Read time: 4 minutes

Leader, Leader, Leader….welcome back.

OpenAI published research last week on who actually uses AI at work, broken down by function. Finance came second from bottom.

Most of the coverage treated that as bad news.
I don’t think it is, but the real problem sits somewhere else entirely!

OpenAI's economics team went through more than 800,000 work messages and asked a simple question of each one: is this person doing their own job, or someone else's?

The answer is that a lot of people are doing someone else's.
Once you strip out the generic admin that everyone does, like scheduling and drafting emails, 43.5% of what's left is work that traditionally belongs to another role. Customer experience teams led the field at 77%, with designers and HR close behind.

Finance came in at 40%, second from bottom, ahead of only engineering.

The obvious reading is that finance is slow off the mark, and almost every write-up went with it.
I'd argue nearly the opposite.

When a marketer borrows a task from another team and the output is 80% right, they ship it and nobody notices.

When a finance manager does the same thing, 80% right is a misstatement in a board pack. Being careful about which work you take on isn't a failure of ambition. It's a reasonable response to being the function that signs things off.

The trouble is that the caution has a price, and Gallup research measured it this quarter.
They asked people how many different things they use AI for, then asked whether it had made any real difference to their productivity.

Among those using it for one or two things, 45% said yes.
Among those using it for seven or more, 90% did.
Same tools, twice the reported benefit, purely from breadth.

And the two most common uses in the whole study were writing and search, which also happened to score the lowest of anything measured.

So here's what that actually means.
Your team almost certainly isn't refusing to use AI.
They're using it for the two things that help least, in the one function where the standard advice to experiment and see what happens is genuinely bad advice.

Which is why we don't tell finance teams to take more risk.
We give them somewhere safe to take it.

So the steps for you are:
Pick a month you've already closed and signed off.
You know what the right answer looks like, which means you can check the output instead of trusting it.
Run the work through, compare it against what you filed, and you either come away with a process you can defend to an auditor or you find out it doesn't work on a month that can't hurt you.

That is the whole difference between the 45% group and the 90% group, and nobody has to gamble on a live close to get there.

Finance prompts don't usually fail on the instruction. They fail on the examples.

We teach a five part structure called RACEF:
Role, Action, Context, Examples, Format.
Most people get to four of those on instinct, but almost nobody writes the Examples, and Examples is the one doing the heavy lifting.

Here's a variance commentary prompt the way it usually gets written:

And here it is with the structure applied:

The second version isn't better because it's longer.
It's better because the Examples line shows the model what good looks like in your house, and the Context line stops it inventing a driver it had no way of knowing.

Try it on a month you've already closed and you'll know within ten minutes whether it's usable.

And here's what usually happens next.
Someone gets a prompt like that working, feels quietly pleased with it, and then never writes a second one. That's the 45% group from earlier, and it isn't a motivation problem.
Working out which task to point it at next, and what good actually looks like when you get there, is a harder job than writing the prompt.

This is the problem we get hired to fix.
Most of that happens in live workshops with finance teams, and we've now rolled the same material out as self-paced modules on our own platform, for people who'd rather work through it in their own time.

63 more prompts, all built for the work that eats your month.
Practice that gets marked rather than just watched.
A live tutor when you're stuck, which I've not seen anywhere else at this price. And an hour with me on your own close, to work out where this fits and where it doesn't!

The teams sitting at 90% aren't cleverer than yours.
They just had a method and a running start!

It's £49 for 7 days with the code LOYALTY50 - Don’t miss out.

OpenAI put voice mode into the ChatGPT desktop app on 23 July, and I gave it an afternoon to see what it could do.

This isn't dictation. You talk, it works, and it carries on working while you walk away. It shrinks into a corner window when you switch apps, it handles you cutting it off mid sentence without losing the thread, and it can read what's on your screen and drive Chrome.

The part most people will miss is that you can say "work on these at the same time" out loud and it splits the job across separate agents, so research runs while something else renders. You won't get that by default.

In a single afternoon of talking, it took a made up coffee brand from a folder of files to finished brand images, competitor research and the beginnings of a hosted site.

My honest verdict is that OpenAI has copied Cowork closely, including the parts that don't work, like still having to choose Chat or Work up front rather than one place that works it out for you. On everything else the two are level, and on voice ChatGPT is ahead. So if Cowork is already working for your team, stay where you are. If you haven't committed either way, voice is a good reason to start here.

Either way it's worth ten minutes. You'll notice at some point that you've stopped typing.