both levels: per-engineer efficiency gains AND portfolio-wide year-over-year improvement. The former tells you if your tools work; the latter tells you if your strategy is compounding.

LLM Suite — Why JPMorgan Built Its Own AI Platform Instead of Buying

LLM Suite is JPMorgan’s fully in-house built AI platform, integrating models from both OpenAI and Anthropic in a model-agnostic architecture that allows the firm to swap or combine underlying models without rebuilding integrations. According to The Digital Banker’s reporting, LLM Suite reached 200,000 enterprise users within just 8 months of launch in summer 2024 — a viral, opt-in adoption rate that surprised even the teams who built it. It was named “Innovation of the Year” by American Banker in 2025. The decision to build rather than buy reflects a calculation that most enterprises get wrong: JPMorgan’s compliance, data residency, and IP protection requirements aren’t edge cases that a vendor product can patch around — they’re architectural constraints that require infrastructure-level control. When you’re running over $3 trillion in assets under management and operating under OCC, SEC, and Fed oversight simultaneously, you cannot accept a vendor’s terms of service as your data governance strategy.

The model-agnostic architecture also provides negotiating leverage and risk mitigation. In 2024, when OpenAI changed API pricing and model availability timelines, JPMorgan’s teams could shift inference workloads toward Anthropic models within LLM Suite without retraining users or rebuilding prompts. For coding specifically, this flexibility matters because different models exhibit meaningfully different strengths across languages and task types — GPT-4 class models have historically performed well on Python and JavaScript completions, while Claude models have demonstrated stronger performance on multi-file reasoning and code review tasks. LLM Suite lets JPMC blend these strengths rather than commit to a single vendor’s roadmap. The 450+ AI use cases in development across the enterprise would be impossible to manage through vendor-controlled platforms without creating unmanageable complexity in contracts, compliance reviews, and incident response chains.

GitHub Copilot + PRBuddy: The Two-Tool Stack Behind JPMC’s Developer Productivity

GitHub Copilot forms the foundation of JPMorgan’s AI coding stack, providing inline autocomplete and code generation across the editors JPMC engineers already use. But Copilot alone doesn’t address what slows engineering teams at scale: the review, approval, and documentation burden accumulating around every pull request. JPMorgan addressed this with PRBuddy, an internal AI tool purpose-built for the pull request workflow. PRBuddy doesn’t just summarize diffs — it analyzes the intent of changes, surfaces potential conflicts with existing patterns, and provides context-aware comments that reduce the back-and-forth cycles between authors and reviewers that typically add days to merge timelines. Together, Copilot handles the generation phase while PRBuddy accelerates the review and ship phase — compressing the full software delivery lifecycle rather than optimizing a single stage.

The two-tool approach reflects a sophisticated understanding of where developer time actually goes. GitHub’s own research suggests that code generation represents only a fraction of total developer time — design, review, testing, debugging, and documentation consume the majority. Copilot addresses the generation fraction; PRBuddy addresses the review and documentation fraction. Notably, the tools also serve as an onboarding accelerant: JPMorgan has found that AI coding assistants function as mentors for junior engineers, providing inline guidance that previously required senior engineer time. This compounds the productivity benefit — senior engineers spend less time on synchronous code review while junior engineers improve faster. For enterprises evaluating AI coding ROI, the compounding effect on junior engineer ramp time may ultimately exceed the direct productivity gains on senior developers.

Performance Reviews Tied to AI Adoption — A New Era of Enterprise Mandate

In March 2026, JPMorgan took a step that no major enterprise had formalized so explicitly: linking individual AI tool adoption directly to career outcomes. The dashboard tracks GitHub Copilot usage at the individual level, categorizing engineers as light users, heavy users, or non-users, and this classification feeds into performance reviews. Engineers who lag in adoption risk negative career impact. The move immediately sparked industry debate — is this smart adoption governance or coercive surveillance? The answer depends on the frame. JPMorgan’s engineering leadership argues the mandate reflects a professional competency expectation, not a surveillance initiative. As coding AI becomes standard infrastructure (92% of US developers had adopted some form of AI coding by early 2026 per CIO Dive survey data), refusing to learn it is functionally equivalent to refusing to use version control — a career-limiting choice regardless of whether any dashboard formally penalizes it.

The more interesting strategic question is whether the mandate produces genuine productivity gains or Goodhart’s Law distortions — where engineers optimize for usage metrics rather than actual velocity improvements. Early evidence suggests JPMorgan is aware of this risk. The dashboard tracks usage categories rather than raw completion counts, reducing the incentive to accept low-quality suggestions simply to inflate numbers. The Claude Code pilot rollout planned for April 2026 introduces agentic capabilities — multi-step code generation, automated testing, PR creation — that require substantively different usage patterns than autocomplete. A governance framework built around measuring Copilot autocomplete acceptance rates will need to evolve when the tooling moves from autocomplete to autonomous agents. This is the governance challenge that every enterprise following JPMC’s mandate approach will face in the next 12–18 months.

Security and Compliance: How JPMorgan Scales AI Coding in a Regulated Environment

Financial services firms face AI coding constraints that technology companies don’t: proprietary trading algorithms represent material non-public information, customer data is governed by GLBA and state privacy laws, and certain code artifacts may constitute regulated communications under SEC and FINRA rules. JPMorgan’s approach to these constraints explains why LLM Suite was built in-house and why standard GitHub Copilot deployment requires additional controls for JPMC engineers. The core principle is data locality — code sent to an external AI service for completion should not be retained, logged, or used for model training. GitHub Copilot Business and Enterprise tiers offer these guarantees contractually, but contractual guarantees require audit, and JPMorgan audits vendor compliance as a standard practice rather than trusting attestation.

The security layer extends to output validation. AI-generated code at JPMC passes through the same automated review gates as human-authored code, including static analysis, dependency scanning, and compliance rule checks. This design — AI as a code contributor rather than a code bypasser — is critical for regulated environments. The 450+ AI use cases in development across the JPMC enterprise include applications touching fraud detection, trade surveillance, and regulatory reporting. Code powering those systems must meet the same standards whether a human typed it or a model generated it. For enterprises emulating JPMorgan’s AI coding program, the compliance infrastructure is as important as the productivity tooling: deploying Copilot without data governance controls isn’t a productivity investment — it’s a compliance liability waiting to materialize at the worst possible moment.

JPMorgan vs. Goldman Sachs: Diverging Enterprise AI Coding Strategies

The two dominant Wall Street technology organizations have taken meaningfully different approaches to AI coding in 2025–2026, and the divergence illuminates the strategic choices available to enterprise engineering leaders. JPMorgan’s strategy is mandated broad adoption: standardize on GitHub Copilot, build internal infrastructure (LLM Suite, PRBuddy) to extend and control it, track individual usage via formal dashboards, and tie adoption to performance reviews. The goal is to raise the baseline velocity of every engineer in the organization, with compliance and governance built in at the platform level. Goldman Sachs, with approximately 12,000 developers versus JPMC’s 60,000+, is pursuing a different bet: autonomous agents over augmentation. Goldman piloted Devin — Cognition’s autonomous coder — for its developers in 2025, with the firm projecting 3–4x productivity gains from AI coding agents versus previous AI tools, per Lucidate and CNBC reporting.

The Goldman approach reflects a higher-risk, higher-reward thesis: if autonomous agents succeed in complex financial systems, the productivity ceiling far exceeds what autocomplete assistants can reach. If they fail or plateau in regulated production environments, Goldman will have invested significant governance effort in a technology that underdelivers. JPMorgan’s approach is more conservative: demonstrate compounding gains from tools that work reliably at scale before adopting higher-autonomy systems. The April 2026 Claude Code pilot at JPMC represents the firm testing a middle path — agentic capabilities with human oversight checkpoints — rather than committing to fully autonomous deployment. For enterprise engineering leaders choosing between these strategies, the key variable is organizational risk tolerance and regulatory exposure: firms in heavily regulated industries have stronger reasons to follow JPMC’s measured augmentation approach before betting on autonomous agents.

Lessons for Enterprise Engineering Leaders: What You Can Replicate from JPMC’s Playbook

JPMorgan’s AI coding program offers eight transferable lessons for enterprise engineering leaders regardless of industry, organization size, or existing AI maturity. First, build model-agnostic infrastructure before committing to any single vendor. LLM Suite’s architecture gives JPMorgan flexibility that firms locked into single-vendor API contracts don’t have — and that flexibility pays dividends every time model pricing or availability changes. Second, measure at two levels simultaneously: per-engineer tool effectiveness AND portfolio-wide business impact compounding year-over-year. Teams that only measure the former miss the strategic picture; teams that only measure the latter can’t diagnose what’s working. Third, close the review bottleneck explicitly. Copilot-style autocomplete addresses code generation but not the PR review cycle where velocity gains evaporate. A PRBuddy equivalent is often the higher-ROI investment for teams already using Copilot. Fourth, treat adoption governance as an instrumentation problem — the same discipline engineering teams apply to system observability applied to tool adoption.

Fifth, build data governance infrastructure before scaling. The compliance work JPMC did before deploying AI coding tools at scale is what prevents the costly rollbacks and regulatory conversations that follow poorly governed deployments. Sixth, use AI tools as junior engineer accelerators, not just senior productivity multipliers — the mentoring function compounds across the organization over time. Seventh, resist proof-of-concept hell: JPMorgan’s leadership has explicitly warned against the trap of running dozens of AI pilots that never reach production. The discipline to move pilots to production, or kill them cleanly, separates firms generating real ROI from those running perpetual experiments with impressive slide decks. Eighth, anticipate the mandate question. As AI coding becomes standard professional competency, the question isn’t whether to require adoption — it’s how to measure meaningful adoption versus metric gaming. Building that measurement framework before rolling out performance-linked requirements prevents the Goodhart’s Law distortions that undermine genuine productivity culture.

FAQ

How many engineers does JPMorgan Chase have working with AI coding tools? As of early 2026, approximately 40,000 of JPMorgan Chase’s 60,000–65,000 engineers have access to AI coding assistants including GitHub Copilot and JPMC’s internal tooling. The firm has the largest known enterprise AI coding deployment in financial services.

What AI coding tools does JPMorgan Chase use? JPMorgan uses GitHub Copilot for inline code generation, an internal tool called PRBuddy for pull request workflows, and LLM Suite — a proprietary in-house platform integrating OpenAI and Anthropic models in a model-agnostic architecture. A Claude Code pilot was scheduled for April 2026.

Does JPMorgan really tie AI tool usage to performance reviews? Yes. As of March 2026, JPMorgan formally tracks individual GitHub Copilot usage through a dashboard classifying engineers as light users, heavy users, or non-users. These classifications feed into performance reviews, and engineers who lag in adoption risk negative career impact.

What are the actual productivity gains from JPMorgan’s AI coding program? Individual engineers using AI coding assistants have seen efficiency gains of 10–20%. At the portfolio level, AI-attributed benefits have grown 30–40% year-over-year since the program’s inception. Code deployments increased more than 70% over two years, and work replanning dropped 20%.

How is JPMorgan’s AI coding strategy different from Goldman Sachs? JPMorgan focuses on mandated broad adoption of augmentation tools (GitHub Copilot, LLM Suite) with individual performance tracking, targeting baseline velocity gains across all 60,000+ engineers. Goldman Sachs is pursuing autonomous agents like Devin, projecting 3–4x productivity gains — a higher-risk, higher-reward approach for its 12,000-developer organization.