Read time: 4 minutes

Leader, Leader, welcome back.

A century ago, a French village farrier spent his days shoeing horses.
Then cars arrived. He didn't fight them. He became a mechanic, and so did his son and his grandson.

Same job, new tool: getting people where they need to go.

Finance is at that moment now.
AI writes the first draft; you check it, fix it and sign it.
This week: how much time that checking really takes, and a prompt that cuts it.

What happened: Datarails surveyed 270 CFOs and finance leaders at US companies with $100m+ revenue. Every one of them uses AI in finance.
On average, their teams spend 26% of the working week verifying or correcting what it produces.

Trust is low where it counts.
Only 5% would let AI produce a board-ready report without human review.
Only 4% trust it with the month-end close.

Why it matters: Look at why the checking takes so long.
65% say their biggest frustration is confident answers built on the wrong data.
56% say colleagues got materially different answers from the same prompt and the same data.
And just 4% have a single source of truth for their numbers.

The jobs news is better: 60% are moving staff to higher-value work, and only 3% are cutting headcount because of AI. Same people, new tools.

My take: That 26% isn't the cost of AI. It's the cost of using it badly.
If 5 people prompt 5 different ways, you're checking five different answers.
A shared, structured prompt for recurring work is the cheapest control you'll ever put in place. (Worth knowing: Datarails sells finance AI software, so read it with that in mind.)

What happened: OpenAI will hide an invisible pattern in the word choices of text that ChatGPT and Codex write for EU users, on every plan, over the coming weeks. It's called textGrain, and it's there because the EU AI Act wants machine-written text to be identifiable. Only approved researchers get the detector.

Why it matters: OpenAI published where it breaks.
Swap 10% of the words and detection drops from about 92% to 66%.
Swap 25% and it falls to 17%. OpenAI itself says a watermark can't prove who wrote something or whether it's right.

My take: Worried about being "caught" using AI on a board pack? Don't be. Hoping a detector will catch a junior's unchecked AI work? Don't count on it.

Detection was never the control in finance. Sign-off is. Your AI policy should say who reviews AI drafts and what they check, not whether AI was used.

Both stories land in the same place: AI drafts, you check.
So let's make the checking faster.

Most people prompt "write the Q3 board commentary" and get back something polished, generic and nothing like them.
The board can tell.

Flip it.
Write the prompt as close as possible to what you actually want to say: what moved, why you think it moved, what you're doing about it, and where you're not sure yet. Don't polish. It's a brain dump, not a draft.

Can't face typing? Record a voice note on the walk back from month-end.

Then use this:

That last line does the heavy lifting. You want an editor who pushes back, not one who smooths over a weak explanation.

Now the step everyone skips. Put the AI's version next to a blank page and rewrite it yourself. Within a minute you'll see what it misunderstood and what you actually meant.

Once it works, save it as a shared skill that your whole team uses for commentary. Same structure, same checks, fewer surprises.
That's how you claw back some of that 26%.

My rule: use AI to raise the quality of your thinking, not the volume of your output. Writing 10x faster doesn't mean you have 10x more worth saying.

This is the RACEF structure from my AI for Finance Leaders course.
The full version goes deeper: variance commentary templates, board pack narratives, and prompt chaining that breaks a 3-hour analysis into 4 sequential prompts. Currently available at a discounted rate for a limited time. Check it out here.