How I would build an AI automation agency from zero

If I were starting again with no audience, no list and no case studies, this is the order I would go in. It is not the exciting order.

One process, one industry

Not "AI automation for businesses". Pick one repetitive process in one type of business. Missed-call follow-up for clinics. Quote chasing for contractors. Review requests for restaurants.

Narrow is not a limitation at the start, it is the only way to get good fast. The second client in the same niche takes a third of the time the first did, and that gap is your entire early margin.

Build it once, for free, on purpose

Find one business, build the thing properly, charge nothing, and ask for two things in return: permission to write about the result, and honesty about what did not work.

You are not buying a testimonial. You are buying the discovery of every edge case you did not anticipate, which is what you will actually sell later.

Charge for the outcome, not the automation

Nobody wants an automation. They want the missed calls to stop being missed. Price against the cost of the problem, not against the hours it took you.

A useful shape: a setup fee that covers the build, then a monthly fee that covers the part everyone forgets, which is that automations break. The maintenance is the business. The build is the entry ticket.

Sell before you scale

The failure mode I see most in new agencies is building infrastructure for a business that does not have clients yet. A brand, a website, a team, a fourteen-step onboarding, all before deal three.

Ten clients in one niche will teach you more about what to build than any amount of planning. Get to ten, then systemise what those ten actually needed.

Then hire for the part you hate

Whatever you have been avoiding is where the first hire goes. For most technical founders that is sales, and for most sales founders it is delivery.

Hire for it before it becomes the ceiling, because by the time it is the ceiling you no longer have the time to train anyone.

The part that takes longest

Everything above is maybe ninety days of work. The part that takes years is judgement: knowing which client will churn on the discovery call, which process is not worth automating, which prospect is describing a symptom.

There is no shortcut to that. It comes from having been inside enough businesses to see the pattern repeat, which is why the first ten matter more than the price you charge them.