Last week I sat in with the finance and operations team at an investment firm. Their tax manager was out of office, and someone still had to calculate tax basis across 16 fund statements.

The person who picked it up did not start with the 16. She gave Claude one prior example, the inputs and the finished answer, so it could learn how the calculation had been done. Then she checked one year for each fund by hand. Only after that did she run the full batch.

About 30 minutes, start to finish.

The sequence is the part worth borrowing: one worked example, one checked run, then scale.

Four days later I was at a research summit, listening to a sell-side research leader explain why his team still keeps AI away from financial forecasting. His reason: you have to be able to "explain the shape of the numbers."

That line fits the tax basis story better than I expected. If you understand the one you checked, and the method is identical for the rest, you understand the batch well enough to trust it. The spot-check is how a non-specialist earns the right to trust the output. It also covers the gap when the person who normally does this work is away.

What if the same pattern works for the calculations your team repeats every month?

Here is how I would try it:

1. Pick one calculation that repeats across many files. Tax basis, fee checks, commission splits, grant budget-to-actual all qualify.
2. Give Claude or ChatGPT one finished example (the inputs and the answer) and say: "Learn how this was calculated. Write the method back to me step by step before you do anything else."
3. Read the method it writes back. Correct anything that is off. Then run it on one new file and check that file by hand.
4. Run the rest. Ask for a final column showing each file's key inputs, so any number that looks odd can be traced in seconds.

Step 3 is where most of the value sits. When the method comes back in plain English, you find out whether it learned the right calculation before 16 files depend on it.

I'm curious what your version would be. Which task on your team has a clear right answer, a pile of files, and one person who dreads it?

Alex