AI Productivity Gains: What Bank of America Actually Measures

September 2026: the number the bank puts its name to

On 24 September 2026, at a BofA Securities conference, Bank of America co-president Jim DeMare used two words that carry more weight than an entire slide: "clearly identifiable".

He applied them to a narrow perimeter: the productivity of people who write code. AI lifts it by roughly 15-20% across most industries, he explained, and that is the gain the bank measures across its own 20,000 developers, now working with coding agents. The account comes from CIO Dive^[1].

The debate over AI productivity gains rarely lands on a figure a bank's leadership will commit to in writing. The rest of the bank's AI portfolio, meanwhile, resists that kind of clean measurement. The gap between the two registers deserves attention.

That asymmetry is the heart of the story.

The question circulating in the room, as relayed by DeMare himself, ran roughly like this: will the returns arrive fast enough? The pressure is rising alongside the spending.

The original idea: put AI where the work was already measured

Bank of America's interesting decision concerns the point of application, not the budget.

Technology units come first for measurable returns, DeMare indicated, and the reason is structural. People who write software work inside an environment that has counted everything for years: commits, tickets, review cycles, release times, defects per thousand lines. The measurement predates AI by a decade.

The before/after already exists, archived, granular, with a stable denominator. The agent enters an instrumented perimeter and its effect surfaces immediately. Finance sees it at the first close.

The same technology, dropped into a poorly tracked process, produces a different experience: the sense of moving faster, with no time series to confirm it. The return may well exist, and it stays invisible.

Here is the first lesson worth copying: instrumentation precedes adoption. Measure the process before introducing an agent and you get a number; introduce the agent and then go looking for the number and you get an anecdote.

The numbers, one by one, with the source attached

CEO Brian Moynihan laid out the economics the week before, at a Barclays conference, as Banking Dive^[2] reports.

• roughly 140 AI use cases moved into production
• stated cost: $400 million
• stated benefit: $800 million
• AI spending budget expected to double next year
• 20,000 developers with coding agents, +15-20% productivity
• Erica: a workload equivalent to about 11,000 people

The two-to-one ratio of benefit to cost is the figure that will circulate through boardrooms. It needs one caveat: that 2× comes from the bank itself, in an investor presentation. An attentive board will ask how it was calculated.

The 15-20% on code, by contrast, is what DeMare qualifies as clearly identifiable. The difference in language between the two items is deliberate, and it deserves to be respected.

A stated figure and a verified figure carry different weight.

Scale matters too: $3.5 trillion in assets. On that base, $400 million of AI spending is a rounding error on the income statement, and the announced doubling becomes a bet with little financial risk. Few organisations enjoy that margin for error.

Erica changes jobs: from customer to colleague

Erica launched in 2018 as a virtual assistant for the bank's retail customers. The move described on 24 September 2026 shifts it to a different front: employee self-service.

DeMare spoke of a considerable drop in requests to the internal help desk. The total workload Erica handles, the bank states, equals the work of about 11,000 people. The number describes volume absorbed, not headcount saved.

Reuse counts for more than novelty. Bank of America set aside the idea of buying a new platform for internal support and backed a system already seasoned on millions of conversations, with its own intent model, its own logs, its own containment metrics.

It is the pattern this desk keeps finding in the most solid cases: organisations that reuse their own assets scale, organisations that buy the turnkey solution stay stuck at the pilot. An assistant already trained on the company's own language starts with an advantage a vendor demo will struggle to close. The advantage lies in the historical data, before the model.

The friction: four bank leaders in five see little

This is where the story becomes genuinely useful.

Accenture found that 20% of banking executives see widespread, durable value from AI initiatives, with everyone else struggling to scale. The figure appears in the same CIO Dive piece.

Translated: four leaders in five describe a picture different from the one coming out of Charlotte. Bank of America sits in the minority, and it gets there by presenting a single line item as certain.

So even in the best case, the safe perimeter stays narrow. Roughly 140 use cases in production, a stated aggregate benefit, and one line the co-president signs off as identifiable: software development.

That is the case's implicit correction: the budget doubles before the return is broadly proven. A $3.5 trillion bank can carry that bet; a small or mid-sized company, hardly. The useful lesson runs the other way: start from the perimeter where the number already exists, widen afterwards.

Headcount falling, layoffs ruled out

Moynihan added the most delicate figure: headcount has gone from roughly 213,000 employees at the start of 2026 to 209,000, with a voluntary attrition rate near 8.5%. The CEO ruled out layoffs.

The arithmetic is worth doing. An 8.5% rate on 213,000 people produces roughly 18,000 departures in a year, while the net decline is 4,000. The difference tells a story of still-plentiful hiring, with partial backfilling of the roles that open up.

For a board, the point is this: AI reaches the income statement through the pace of rehiring, before it reaches it through a savings line.

The mechanism is slow, quiet and reversible. Anyone expecting an immediate cut will misread the case.

For a manager the meaning shifts again. The work the agent removes is the repetitive ticket, the reset request, the hunt for an internal procedure: activities that eat whole days and build little expertise. That is where the value sits for the people who stay.

What you can take away from this story

The playbook comes down to four moves, repeatable at any scale.

1. Pick a process that already produces time-series data: tickets, cycle times, error rates.
2. Fix the "before" on three months of history, with a clear denominator.
3. Reuse an existing internal system before considering a purchase.
4. State which gain is identifiable and which remains an estimate.

A small-business founder reads this as permission to start small: the development perimeter, or internal support, is worth a pilot of a few weeks. The 15-20% on code is an industry range, applicable to a team of eight as much as to a team of twenty thousand.

A CTO takes away a technical spec: the agent goes where telemetry exists, and telemetry is built first. A board takes away a sterner yardstick, with a precise question to put to every presentation: which part of the benefit is identifiable, and which part remains an estimate?

The open question stands, and it holds for any organisation: how many of the processes you would like to hand to an agent produce a number today that you could compare six months from now?

The answer decides the return long before the choice of model does.