Banks Start Reporting AI Returns, but Proof Remains Thin

Large banks are moving beyond reporting how much they use AI to showing what they gain from it. Bank of America's chief executive says the bank spends $400 million on its AI programme and generates $800 million in benefits. Société Générale has included up to €600 million in AI-related savings in its 2029 plan. Yet an industry index finds that only a small fraction of the roughly 1,100 AI use cases deployed by banks have demonstrated specific financial gains. Banks are becoming more open about AI's financial impact, but clear evidence of returns remains limited.

That matters for anyone judging whether bank AI spending is paying off, because the numbers are mostly self-reported, use different definitions, and in some cases are not tied to a stated time period.

What banks are now reporting

According to PYMNTS, Bank of America CEO Brian Moynihan said at a Barclays conference on September 14 that the bank has 130 to 140 AI projects "at a cost of $400 million, generating a benefit of $800 million," without specifying the period. The bank plans to double its AI expense budget next year. PYMNTS adds that it spends $4 billion a year on new technology initiatives out of a $13.5 billion technology budget, that its Erica assistant had 24.6 million active users in the second quarter, and that a generative AI tool used by more than 18,000 customer service staff shortens average calls by nearly a minute.

Other banks have put AI into forward-looking financial targets. PYMNTS reports that Société Générale expects about 1.9 billion euros in gross savings from 2026 to 2029, with AI accounting for up to 600 million euros, and a cost-to-income ratio below 55% in 2029. Commerzbank plans to invest about 600 million euros in AI between 2026 and 2030, and CEO Bettina Orlopp has said AI should save about 350 million euros a year by 2030, with about 500 million euros in yearly AI value expected from that year.

These are plans and management claims, not audited outcomes. The Société Générale and Commerzbank figures in particular are projections.

The independent view: broad adoption, narrow proof

Evident Insights, which publishes an annual AI index of the 50 largest banks, offers a more cautious reading in a preview of its 2026 index published on October 1. It reports that about 1,100 use cases with some impact were rolled out over five years, but only 1% show a specific financial benefit, and the top ten banks account for half of that group.

The same preview says the gap between the first- and tenth-ranked banks has narrowed by 25% compared with last year, and that this year's twentieth-ranked bank scores as high as last year's tenth. Evident's interpretation is that differentiation now depends less on the volume of AI activity, such as team size or number of use cases, and more on how much AI changes the business. At each of the top three banks, it says, more than 80% of engineers and more than 75% of staff use AI tools or copilots regularly.

Reconciling the Bank of America numbers

The two sources describe the same bank slightly differently. Evident's preview says Bank of America has 130 generative AI use cases in production, up from 114 in July, and describes returns as the CEO characterising annual value on a $400 million investment, with a reported figure of about 0.7% of revenue. PYMNTS quotes a cost and benefit pair and a range of 130 to 140 projects. Both agree on the $400 million figure, but neither clearly states the time basis, and "projects" and "use cases" may not be the same unit, so comparisons across banks should be made carefully.

Evident reports that Bank of America is the 13th bank in its index to disclose actual or projected financial returns from AI, up from eight banks a year earlier. It also reports that Lloyds generated approximately $133 million in AI-driven value in 2026, equivalent to about 0.5% of its 2025 revenue. Meanwhile, Commonwealth Bank of Australia expects to break even on its AI investments next year, provided it meets its targets.

Agents and the deposit question

The same preview flags a risk that sits on the other side of the ledger. It cites Apollo economist Torsten Slok warning that AI agents could move customer cash toward higher-yielding accounts, eroding the cheap deposit funding banks depend on and, in extreme cases, contributing to a run. It also cites an Experian survey in which 54% of respondents were comfortable with agents applying for credit on their behalf, while only 5% would grant full autonomy and 75% would be more comfortable with AI tied to a trusted institution.

Banks are responding on two fronts. Evident reports that HSBC has launched HSBCnio, which lets corporate clients connect their own AI agents to banking data such as balances and payments, and that JPMorgan Chase is building a product called Smart Cash that automatically moves funds between checking and higher-yield investments. Lloyds' chief financial officer, William Chalmers, is quoted as saying the bank would compete on value, including pricing, security and AI-enabled rewards.

The strategic logic is straightforward. If customers' agents shop for yield, banks that expose their own data to those agents, or that automate the optimisation themselves, may retain relationships that would otherwise erode. Whether that works is untested, and the Apollo scenario is a warning rather than an observed trend.

Why the measurement problem persists

PYMNTS' wider survey of 60 senior technology executives at US companies with at least $1 billion in revenue offers context from outside banking: 90% to 100% report some return on AI spending, but only 5% to 10% report full payback, and at least half of firms in every industry group expect full payback to take five to six years. Three-quarters of CFOs measure generative AI returns by improved margins. That survey is small and covers large American firms in general, so it should not be read as a measure of banks specifically.

Evident quotes Wells Fargo's Saul Van Beurden saying the bank needs to see that use cases are "the highest-impact use cases in terms of ROI." That reflects the underlying difficulty: copilots save time that is hard to convert into cost, and revenue gains are hard to attribute to a single tool.

What to watch next

• The full index. Evident's complete 2026 ranking and its roundtable on October 13 should show which banks it judges to be converting AI activity into results.

• Consistency of disclosure. Whether more banks adopt comparable metrics, such as value as a share of revenue, or continue to report in incompatible ways.

• Fulfilment of targets. Société Générale and Commerzbank's multi-year savings goals will be tested in annual results.

• Deposit behaviour. Any evidence that agent-driven switching is moving balances would turn a theoretical concern into a funding issue.

For now, the strongest claim the evidence supports is that banks are getting better at saying what AI is worth, not that the worth has been proven across their portfolios.