Scarcely any metric is currently cited in the board presentations of European banks as frequently as the number of Artificial Intelligence use cases (AI use cases) at JPMorgan Chase. It appears as a benchmark, as a wake-up call, as proof of how far ahead the largest US bank is of the rest of the industry. The most popular version reads: 450 AI use cases in production. The figure is memorable, it is impressive – and in this form it is not accurate. Anyone who draws the wrong conclusions from it optimises the wrong variable.
This is not pedantry. The difference between a use case running in production and a proof of concept (PoC) in the pipeline determines whether a bank realises value or simulates activity. This is precisely where this article begins: it reconstructs what JPMorgan has actually demonstrated, separates the robust figures from the secondary-source folklore, and derives from them what European and German institutions should realistically measure themselves against – and what they should not.
What: Contextualising the much-cited "450 AI use cases" at JPMorgan Chase – separating production from the experimentation pipeline, assessing the reported value contributions
Robust lineage: more than 300 in production (Investor Day, May 2023) → more than 400 in production (Jamie Dimon's shareholder letter, April 2024) → around 450 proofs of concept in the pipeline (May 2025)
Value contribution: Jamie Dimon in October 2025: roughly two billion US dollars of benefit for roughly two billion US dollars of expense
Adoption: LLM Suite – the internal generative AI tool – onboarded to 200,000 users within eight months; around half of 232,000 entitled employees use it actively
Relevance: The distance between JPMorgan and European banks lies in cloud maturity, data architecture and governance – not in the use-case count
The Anatomy of a Number
What JPMorgan actually said
Tracing the primary sources yields a clear but different story from the headline. At the Investor Day in May 2023, then Global Chief Information Officer (CIO) Lori Beer spoke of more than 300 AI use cases in production, spread across risk, marketing, customer experience and fraud prevention. In the shareholder letter of April 2024, Chief Executive Officer Jamie Dimon cited a figure of more than 400 use cases deployed in production. Both statements are covered by official JPMorgan documents and refer explicitly to production status.
The figure 450 originates from a different layer. It appears in May 2025 in a report by the trade publication Tearsheet, which quotes Katie Hainsey, Managing Director for AI, Machine Learning and Data and Analytics at JPMorgan Chase. There, the reference is to 450 proofs of concept in the works – explicitly an experimentation pipeline, not a production inventory. The jump from "450 proofs of concept in the pipeline" to "450 use cases in production" is a compression by secondary aggregators, not a statement by the bank. The accurate formulation reads: more than 400 in production as of early 2024, plus around 450 additional projects in the experimental pipeline as of mid-2025.
This self-description from inside the bank is remarkably candid. It confirms what experienced transformation leaders know: the number of use cases is an output metric, not a proof of value. Even the industry's AI leader concedes a gap between technical possibility and realised business impact.
The value contribution: robust and cautious at once
The value figures are more robust than the use-case count. At the 2023 Investor Day, Lori Beer formulated a raised target of 1.5 billion US dollars of value by year-end, explicitly increased from an earlier one billion. At the 2024 Investor Day, JPMorgan assigned AI use cases a value of one to 1.5 billion US dollars, concentrated in customer personalisation, trading, operational efficiency, fraud management and credit decisioning. Finally, in October 2025, Jamie Dimon refined the balance in a television interview with Bloomberg.
The statement is instructive in two respects. First, it confirms an approximate parity of expense and benefit – not a multiplier, but a balanced ratio at an AI budget of around two billion US dollars within a technology budget of about 18 billion US dollars. Second, Dimon deliberately avoids overstatement: he says "about", calls the effects "just the tip of the iceberg", and at the same time acknowledges workforce consequences. This calculated sobriety is itself a message – it contrasts with the return promises with which AI initiatives are sold elsewhere.
The Real Engine: Platform, Not Project Counting
LLM Suite as the access layer
The strategically underestimated variable is not the number of use cases but the infrastructure that makes them scalable in the first place. JPMorgan's internal generative AI tool, the LLM Suite, launched in the summer of 2024 and reached 200,000 onboarded users within eight months. According to Teresa Heitsenrether, Chief Data and Analytics Officer (CDAO), around 232,000 employees are entitled, and more than half use the tool actively. The American trade outlet American Banker named the solution Innovation of the Year in the generative AI category in 2025.
What is decisive is the architectural idea behind it. The LLM Suite is an abstraction layer that bundles different language models behind a single, governed interface instead of letting each business unit experiment in isolation. Derek Waldron, Chief Analytics Officer, describes the target picture as an "AI hub" for employees. It is precisely this bundling that explains why a high use-case count can emerge at JPMorgan at all: not because 450 teams have built 450 isolated solutions, but because a shared platform lowers the marginal cost of the next use case.
Where the use cases flow
According to the Investor Day disclosures, the focus of the productive use cases lies in fraud prevention, marketing personalisation, trading, operational efficiency, software engineering and credit decisioning. In software engineering, more than 40,000 developers use AI-supported coding assistants according to investor disclosures; the reported productivity gains range, depending on the source, between 10 and 20 per cent. In wealth management, a specialised variant supports advisers with research. The precise percentages of individual effects should be treated with caution throughout – they stem predominantly from secondary sources, not from verified primary transcripts.
This outlines the sober finding. JPMorgan does not have a magical use-case factory but a rare combination: an accessible platform, a consistent data architecture and governance that permits productive deployment without ignoring model risk. The number is the visible result of these preconditions, not their cause.
The European Perspective
The gap is structural, not numerical
For a German savings bank, a Landesbank or a mid-sized institution, the relevant question is not "How many use cases do we have?" but "How many are running in production and show measurable impact on results?". The AI Index 2025 published by Evident Insights ranks JPMorgan first for the fourth consecutive time. More telling than the top position is the distribution: North American banks achieve, on average, around 20 per cent higher scores than European ones, and only ING and UBS appear among the top ten. The lag is therefore not an isolated case but a structural feature of the European banking sector.
The cause lies upstream of the use-case count. Over years and with double-digit billion-dollar amounts, JPMorgan invested in cloud migration, data architecture standardisation and model risk governance before the number of use cases became a meaningful metric. The platform layer that reaches 232,000 non-technical users with low friction, and a governance model led through the CDAO function, are the actual differentiating features. It is precisely these preconditions that many European institutions lack – not the willingness to define use cases.
The regulatory asymmetry
There is also an unequal starting position in governance. JPMorgan operates under the US supervisory framework for model risk, which has been established over years and builds on existing banking practice. European institutions additionally face the staged application of the Regulation on Artificial Intelligence (EU AI Act): for high-risk applications – such as credit decisioning or employment-related systems – tightened obligations apply from August 2026, which must be integrated with the requirements of the General Data Protection Regulation (GDPR) into a consistent framework. This creates an asymmetry: JPMorgan invests governance bandwidth predominantly in deployment, European institutions first in compliance infrastructure. Anyone who does not actively steer this sequence risks the compliance burden permanently delaying scaling.
What German Institutions Should Do Now
The lesson from the JPMorgan case is not "define more use cases". It is considerably more uncomfortable: create the preconditions under which use cases become scalable at all. Four measures are priorities for this:
Immediately: Before a bank counts its AI use cases, it must define what "in production" means – with criteria for model approval, ongoing monitoring and demonstrated impact on results. A use-case list without this precision measures activity, not value. It is precisely this vagueness that turns "450 proofs of concept" into a headline of "450 use cases in production".
By Q4 2026: Instead of promoting isolated solutions per business unit, a shared access layer analogous to the LLM Suite should be prioritised – with central model governance, uniform logging and low-friction access for non-technical employees. The marginal cost of the next use case falls only once this layer is in place. Without it, every use-case initiative remains an individual project with individual costs.
2026 to 2027: The gap of European banks documented by Evident is predominantly an infrastructure finding. Institutions should honestly assess what share of their applications is migration-ready and where the data architecture prevents consistent model provisioning. This stocktake belongs before any ambitious AI roadmap, not behind it.
Before August 2026: The EU AI Act obligations for high-risk applications and the GDPR integration tie up capacity that does not arise in the US in this form. Institutions should decide early which use cases justify the regulatory investment and build an integrated governance framework instead of processing AI Act and GDPR requirements sequentially. Anyone who does not steer the sequence lets compliance dictate strategy.
Risks and Open Questions
JPMorgan's success, too, should not be read without reservations. First, the question of end-to-end impact remains open. Derek Waldron puts it unequivocally internally: an hour saved here and three there raise individual productivity but, in end-to-end processes, often merely shift bottlenecks instead of resolving them. Point productivity gains do not automatically add up to business value.
Second, the comparability of counts across institutions is low. BNP Paribas itself reports more than 750 AI use cases in production – a higher figure than JPMorgan's published order of magnitude. Whether both count the same thing is unclear in the absence of a uniform definition. This precisely underlines the core finding: as long as "use case" is not a defined term, the number is unusable as a benchmark.
Third, the workforce question. Jamie Dimon publicly stated in February 2026 that AI is already changing the workforce, and announced a substantial redeployment of displaced employees. Operational and support functions shrank in 2025, client-facing roles grew. For European institutions with a different labour-law and co-determination reality, this path is not transferable one to one – a point that regularly gets lost in the enthusiasm for productivity figures.
The strategic consequence for German capital-markets and banking actors is thus clearly outlined. The "450 use cases" are neither a myth nor a direct yardstick. They are the visible result of years of investment in platform, data and governance – and their popular reading is shifted by one decisive nuance. Anyone who understands the number as the goal optimises the wrong thing. Anyone who understands the preconditions behind it as the goal begins in the right place. The most expensive misinterpretation would be to copy the output metric and skip the input work.
Christian Schablitzki
Strategy & Management Consultant · Agentic AI expert for financial institutions
Over 20 years in investment banking and derivatives trading, followed by more than 10 years advising financial institutions. Currently a Partner at Infosys Consulting in Germany. Certified in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.
LinkedIn profile →Keep reading – every 14 days in your inbox.
Capital-markets insights, regulatory updates and AI trends. Concise, well-founded, free.
GDPR-compliant. Unsubscribe at any time.