← The notebookBuilding · 5 MIN READ

Where to start with AI in your business

People ask me this all the time. They've tried ChatGPT. Or someone on the team is already using it quietly, because nobody set a rule. Or a board member heard a talk and now there's pressure to "do something with AI."

The question underneath is usually the same. Where do we actually start?

Not with a tool. Not with a 40-page strategy. Start with the work your company already does, and one number you can measure.

I already wrote a short note on this: do the work manually, then automate it. That still holds. This is the longer version, for a company that wants a first real step.

Look at a normal week, not a trend.

Walk through how the business actually runs. Where does time go? What gets repeated every day or every week? What sits in a queue? What costs you money when it goes wrong?

You're looking for a task with three traits.

It happens often. A weekly task teaches you more than a quarterly one.

You can measure it today. Time, errors, backlog, hours. Something with a before.

A person can check the output before it reaches a customer. First projects should be recoverable. If the model gets it wrong, someone catches it.

Good first tasks look like this: drafting the first version of a common email, pulling the same fields out of incoming documents, summarizing a call into a standard format, sorting routine questions so the right person sees them.

"We need AI in marketing" is not a first task. That's a department. Shrink it until you can say the job in one sentence, with an input and an output.

Write down how the work works today.

Before anyone signs up for a new product, write the current process in plain language. What happens first? What does a good result look like? What happens when something goes wrong?

If you can't explain the work, AI will just make the mess faster.

This is the part people skip because it feels slow. It isn't. A page or two will show you whether the task is even ready. Sometimes the real problem is that five people do the same job five different ways. Fix that first, then decide what to automate.

At Ecommerce Scaling, I wrote the operating procedures first and then automated them. Writing the process down forces you to confront the steps. I still start there.

At Cleveland Clinic Innovations I taught more than 70 staff how to use AI safely. We started with the work already on their desks, not a tour of features.

Pick one number, before you change anything.

Time per task. How often it goes wrong. Hours spent each week. How long a customer waits. One number, with a before.

If you don't have a baseline, you will never know if it worked. You'll have a feeling. Feelings are how tools stay in the stack after they've stopped helping.

Write the target too, and the result that would make you stop. "A little faster" is not a target. "We cut drafting time, and the error rate did not go up" is closer.

Name one owner. Not a committee, and not "the team." Someone has to be accountable for the outcome, not just for picking the software.

Set a few rules so the team isn't guessing.

Your people are already experimenting. That's normal. Give them a safe lane instead of pretending they aren't.

You need four things on one page:

Which tools are approved.

What may not go into them. Customer records, financials, contracts, anything confidential.

When a human has to review the output before it leaves the company.

Who owns the pilot.

If the rules need a binder, the first project is too big.

Run a short pilot, then make a real decision.

A few weeks is enough for a narrow task. Use real work, not a demo. Keep the group small. Have someone review every output at the start.

At the end, look at the number you wrote down. Scale it, fix it, or stop.

Stopping is a good outcome. It means you didn't roll a weak process out to the whole company.

Then pick the next task and do the same loop. After two or three of these, you'll know more than a strategy offsite would have taught you.

You don't have to do the first pass alone.

I think there's value in trying the first one yourself. You learn the work that way. The questions get sharper.

If you're running an established company and you want a team that starts with your numbers, that's the work we do at AI Answered. I'm Co-founder of AI Answered. We sit with how the business actually runs, agree on a number we can track, and build around that. We price on the value we can deliver, not a packaged plan.

And if people on your team are new to this and just want a plain-language place to begin, Noah Kruthaupt and I wrote AI for Boomers for that. The point of the book is the same as the point of this: technology should make the work simpler, not add another layer of confusion.

That's the start. One task. One number. A short test. An honest decision. Then the next one.

If you're choosing who to hire for that first step, here's what to look for in an AI consultant. And if the company hasn't measured itself yet, AI won't fix a business that doesn't know its numbers.

More from the notebook ↗