The CEOle squirrel holding an hourglass as its sand pours into a row of empty meeting chairs

Twelve hours a week is not a return

If cutting a role is the most common AI business case, saving hours is a close second. I've sat through a lot of AI demos (enough of them that I should be earning loyalty points), and most of them are the Instagram version of AI: it's magical, it's going to solve all your problems, it's going to save so much time. Most of them don't even give you a number, they just promise the world. The ones that do usually say something like "this will save your team twelve hours a week," and I actually believe that one, because the hours are real and nobody's making them up.

The problem is what happens next, because the time gets eaten up in a hundred little ways and nothing on the financial statement moves.

Why everyone is happy and nothing changed

This one confused me for a while. Ask around and most people using AI will tell you it's helping them, and at the same time PwC asked 4,454 CEOs for its January 2026 CEO survey and 56% of them said AI hadn't brought them more revenue or lower costs. Both can be true, because feeling like it helps and the business actually changing are two different things.

Nobody's lying on the survey, but ask the CFO what the hell actually changed and the room gets real quiet.

The sentence test

Graeme Crouch on our team uses a test that I've now stolen for every AI conversation I'm in. Your use case has to finish one of these sentences:

  • We'll be able to offer something we can't offer today
  • We'll be faster at something specific
  • We'll be more consistent at something
  • We'll absorb more volume without adding people
  • We'll free a specific person to do a specific thing

"We'll save twelve hours a week" doesn't finish any of them. That's just repeating what the feature does. If you can't say what those hours are for, it's still just a demo.

I get asked "what's the ROI on Copilot?" all the time, and it's the wrong question. First, which Copilot, because Microsoft calls a lot of things Copilot (I'm pretty sure the coffee maker is next). But mostly it's wrong because you're starting with the solution in mind. What's the problem you're trying to solve? Figure out what it's good for and what it's not good for, pilot it, test it, and see if it's actually better. If so, awesome, keep improving it, and if not, don't. You don't have to use AI for everything. Sometimes the answer is automation, sometimes it's plain old code that does the same thing every time, and sometimes it's still a human, because AI may be intelligent but it is not wise.

We did this about a hundred times

I can't really point fingers here. Back in the MSP days we rolled out tools and process changes all the time, and we didn't have the operational maturity to know if any of it worked. We didn't know what good looked like or how we were going to measure it, so we made a lot of improvements without knowing if they paid off. We probably did it 100 times, which is not good, and I'm rounding down.

That was before AI, and AI isn't any different. This is more of a business thing than an AI thing. If you're going to change a process, how do you know it'll work, and how do you know it's better than the last one? Sometimes you can measure it and sometimes you can't, but either way it comes down to planning ahead of time.

Where my hours actually go

Since I'm asking you, it's only fair I answer it too. Some of the time AI saves me goes into what I call gray space, which is just time to think. As a CEO part of my job is to wonder about the future and how we're going to get there, and that's hard to make room for. When I'm in the car I now brainstorm with ChatGPT (I named it Scout), and when I get home it reminds me what the thing was and can add it to my Microsoft To Do list. It could send the email too, but I usually don't let it, because I don't like how any of the AI writes them. With my ADHD, a brainstorm partner in the car is the improvement I actually needed, no agent required. People are talking way too much about agentic right now, in my opinion.

And to be fair, at TruNova we don't really measure before and after on everything either, because that adds a lot of extra time. What we do is be careful about what we change. Our CFO uses Claude to validate, check and correlate things while she works, so she doesn't have to review every spreadsheet by hand. She won't use it for the month-end close, because AI isn't deterministic and accounting should be. And sometimes she spends two hours arguing with Claude to get it working, so there are definitely pros and cons. AI is not magical, at least not yet.

Write three numbers down first

The fix sounds obvious and almost everybody skips it, me included. Before you turn anything on, declare a metric (a number you can count, not a dollar figure) and write down where it is today. It doesn't have to be a big study.

So instead of "we'll improve invoice processing," write something like this: today we process this many invoices a month, this percentage come back for correction, and it takes this many days to close the month. Check them again in ninety days. You either moved them or you didn't, and both of those tell you something.

I've sat in meetings where everyone felt good about AI and nobody could say what changed, and I'd rather have a boring page with three numbers on it.

Write the three numbers down first, and then go save all the hours you want (just don't spend them all in meetings about AI).

Follow along with the RSS feed, or talk to me about it on LinkedIn.