Most of the business owners I talk to have the same relationship with AI right now. They know it matters. They’ve tried a few things. Someone on the team uses ChatGPT for emails. And there’s a nagging sense that they should be doing something bigger with it — a strategy, a rollout, a transformation — without any clear idea of what that would even look like on a Tuesday afternoon.

So they do what sensible people do with a big, vague obligation. They put it off.

I listened to an episode of Peter Diamandis’s Moonshots podcast recently — a conversation with Salim Ismail, who wrote Exponential Organizations — that cut through this better than anything I’ve heard in a while. It’s called The Organizational Singularity and it’s pitched at Fortune 500 boards. It talks about digital twins and agent passports and running a power plant with a tenth of the staff. Big, dramatic stuff.

But buried in the middle of it is a plan that is small, concrete, and almost boringly practical. And I think it’s the right plan for nearly any business, including the five-person ones. Here’s my translation.

Why the big-bang approach fails

Salim opens with a number: more than 80% of AI projects inside companies fail. His explanation is the best I’ve heard. Companies take a workflow that was designed for humans handing things to humans — with all its approvals, check-ins and hand-offs — and they inject AI into it. They automate the bottlenecks instead of removing them.

His analogy: when television arrived, the first thing anyone did was sit radio announcers in front of a camera. It took years to work out that the new medium needed a different show.

That’s exactly what “let’s add AI to our process” produces. A radio announcer on TV. You get a slightly faster version of a process that was never designed for this, and everybody concludes AI doesn’t really work for their business.

The plan: carve off one process

The alternative is not a strategy. It’s an act of surgery. Here’s the version I’ve been giving people since I heard it.

1. Pick one process. Not your most important one. Your most predictable one. Something cookie-cutter that you know exactly how it works and that runs the same way every time. Invoice processing is Salim’s example; for a smaller business it might be onboarding a new client, chasing overdue accounts, writing up the weekly report, quoting a standard job, or turning a site visit into a proposal. The test: could you write down every step on one page? If not, it’s not the first one.

2. Give it to two or three people, and give them ninety days. Not a committee. Not a vendor. A couple of people who are curious, a bit impatient, and close enough to the work to know where the bodies are buried. Their job is to rebuild that one process in an AI-native way — which usually means far fewer steps than the original, with AI doing the reading, drafting, checking and chasing, and a human looking at the result.

3. Copy it, don’t move it. This is the part people skip. The old process keeps running exactly as it is. You are not touching the thing that makes money. You are building a second version alongside it, with a copy of the same inputs.

4. Run both in parallel until you trust the new one. Every week, compare the outputs. Where does the AI version get it wrong? Fix that, run it again. Salim’s rule is to wait until the new version is improving faster than the old one ever could — and then wait a few more weeks anyway. You’ll know when you’re there, because the comparison stops being interesting.

5. Cut over. Then pick the next process. Retire the old version. Take what those two or three people learned, and point them at the next workflow. Then the next one. That’s it. That’s the whole strategy. It doesn’t need a name.

What surprised me about this

Three things stood out that I keep repeating to people.

The hardest step is writing the process down. Salim calls it tacit knowledge: the dozen little things the person who runs the process does without noticing, that aren’t documented anywhere. AI can’t do a step nobody has ever described. In my experience this is where at least half the ninety days goes, and it’s worth every hour, because the business gets a written process out of it whether the AI part works or not.

Simplify before you automate. His advice to a company with too much internal friction: if a process takes ten steps, brute-force it down to three before you bring AI near it. Automating a ten-step approval chain just gets you a very fast ten-step approval chain. Most of the steps exist because humans needed to hand things to other humans. Take those out first and the automation is often trivial.

The people don’t go away, they move up. The version of this that ends up in headlines is “run the company with 20% of the staff.” What happens in practice in a small business is different: the same people stop doing the copy-and-paste part and start doing the judgment part — checking exceptions, fixing the cases that don’t fit, improving the process. If you’re a small firm, that’s not a threat. It’s the thing you never had time for.

There’s one line from the episode I’d put on a wall: building the feature is cheaper than having the meeting about the feature. When execution gets that cheap, the meetings are the expensive part. Most of what a ninety-day sprint like this actually removes is meetings.

If you’re under fifty people, you have an advantage

The big companies in this conversation have to do all of this at the edge — a separate team, a separate entity, board sign-off, because the corporate immune system will kill anything disruptive inside the main organisation. Salim has seen the inside of half the Fortune 500 and says he’s never seen any other approach work.

Then he says something almost in passing: if you’re under fifty people, you don’t need any of that. You’re on first-name terms with everyone. You can just do it.

I’d underline that. A five-, fifteen-, or thirty-person business can run this whole loop inside a quarter, with no separate entity and no change-management programme. The founder picks the process. Two people rebuild it. Everyone sees it working in parallel and the trust builds by itself. That’s a structural advantage over every big competitor you have, and it expires the moment they finish their edge programme.

Where to actually start this week

If you want the single most useful hour you can spend on this, it’s not choosing a tool. It’s sitting down with the person who runs your most predictable process and writing down every step they take — including the ones they think are too obvious to mention. You’ll have the list. And then, honestly, the first AI version of it is often a single well-written prompt and an afternoon.

If you’d like a second pair of eyes on which process to pick, or on the first version of it, that’s exactly the kind of thing I do. Send me a description of how the job works today, in whatever state it’s in, and I’ll tell you straight whether it’s a good first candidate and what the ninety-day version might look like.

Start somewhere. The rest follows.