From Sanskriti Khandelwal | Product & Market Analysis
AI Use-Case Counts Are a Vanity Metric: What JPMorgan's 400-Plus Leaves Out
On this page
In April 2024 JPMorgan told shareholders it had more than 400 AI use cases in production. By October 2025 its chief executive was describing the programme a different way: about $2 billion of benefit for about $2 billion of expense. That shift, from how many to how much, is the most useful lesson in AI portfolio management that any large adopter has published, and most enterprises have not copied it.
Key takeaways
- JPMorgan's headline AI number moved from a count to a value. Its letters reported more than 300 use cases in production in 2023 and more than 400 in 2024. By late 2025 the bank's chief executive was citing about $2 billion of benefit against $2 billion of expense.
- Most bank AI announcements carry no outcome at all. Of 173 new AI use cases announced by 50 large banks over 12 months, Evident found that only 30% disclosed any associated impact metric.
- The average value per use case is small, and the average is misleading. About $2 billion spread across more than 400 use cases is at most $5 million each, while HSBC now refuses to prioritise new use cases worth less than $100 million.
- A useful AI portfolio scorecard has six lines, and the count is not one of them. Track realised value against a baseline, full programme cost, value concentration, a funding threshold, time to first value and use cases retired.
This post is written for the head of delivery or operations who owns an AI portfolio and reports on it every quarter. That reader needs a scorecard template more than another argument about AI returns.
The short answer. An AI portfolio review should track six things. Start with realised value measured against a control group or recorded baseline, and the full programme cost. Add the share of value from the top 10 use cases and the minimum value a use case must promise before it is funded. Finish with time to first measured value and how many use cases were retired.
Why AI use-case counts became the default progress metric
A use case is a specific business problem an AI system is applied to, such as flagging a suspicious payment or drafting a credit memo. Counting them is the most common way large enterprises describe AI progress to investors, boards and staff.
The count spread for understandable reasons. In the first years of a programme there is no value to report yet, and leadership still wants a signal that something is happening. A rising count is a credible sign of activity, budget being spent and teams being formed.
The trouble starts when the count outlives that early phase. A programme in its fourth year that still leads with how many use cases it runs is telling you it either has not measured value or does not like what it found.
A count is the cheapest number a programme can produce
Producing a count requires an inventory and nothing else. Nobody has to agree on a baseline, run a control group or reconcile a benefit claim with finance. That cheapness is exactly why it is attractive, and exactly why it carries so little information.
A count also rewards the wrong behaviour. If the number on the slide goes up when a team splits one workflow into three use cases, teams will split workflows. I would not call that dishonest. It is what any metric produces once people are judged on it.
What JPMorgan disclosed about AI use cases, year by year
JPMorgan is the right case to study. It has disclosed more about its AI programme, for longer, than almost any other large company. The sequence below uses only figures the bank or its executives stated publicly.
| Date | Where | Count disclosed | Value disclosed |
|---|---|---|---|
| April 2023 | Shareholder letter | More than 300 in production | None in the letter coverage |
| May 2023 | Investor day, Lori Beer | More than 300, up 34% year on year | Target raised from $1 billion to $1.5 billion of business value by end of 2023 |
| April 2024 | Shareholder letter | More than 400 in production, 2,000-plus AI and ML staff | None reported in coverage |
| September 2024 | Barclays conference, Daniel Pinto | None | Projection raised from up to $1.5 billion to close to $2 billion |
| October 2025 | Bloomberg TV, Jamie Dimon | None | About $2 billion of benefit for $2 billion of expense |
| April 2026 | Shareholder letter | None found in the text we reviewed | None found for JPMorgan's own programme |
Counts and values are as stated by the bank or reported from its letters and events. "None" means the source we read did not contain one, not that the bank has never stated one elsewhere. The 2024 and 2025 value figures are executive statements, not audited line items.
The count stopped leading the story
The 2023 and 2024 shareholder letters led their AI passages with a use-case count, as coverage at the time reported. When we read the web versions of the April 2025 and April 2026 letters, we did not find a use-case count in the text we could review. The AI passages in the 2026 letter are about infrastructure, competition and how clients will use AI.
Meanwhile the value figures kept arriving through other channels. In September 2024, at an industry conference, Daniel Pinto raised his estimate of what JPMorgan's AI use cases could deliver. It went from up to $1.5 billion to close to $2 billion, according to Banking Dive's report of his remarks. He added that "a lot of that is related to prevention of fraud."
That second sentence is the most important thing anyone at the bank has said about the portfolio. It means the value is concentrated, which is what the count conceals.
What the $2 billion does and does not include
In October 2025, Jamie Dimon told Bloomberg TV that "for $2 billion of expense, we have about $2 billion of benefit". He was candid about the limits: "Some we can do in real detail: We did this, we reduced headcount, we saved this time and money. Some you can't: It just improved service."
Read that carefully and it describes a portfolio with two kinds of value. One kind is measured and attributable. The other is real but asserted. The public figure blends them, and no outside reader can see the split.
It also makes a point that most AI program metrics miss entirely. Dimon stated the cost next to the benefit. A value figure with no cost beside it is half a number. The version that breaks even on its own stated terms is a more honest disclosure than a larger gross figure would have been.
Three banks that publish a value next to the count
JPMorgan is not the only large adopter that discloses both numbers, and the others show what a stronger version looks like. Evident tracks AI disclosure at 50 large banks. In September 2024 it reported that 38 of them had announced at least one AI use case in the prior year. Only 21 reported any outcome, and only two, JPMorgan and DBS, specified a total realised dollar return.
DBS measures against a control group
DBS reported economic value from data and AI of SGD 370 million in 2023, according to its annual report as quoted by Evident. The figure rose to S$750 million in 2024 and to about S$1 billion in 2025. The bank's 2025 annual report attributes that to over 2,000 AI models across more than 430 use cases.
The method matters more than the total. DBS derived the figure by comparing outcomes between customers offered an AI-driven solution and those in a control group, as The Business Times reported in February 2026. That is an A/B test, the same discipline a product team uses for a checkout page, applied to an AI portfolio.
Even DBS is signalling the method has limits. In February 2026, chief executive Tan Su Shan said the bank "might still try to capture the economic value (of AI) based on what we've been doing in the past, which is A/B testing". Then came the caveat: "but I suspect there will be a lot more in terms of capacity building". Capacity gains are harder to put in a control group.
HSBC sets a $100 million floor
HSBC took a different route. On 17 June 2026 it expanded its partnership with Google Cloud. It said the deal should enable more than 200 new AI use cases over two years. It also said that only use cases estimated to return more than $100 million, in revenue or efficiency, will be prioritised for delivery, according to The Asian Banker.
This is the most interesting portfolio rule any bank has published. It turns the count from a goal into a by-product. HSBC is not saying it wants 200 use cases. It is saying it will fund the ones that clear a value bar, and 200 is roughly how many it expects to find.
The arithmetic needs a warning attached. JPMorgan's count dates from April 2024 and Pinto's near-$2 billion projection from September 2024. The two numbers are five months apart and one is an estimate. Treat the result as directional. It is still enough to show that "400 use cases" and "$2 billion" describe a portfolio where most entries are small.
What an AI use-case count hides
A count flattens every entry to the same weight. A fraud model worth hundreds of millions and a meeting summariser used by one team both add one to the total. Pinto's remark about fraud suggests JPMorgan's value distribution looks like most portfolios: a handful of large wins carrying a long tail.
The sector data points the same way. In its August 2025 report, Evident found that 68% of bank AI use cases with reported outcomes were concentrated in two areas, front office and IT and security. It also found that 85% of generative AI use cases were restricted to internal functions. Only 9 of 50 banks had documented an agentic use case in production or pilot.
None of that is visible in a count. A board told the bank runs 400 use cases cannot tell whether value is spread evenly, concentrated in one model, or sitting in pilots that never shipped.
Nobody publishes the counting rule
The deeper problem is definitional. DBS reports more than 2,000 models and more than 430 use cases, so its own numbers show one use case can contain several models. Other banks report "use cases in production", "active generative AI use cases" or new use cases announced in a period. These are different units.
When Fintech News Singapore compared the three Singapore banks in October 2026, DBS reported more than 430 use cases while OCBC and UOB each put theirs at more than 300. Without a shared definition, those three numbers cannot be ranked against each other, and I would not try. The same applies inside your company if two divisions count differently.
This is the same failure covered in the forensic reading of AI ROI case studies: a figure without its construction rules is an assertion. The count is simply the least examined example of it.
An AI portfolio scorecard built from real disclosures
The scorecard below borrows one element from each bank that has published something better than a count. None of them publishes all six lines. That gap is the opportunity for an internal review, where you control the data.
| Metric | What it measures | Public precedent | Weak version to avoid |
|---|---|---|---|
| Realised value against a baseline | Value the portfolio produced, measured against a control group or recorded pre-launch baseline | DBS A/B testing against control groups | Projected value restated as achieved |
| Full programme cost and net value | Model, platform, people and vendor spend, set against realised value | Dimon's "$2 billion of expense" | Gross benefit with no cost beside it |
| Value concentration | Share of realised value from the top 10 use cases | Pinto naming fraud as a large share | An average value per use case |
| Funding threshold | Minimum estimated value a new use case must show to be built | HSBC's $100 million floor | Funding anything with a sponsor |
| Time to first measured value | Months from approval to the first value reading that passes the baseline test | No bank in our review publishes this | Months to "go live" |
| Use cases retired | Number killed, merged or replaced in the quarter, with the reason | No bank in our review publishes this | A count that only ever rises |
The last two rows are where an internal review can beat every public disclosure. A portfolio that never retires anything is not being managed, it is being accumulated. If your team does not yet have rules for when to stop, the criteria in the guide to kill criteria for AI projects are a good place to begin.
How to run a quarterly AI portfolio review
The scorecard only helps if a meeting uses it. Here is the shape I would give a quarterly review for a portfolio of 20 or more use cases. It assumes someone owns the inventory and someone in finance will sign off the value line.
Two weeks before. Ask every use-case owner for one value reading with its baseline, one cost figure and a status of scale, hold or retire. Owners who cannot produce a baseline report that fact instead. A missing baseline is a finding, not an administrative gap.
Build the concentration view. Rank use cases by realised value and calculate the share held by the top 10. If one or two use cases carry most of the value, the review should spend most of its time on them. Their risk is the portfolio's risk.
Apply the threshold to new requests. Your floor will not be $100 million, but you need one. Pick a minimum estimated value relative to your programme budget and decline proposals below it unless they serve a stated platform purpose.
End with retirements. Close the meeting by naming the use cases you are stopping and why. Sunk cost is the force working against this, which is examined in the piece on why AI pilots keep getting funded. A quarter with zero retirements in a portfolio of dozens should be questioned.
Who owns this matters too. If a central AI team both builds use cases and grades them, the review will drift toward the count. The count is, after all, the team's output. The tension is set out in the analysis of AI centres of excellence as bottlenecks. Put the value line under finance.
Where this argument is weakest
The case against counts is strong, but it is not complete. Three objections deserve a straight answer.
Counts are a fair measure in the first year
A programme with six months of history has nothing to measure yet. In that window, a count of use cases in production is a reasonable proxy for capability being built. JPMorgan itself led with counts in 2023 and 2024 before it had a stable value figure. The argument here is against counts as the headline in a mature programme, not against counting.
Value figures are self-reported as well
Every value number in this post is the company's own claim. Dimon conceded some of JPMorgan's benefit cannot be measured in detail. DBS's control-group method is described in press coverage, not in a published methodology note we could inspect. Replacing a count with a value figure swaps an uninformative number for an unverifiable one, unless the method is published alongside it.
A value threshold can also be gamed. If HSBC's teams know the bar is $100 million, estimates will cluster just above it. The threshold works only if someone compares realised value with the estimate a year later, and HSBC has not said that it will.
Our reading of the letters is incomplete
The observation that JPMorgan's recent letters do not lead with a count rests on web versions we could only partly review. The bank also publishes AI figures through investor days and interviews too. A count may appear somewhere we did not read. The direction of the shift is clear from the sources in the table. The absence of a count in 2025 and 2026 is a weaker finding.
Frequently asked questions
How many AI use cases does JPMorgan have?
JPMorgan's April 2024 shareholder letter reported more than 400 AI use cases in production, up from more than 300 a year earlier, run by over 2,000 AI and machine learning staff. Later public statements have led with value instead of a count. In October 2025 Jamie Dimon told Bloomberg TV the bank sees about $2 billion of benefit for $2 billion of AI expense.
Is the number of AI use cases a good metric?
It is a fair measure of activity and a poor measure of value. A count does not say how a use case is defined, how much each one earns, or how many were retired. Across 173 new use cases announced by 50 large banks, Evident found only 30% disclosed any impact metric. Use the count as a denominator in your reporting, never as the headline.
How do you measure the value of an AI portfolio?
Measure realised value against a control group or a recorded baseline, then set it against the full programme cost. DBS reports its figure this way, comparing customers offered an AI-driven solution with a control group that was not. Report the total, the share from your top 10 use cases, and the method used, so a reader can check how the number was built.
What metrics should an AI program track?
Track six metrics. They are realised value against a baseline, total programme cost, value concentration in the top 10 use cases and a minimum value threshold for funding. Add time from approval to first measured value and the number of use cases retired each quarter. Adoption figures such as weekly active users belong in the review too, but as a leading indicator rather than proof of return.
How much value does JPMorgan get from AI?
JPMorgan executives have put the figure at about $2 billion. Daniel Pinto raised his projection to close to $2 billion in September 2024, naming fraud prevention as a large contributor. In October 2025 Jamie Dimon said the bank has about $2 billion of benefit for $2 billion of expense. He conceded that some benefits, such as better service, cannot be measured in detail.
Where to start this week
Open your current AI inventory and add one column: realised value with its baseline. Fill it in for your 10 largest use cases only. The number of blank cells is your real portfolio status. It is a more useful slide for your next leadership update than the total count.
Then write down your funding threshold, even if it is provisional. Send it to whoever approves new AI work, and ask that every proposal from now on state its estimated value against it. Banks that already disclose AI value, covered in the comparison of AI adoption in banking and healthcare, got there by forcing that question early.
Related on value measurement
If your portfolio already reports a value figure, test it against the checks in our AI ROI case study forensics before it reaches a board pack.
References
- Markets Media, JPMorgan Chase raises target for value realised from AI, 23 May 2023. Used for the 2023 investor day count, the 34% growth figure and the $1.5 billion target.
- Banking Dive, JPMorgan's Dimon wants more cloud to fuel AI, analytics, 15 April 2024. Used for the 400-plus use cases and 2,000-plus staff in the 2024 letter.
- Banking Dive, JPMorgan Chase rolls out LLM Suite, 12 September 2024. Used for Daniel Pinto's value projection and fraud comment.
- Entrepreneur, JPMorgan Chase is saving $2 billion a year using AI, 8 October 2025, reporting a Bloomberg TV interview. Used for the Dimon quotes.
- Evident, Find me the AI money, 19 September 2024, and Evident AI Outcomes Report 2025, 1 August 2025. Used for disclosure rates across 50 banks and DBS's 2023 figure.
- DBS, 2025 annual report, CIO statement, and Fintech News Singapore, How DBS, OCBC and UOB are betting on AI in 2026, 14 October 2026. Used for DBS, OCBC and UOB counts and the S$1 billion figure.
- FCC Singapore, republishing The Business Times, DBS says AI push is paying off, 9 February 2026. Used for the S$750 million 2024 figure, the control-group method and the Tan Su Shan quote.
- The Asian Banker, HSBC expands Google Cloud AI partnership with $100 million use case threshold, 20 June 2026. Used for the 200 use cases and the $100 million floor.
The weakest part of this source base: every value figure is self-reported by the company. Several reach us through press coverage rather than the primary filing, and no bank publishes the definition it uses to count a use case. Figures are current as of 18 October 2026.
Related reading