Notes on AI in treasury: what worked, what quietly did not

We spent six months running AI assistants alongside a treasury team at a mid-sized corporate. The goal was not to replace anyone. It was to find out which small, boring tasks could be safely handed over.

What worked

The wins were unglamorous and that is exactly why they mattered.

• Reconciliation notes. Drafting the first explanation for why two balances differ, which a human then confirms or corrects.
• Policy lookups. Answering “are we allowed to do this?” with a link to the paragraph that says so.
• Summaries of counterparty news. A daily digest, with sources, that the team can skim in two minutes.

The pattern behind the wins

Each successful task had three properties: a clear right answer, a human who could check it quickly, and a low cost when it was wrong.

What quietly did not

Anything requiring the assistant to decide, rather than draft, produced results that looked plausible and were occasionally wrong in ways that took longer to find than to do by hand.

A confident wrong answer is more expensive than an honest “I am not sure”.

Our working rules

1. The assistant drafts; a named person approves.
2. Every output links to its sources.
3. Anything touching money movement stays behind a human approval step.
4. We log every suggestion and whether it was accepted, so we can measure usefulness rather than guess.

What we will try next

Cash-flow forecasting is the obvious next candidate, but we are approaching it carefully. A forecast is only useful if the team understands why it says what it says, and that is a higher bar than a reconciliation note.