Kroger Cuts Jobs for AI as Bank of America Hires to Fix What It Found

On October 12 Kroger, the largest standalone US supermarket chain, told staff it would cut about 3,000 corporate jobs over three years as it expands its use of AI, per Reuters. Two days later Brian Moynihan, chief executive of Bank of America, told CNBC that the bank had added about 1,000 contractors to patch software flaws found by AI models. One company is counting a saving and the other is paying a new bill, and most AI business cases count only the saving.

The business case most companies wrote

The usual case for AI has two parts. On one side, the subscription: a fixed fee per person per month. On the other, the hours the tool removes, counted as roles not refilled or outside work no longer bought. Subtract the fee from the saving and you have a payback period.

That case was reasonable. The fee was flat and predictable, the saving could be counted in headcount, and nothing else in the budget moved because of the tool. Kroger's plan is the case in its cleanest form. Chief executive Greg Foran's memo promises fewer layers, fewer handoffs and less duplicated work, with the savings going to lower prices and store staff.

Read from the buyer's side, the week that this week's Newsletter covered in Anthropic's $16.8 billion quarter and the new limits on AI added four costs the usual case leaves out.

Line one: fixing what the AI finds

The Bank of America contractors joined in the second quarter and were still working in October, alongside about 3,000 of the bank's own staff. Among the models that found the flaws, Moynihan named Mythos, which Anthropic released in April only to a limited group of banks, tech companies and government agencies. On the earnings call he put third-quarter spending at "a couple hundred million dollars plus", a figure that also covers anti-money-laundering work, so it cannot be read as the cost of AI alone.

The tool did what it was bought for: finding flaws before an attacker does. Each flaw it found still had to be fixed by a person, and at Bank of America that meant a thousand extra people for at least two quarters.

Line two: usage that grows with the user

From October 1, Accenture funds $29 per employee per month for Anthropic's business plan, Claude Enterprise. That is according to an internal memo reported by 404 Media, a technology news site. Anything above the allowance must be charged to a project or team budget, and some staff had been running bills of up to $1,000 a month. This is a single report, but it fits what the firm said on October 1, covered in Accenture and Concentrix Show Who Keeps the AI Saving: it is handing AI productivity to its clients, and it now wants the cost of that productivity charged to the client work that earns it.

The price itself is moving too. On October 13 Google made Gemini 4 Argon, its newest model, generally available at an introductory $2 per million input tokens (the units AI services bill by). After November 30 that rate doubles to $4, per Google. A budget built on this month's rate is already out of date.

Line three: the person a rule requires

On October 13 Brazil published a decree that rewrites its rules for customer service lines from November 2. Chatbots may still sort and resolve requests, but they cannot stop a customer from reaching a person, and human agents must be reachable from 8am to 8pm every day, per Tecnoblog, a Brazilian tech news site. The rules now extend to digital banks and payment companies.

A company that planned to run its Brazilian support on software alone now has to pay people to answer as well, in Latin America's largest market, and a decree does not expire the way a backlog does.

Line four: the evidence an auditor will ask for

Also on October 13, the American Institute of Certified Public Accountants, the US accounting profession's standards body, published its first guidance for auditing companies that use AI in financial reporting. The guidance is not binding, but its message is plain: using AI does not shift management's responsibility for the accounts. A finance team that uses AI to prepare its month-end accounts should expect its auditor to ask how those systems are controlled, and someone has to document the answer.

The strongest case against

The best objection is that these are transition costs. Bank of America is clearing a backlog of old flaws, and Moynihan expects the cost to fall as the patching progresses. Accenture's central allowance is $29 a person a month, small next to what a consultant bills a client in an hour. And Kroger's 3,000 salaries, once removed, are a saving that comes back every year.

That is fair, and it marks the limit of this week's evidence. It shows these costs exist; it does not show they outweigh the saving. The bank's figure cannot be separated from its compliance spending, the Accenture memo rests on one outlet, and the Brazilian rule covers one market. What the objection misses is that three of the four costs recur. The decree is permanent, audits come every year, and usage grows with every employee who finds the tool useful. Only the remediation backlog looks like a one-time bill, and at Bank of America it has already lasted two quarters.

Who should be uncomfortable

Finance leads and operating chiefs who approved AI on a two-part case, subscription against headcount. Companies selling fully automated customer service to Brazilian digital banks, whose promise of a support line with no staff stops being legal on November 2. And any large bank that has not yet run these scans on its own code.

The move is to rewrite the case with four costs next to the saving: a usage budget per person and a rule for who pays when a project runs over; a budget for fixing what the tools will find; the human staffing a regulator or a customer requires; and the documentation an auditor will ask for. The things to watch are Brazil's first month under the decree from November 2, the Argon price on December 1, and the department breakdown Kroger has promised "in coming months."

Price the bill next to the saving

Kroger's cut and Bank of America's hiring came out of the same technology in the same week. One plans to remove work the company no longer needs; the other created work the company never planned for. Neither is the whole story of AI on its own. The companies that will look good in 2027 are the ones that priced both this year, and that can show their board the costs next to the saving before an auditor, a regulator or a vendor's invoice does it for them. A business case that only counts the headcount was written for the tool as it was sold, and the bills now arriving are for the tool as it is used.