Bank of America brought in 1,000 consultants to fix what Mythos found. Count your fixing hours first

An AI that gets the first pass on routine work is only half of an AI-first company. The other half is the people who finish what it hands back. That second half is the one most owners forget to plan. If you want the full picture of how the two halves fit together, see the AI-first company guide for owners. This post covers what changed this week.

What the bank said

On Bank of America’s third-quarter earnings call on October 14, CEO Brian Moynihan said the bank had added 1,000 extra staff to work on remediation tied to its Mythos implementation, “not in the headcount but in consultants.” Mythos is Anthropic’s model for finding software flaws, according to The Register. The Register also reports that Moynihan said the workers were hired in the second quarter and are still helping today.

He said a “couple hundred million plus” of third-quarter expense went to two things together: anti-money-laundering remediation and the Mythos remediation. He did not say how much went to each, as The Register also points out. He said he expects the Mythos remediation costs to run down.

On the same call he gave the bank’s wider AI picture. All of these are company-reported:

• About 130 AI projects are in place, and he said all of them have a payback.
• He said the benefits run two to three times the spend over time.
• He expects benefits to grow from $250 million to $300 million this year to $700 million next year.
• Headcount is down about 3,000 from a year ago.

So the same company says AI is paying back in one place and that Mythos remediation brought in outside staff in another. Both can be true at once. That is the point.

Finding got cheap. Fixing did not.

A model that reviews code, contracts, invoices or support tickets can go through a backlog in an afternoon. Every issue it flags still needs a person to read it, decide what to do, make the change and check it. The model sped up the finding. It did not speed up the deciding.

A large bank can rent a thousand people when the pile gets too big. A business of one cannot. If you point a finder at five years of files without counting first, you get a list you cannot clear. Then one of two things happens. The list goes stale, or your best people drop their real work to chase it.

A test you can run this week: the fix-queue sum

Before you run any AI review over a backlog, do this sum on paper. The numbers below are made up to show the method. They are not from a client.

1. Sample first. Run the review on a small slice, such as one month of invoices or one folder of contracts. Say it flags 30 items.
2. Time the fixes. Clear ten of those items yourself. Write down the minutes each took, including checking. Say the average is 15 minutes.
3. Scale up. The full backlog is 20 times the slice, so expect about 600 flags. At 15 minutes each, that is 150 hours.
4. Compare with real capacity. Say the person who would do this work has 6 free hours a week. 150 divided by 6 is 25 weeks.
5. Decide. Twenty-five weeks is half a year of one person’s spare time. That is too long to leave a known list open.

Now you have a number to make a decision with, before you have a list you cannot finish.

A decision rule

Run the full review only when you can clear what it finds within about four weeks with the people you have. If you cannot:

• Narrow the scope. Review only the newest work, or only the clients with the most at stake.
• Sort before you fix. Have the model rank its flags by impact, and have a person check the ranking. Then work from the top only.
• Batch the boring fixes. If many flags are the same kind of problem, fix the pattern once instead of item by item.
• Buy finishing time on purpose. If you bring in outside help, set an end date and a fixed scope. The bank, for its part, said it expects those costs to run down.

The four-week line is a rule of thumb, not a standard. Pick your own, but pick it before you press run.

What to do this week

• List every AI review or audit you have running or planned.
• For one of them, run the fix-queue sum on a small sample.
• Write down who finishes the flags and how many hours a week they really have.
• If the sum comes out longer than your line, narrow the scope before the full run.

Counting the finishing work is the part of a build owners most often skip. A consult block starts with who decides and where the hard limits are, then the one or two workflows where AI saves the most time or money.