There is a useful way to think about what JPMorgan Chase is doing with artificial intelligence, and it is not the way most banks talk about it. Most financial institutions frame AI as a productivity tool, a way to shave costs, automate the tedious stuff, and let staff focus on higher-value work. JPMorgan frames it as a structural rebuild of the entire enterprise. That distinction matters, and the $20 billion annual technology budget they are approaching is the clearest signal that they mean it.
The question worth asking when watching this from the outside is what an organization actually looks like when it treats AI adoption as an architectural decision rather than a departmental initiative.
Behind the $20 Billion Number
Before diving into the specific implementations, it is worth grounding the numbers. JPMorgan's technology budget was approximately $17 billion in 2024, with roughly $1.3 billion of that dedicated specifically to advancing AI capabilities. By 2025, the overall technology spend had climbed to $18 billion, and the bank is on a trajectory toward $20 billion annually.
The $1.3 billion AI allocation is the part that gets headlines, but the smarter read is the broader technology infrastructure underneath it. The bank runs approximately 80% of its applications in public or private cloud. It has over 40,000 engineers on payroll. That infrastructure is what makes AI deployment at scale possible, and it did not materialize overnight. JPMorgan has been investing in data infrastructure and machine learning for over a decade, which is why their current AI outputs are measurably ahead of most competitors who are trying to bolt generative AI onto legacy stacks.
Daniel Pinto, the firm's President and COO, has been public about expecting AI use cases to deliver between $1.5 billion and $2 billion in tangible annual business value, a figure tracked and attributed across specific lines of business with the same rigor applied to any financial target. The bank has reported nearly a 60% year-over-year increase in value from AI and machine learning, which suggests the measurement framework is well past the aspirational stage.
COiN: Where the Story Really Begins
The AI initiative that defined JPMorgan's early ambition was COiN, short for Contract Intelligence, a contract review system launched in 2017 that had nothing flashy about it and delivered extraordinary results anyway.
The problem it solved was straightforward but expensive: commercial loan agreements are long, dense, and riddled with clauses that require careful human review before any deal closes. JPMorgan's legal and compliance teams were spending approximately 360,000 hours per year reviewing these documents, the equivalent of a substantial team doing nothing but reading contracts year-round.
COiN uses natural language processing and pattern detection to extract over 150 specific attributes from commercial lending agreements and mergers-and-acquisitions paperwork. It handles messy scans, varied layouts, and the phrasing variations that lawyers introduce across different firms and jurisdictions. The result: it processes in seconds what previously took thousands of hours, and does so with lower error rates than manual review.
The design approach behind COiN is worth examining closely. The team scoped it deliberately: one high-volume, well-defined task with measurable outputs, a model trained on their own historical document corpus, validated against known outcomes before deployment. That methodology became the template for how JPMorgan approaches AI implementation across the organization.
LLM Suite: Putting Generative AI in Front of 200,000 Employees
When JPMorgan announced its internal generative AI platform in 2024, the number that stood out was 200,000. That is how many employees were onboarded to the LLM Suite within eight months of launch. To put that in context, it is roughly two-thirds of their global workforce.
The LLM Suite functions as a compliance-ready, internally governed environment that allows employees to use large language models for their specific job functions without exposing proprietary data or violating regulatory boundaries. Software developers use it for code review and writing unit tests. Investment bankers use it for preparing client presentations and analyzing earnings transcripts. Legal teams have bespoke configurations for contract analysis. Wealth advisors use it to pull research and synthesize portfolio commentary.
The efficiency impact is documented: employees report saving between three and six hours per week. Across 200,000 people, even at the low end of that estimate, the cumulative capacity returned to the organization each week is enormous.
The infrastructure decision behind it is what makes this worth examining. Rather than granting broad access to commercial tools, JPMorgan built a governed platform with role-specific configurations, data segregation, and compliance guardrails baked into the architecture. The rollout used a "learn-by-doing" training model rather than traditional e-learning, which accelerated adoption considerably. Any organization wrestling with the tension between enabling employees and maintaining control will find JPMorgan's approach here instructive.
Fraud Detection: The Business Case That Writes Itself
If you need a single number to justify AI investment to a CFO, JPMorgan's fraud detection results are a reasonable place to start.
The bank's AI-based fraud prevention systems currently prevent approximately $1.5 billion in losses annually. A separate measure puts the annual savings from the fraud prediction system alone at $250 million. The systems detect fraudulent activity approximately 300 times faster than traditional rule-based approaches, and have achieved a 50% reduction in false positives compared to earlier methods.
That false positive number matters as much as the headline fraud prevention figure. In payment fraud, a false positive means a legitimate transaction gets blocked. That creates friction for customers, generates support calls, and in high-value business banking contexts, can damage relationships. Cutting false positives by half carries direct commercial weight as a customer experience improvement, and that matters as much to the business as the fraud prevention headline.
The fraud detection work at JPMorgan has been running for over a decade, and it illustrates something important about how AI compounds over time. The models get better as they ingest more transaction data. The features the models use become more sophisticated as the data science teams refine them. An organization that started seriously investing in machine learning in 2014 now has models trained on years of labeled outcomes that a competitor starting today cannot replicate in twelve months.
DocLLM: Building a Moat Around Document Intelligence
One of JPMorgan's more technically significant projects is DocLLM, a proprietary large language model built specifically for document intelligence across their business workflows.
The challenge DocLLM was designed to solve is one that any financial institution faces: financial services runs on documents. Invoices, onboarding packets, legal agreements, insurance forms, regulatory filings. Financial documents carry complex layouts, tables, embedded images, inconsistent formatting, and contextual information embedded in the spatial relationships between elements on a page as much as in the words themselves.
Standard language models trained on clean text struggle with this. DocLLM is built with layout awareness, meaning it processes both the textual content and the structural layout of a document simultaneously. On standard benchmarks, it outperforms GPT-4 combined with OCR on document understanding tasks.
The applications run across legal review, client onboarding, invoice processing, and contract workflows. The strategic value extends well beyond efficiency. By building this capability internally, JPMorgan owns a proprietary model trained on their specific document types, their specific workflows, and their historical labeled data. That is a technological moat that is genuinely difficult for a competitor to replicate by purchasing an off-the-shelf solution.
Coach AI and the Wealth Management Play
The wealth management division offers one of the clearest examples of AI driving direct revenue impact rather than just cost reduction.
Coach AI is an AI-powered tool that helps wealth advisors access information, surface relevant research, and pull together client-facing materials significantly faster. The reported improvement is that advisors can access information 95% faster than before the tool was deployed. The business outcome: a 20% year-over-year increase in gross sales in the Asset and Wealth Management division between 2023 and 2024.
It would be an overstatement to attribute all of that revenue increase to Coach AI. Markets were favorable during that period, and there are other variables at play. But the directionality is telling. When you make advisors more effective by giving them faster access to better information, they close more business. The efficiency gain translates to commercial output.
Coach AI illustrates something that often gets lost in infrastructure-heavy AI conversations: the tools that give frontline revenue generators faster access to better information tend to have the highest ROI and the shortest path to business impact.
IndexGPT: Building Investment Products at Machine Speed
IndexGPT is a more experimental project and worth including precisely because it shows where JPMorgan is pushing into genuinely new territory.
Traditional thematic investment products, the kind that let investors gain exposure to trends like clean energy or artificial intelligence as a sector, are built through analyst-driven research processes. Analysts identify relevant companies, define the investment thesis, and construct the basket. That process takes time, is resource-intensive, and is inherently limited by what a team of analysts can monitor.
IndexGPT uses GPT-4 combined with a separate natural language processing model that scans news articles continuously. The system generates keyword sets associated with a given investment theme, uses those keywords to identify relevant companies in near real time, and constructs thematic investment baskets accordingly. The speed advantage alone changes what is possible in terms of how quickly the bank can create new products and how dynamically those products can respond to shifting market narratives.
The seriousness with which JPMorgan treats IndexGPT as a proprietary capability, rather than an internal prototype, is evident in how it has been built and positioned within the firm's broader product strategy.
40,000 Engineers, One AI Copilot
JPMorgan's AI investment runs deep into its own engineering organization. The bank has deployed AI coding assistants to over 40,000 of its engineers, and the reported productivity improvement ranges between 10% and 20%.
In absolute terms, that is a significant number. If you have 40,000 engineers and they become 15% more productive on average, you have effectively added the equivalent of 6,000 engineers without hiring anyone. At technology industry salary levels in the United States, that is a multi-billion-dollar annual value figure before you account for time-to-market improvements on new products.
The coding assistant deployment also compounds over time in ways that are easy to underestimate. Faster code review cycles mean bugs get caught earlier. More consistent code quality means fewer incidents in production. Junior engineers with AI assistance close the experience gap with senior engineers faster. The downstream effects on system reliability, security posture, and talent development are real, even if they are harder to quantify than a simple productivity percentage.
Reading Between the Numbers
The natural temptation when reading about JPMorgan's AI program is to extract principles. Invest in data infrastructure. Build governance before deploying tools. Measure business value, not adoption. These things are true, but they are also the kind of advice that gets nodded at in steering committees and then quietly set aside when the pressure to ship something visible arrives. The more useful exercise is to look at the specific architectural decisions behind JPMorgan's results and ask whether your organization has made the equivalent ones.
The data labeling debt is probably larger than you think. JPMorgan's fraud models are exceptional because they have been training on a decade of labeled transaction outcomes at a scale no competitor can replicate today. Every fraud case investigated, every false positive reviewed, every outcome recorded, has fed back into models that keep getting sharper. If your organization has not been systematically labeling and retaining structured outcome data in your highest-value domains, buying better tooling will not close that gap. It closes slowly, through disciplined data practice, or not at all. The more productive question to ask your team is: what outcome data have we been discarding that we cannot get back.
The LLM Suite architecture is more interesting than the LLM Suite itself. Most enterprises that have attempted broad generative AI deployment have run into one of two failure modes: either they locked things down so tightly that adoption stayed in single digits, or they opened access broadly and watched sensitive data flow through commercial APIs in ways that created compliance exposure. JPMorgan avoided both by solving a harder problem upfront. The LLM Suite uses role-based model configurations, meaning a wealth advisor and a software engineer and a legal analyst are not using the same tool with the same prompts and the same data access. Each function got a purpose-built configuration with appropriate data segregation. That is a substantial engineering investment up front, and the payoff is why the deployment hit 200,000 users in eight months instead of stalling at a few thousand power users while the security review dragged on. If your current AI rollout is stalled, the bottleneck almost certainly sits in the governance layer nobody wanted to build first.
Where the value actually lives matters more than the build-versus-buy debate. DocLLM makes this concrete. JPMorgan built it because their document corpus, their specific layout structures, their labeled historical workflows, is the asset. A commercial model trained on generic data does not understand the spatial relationships in a JPMorgan credit agreement the way a model trained on thousands of those specific documents does. The build decision was driven by data strategy, with cost as a secondary consideration. Most organizations reach for off-the-shelf solutions by default and build only when the commercial option clearly fails. JPMorgan inverts that logic in domains where their proprietary data is genuinely differentiated. The exercise worth doing in your own organization is identifying which two or three domains have data assets unique enough that a purpose-built model would outperform anything you could buy, and whether you are investing there or spreading effort across lower-leverage areas.
The 40,000 engineers with coding assistants deserve more scrutiny than the 10-20% productivity headline. That range is almost certainly an average across very different use cases. Senior engineers working on novel, ambiguous problems likely see smaller gains than junior engineers doing well-defined, repetitive tasks. What matters more than the average is what JPMorgan is doing with the reclaimed capacity. Productivity gains from AI tooling tend to disappear into the same backlog they came from unless there is deliberate organizational design around where that capacity gets redirected. The banks and technology organizations that will pull ahead are the ones redeploying that capacity toward harder problems, not simply maintaining the same output with smaller teams.
The Part Worth Borrowing
JPMorgan has structural advantages that are worth naming plainly. The $18 billion annual technology budget. The 40,000 engineers. The decade of clean, labeled financial data at a scale very few organizations can match. A strong strategy alone will not close those gaps.
The methodology transfers even when the budget stays modest. COiN was a focused NLP system built around a specific, high-volume task with measurable outputs. The LLM Suite governance architecture is replicable at smaller scale. The discipline of attributing business value to specific AI investments rather than reporting deployment statistics is a practice available to any organization willing to do the harder measurement work.
JPMorgan reached $20 billion in annual technology investment by spending a decade making carefully scoped bets, measuring which ones actually worked, and compounding the institutional knowledge of what good AI deployment looks like in their specific context. The organizations that will close the gap are the ones building that institutional knowledge now, in their own context, rather than waiting for the tools to get good enough to make up for the absence of it.
