Finance work that has to be right, done by people with no time to do it.

I have spent my career on the wrong side of the close. As an auditor, as a diligence lead, and then as a CFO, the pattern never changed: a small number of people doing a very large amount of careful, repetitive, unrewarding work, and a month-end that arrives before the last one is properly finished.

The uncomfortable part is that almost none of it is judgement. Matching a purchase register against GSTR-2B is not judgement. Chasing a vendor for a missing invoice is not judgement. Retyping the same figure into a third spreadsheet is definitely not judgement. But it all has to be right, because it ends up on a return, in a board pack, or in front of an auditor — and being right is why it stays manual.

Software has been sold to Indian finance teams for twenty years on the promise of fixing this, and mostly it has moved the work rather than removed it. What has actually changed is that a system can now read a document, reason about a ledger, and explain what it did. That is the first time the hard part has been automatable rather than just the data entry.

So the test I hold us to is the one I was held to: can a reviewer open what the agent produced, follow it back to the voucher, and sign it? If not, it is not finished. We would rather the agent say it cannot ground a number than estimate one.

Most people building agents right now are writing prompts. That works for a demo and fails in an audit, because a prompt is an instruction and an instruction is not a guarantee. Finance is the domain where that difference is expensive.

So we build the harness, not the paragraph. The arithmetic is code — amounts, dates, invoice numbers and tax heads are matched by rules that return the same answer on every run. The model is used where judgement is genuinely required, like naming the right ledger out of a messy chart of accounts or explaining why a break exists, and it is given the evidence to do that rather than being told to try harder. Each agent gets its own narrow set of tools, so an agent that should not be able to post a voucher simply cannot.

The other half is that everything is written down. Every agent step, every tool call, every source document, and the reasoning behind a match or a mismatch goes into an audit log. Internally the rule is blunt: if it was not logged, it did not happen. A run can be paused, read, corrected and resumed rather than restarted, because that is what you need when a real person disagrees with the machine halfway through.

None of this is the flashy part of AI. It is the part that decides whether a finance team can actually put it in front of their auditor.