How to Start an AI Automation Business: Five Steps, in This Order

Every few months a new AI tool arrives and a crowd of people starts learning it from scratch. Nick Puru has watched that cycle from the inside: his channel moved from Make to n8n to Claude, and his agency, Reprise, kept selling the same kinds of fixes the whole time. His argument is simple. The tool is not the job. Knowing how a business makes and loses money is the job, and the tool is just what you use to act on it. If you are working out how to start an AI automation agency, his recent video gives you the order. One slide lists five steps for anyone starting from zero. That slide is the spine of this guide. Each step title below is copied exactly from his screen; the explanations are ours, built from what he said.

How to start an AI automation agency: stop learning tools, learn one business. Nick Puru's five steps for starting an AI automation agency from zero, straight from his slide.

How to start an AI automation agency when the tools keep changing

Owners do not buy Claude, n8n or Zapier. Puru sorts what they actually pay for into four buckets: hours back, lower costs, more revenue, and an edge a rival can't easily copy. The request rarely arrives in those words. "Can you automate our customer service?" usually means staff answer the same questions all day (time). "We need better lead tracking" usually means leads are slipping away (revenue). Part of the work is translating the ask into the bucket, because an owner values plugging a customer leak far above saving an hour on a report.

The demand is real and mostly unmet. A Goldman Sachs 10,000 Small Businesses Voices survey of 1,256 owners was published March 17, 2026. It found 76% already use AI. Just 14% say it is fully built into their core operations. The software is already in the building. What is missing is someone who turns occasional use into a business that runs on it.

Tourist, builder or AI operator?

Puru splits people working in AI into three groups. Tourists collect tools and know each feature the day it ships, but have shipped nothing of value. Builders get something working for themselves, an inbox sorter or a report writer, that nobody pays for. Operators have put a system inside a real business that pays for the result and keeps running when they are not in the room.

His quick test: look at your last seven days. Count the hours spent learning or building, then count the messages you sent to business owners. If the second number is near zero, you are a tourist or a builder, however good your builds are. The five steps below are how you move to the third group.

He also ranks the work in five levels: running the tool, connecting systems, designing the fix, finding the real problem, and changing how the business runs. The tools are swallowing the bottom levels fast. The top two, finding the real leak and changing the operation around it, are where the money is moving.

Why the operator path holds up

Here's the reality. I've built, scaled, and sold three companies, generated over $85 million in the last seven years in my own companies. None of them won because of a tool. They won because someone understood exactly where the business leaked money and fixed that first. Puru's data points the same way.

Evidence Source
76% of small businesses use AI; 14% have it fully built into core operations Goldman Sachs 10,000 Small Businesses Voices, 1,256 owners, Mar 17, 2026
Of about 360 owner requests to his agency, the top asks were follow-ups (152), weekly reports (127), inbox and phones (92) and client onboarding (91) Nick Puru, Reprise's own inquiry data, as stated in the video
68 of 192 teams on his growth plan form ticked "not sure yet, show me where AI helps most" Same

How to start an AI automation agency: the five steps

These are Puru's steps, in his order. He is clear that the order matters: each one sets up the next.

Step 1: Pick one kind of business and stay with it

Choose one industry you can actually get a meeting in, such as property managers, accounting firms or clinics, and commit to it. Without prior experience, learning one industry properly can take a year or two: who answers the phone, how quotes are built, where jobs stall. That knowledge compounds, and nobody can copy it from a tutorial. Switching niches every month resets it to zero.

Step 2: Find one problem you can measure before you speak to them

Walk in with a number, not a pitch. Good candidates are visible from the outside: whether anyone answers the phone after hours, or how long a quote takes to come back. Puru's favorite is the contact form test. Fill in the business's inquiry form like an ordinary customer and start a timer. If the reply takes five hours, you have found the problem, and the exact time is your opening line. A buyer who is shown their own leak listens differently from one who is offered a chatbot.

Step 3: Write down how the business works before you build anything

The answer to the problem rarely lives in the model, since every business can rent the same one. It lives in what the business knows: how it prices work, whose calls get returned first, how the owner handles an upset client. In most small firms that sits in two or three heads. Get it out with roughly one conversation per department head. Put everything in one folder of plain files: a rules file for how the work is done, a skill for each repeated task, and the list of apps the agent is allowed to touch. Puru keeps a CLAUDE.md and an AGENTS.md side by side saying the same thing, so any agent can work from the same folder.

Step 4: Build one fix for one business, and measure what changed

Pick a daily task that touches money or time. Record the number before you start and again after go-live. Define success as a pass mark rather than a promise; his proposals state that the system must "sort at least 90% of a test set of emails correctly." Keep the first phase small enough that a nervous owner can say yes. In one of his examples, an owner worried about a 35-million-record database got a first phase that ran on a test copy for $7,000 instead of $22,000.

Step 5: Sell the same fix to the next business like it, priced on the result

Now take the proven fix to the next firm of the same type and price it on what it changes, not on your hours. Hourly pricing punishes you for getting faster. Price against the cost the fix removes or the capacity it frees, and add a monthly retainer for keeping it running (he cites $500 to $1,500 monthly, sometimes more). We walk through both pricing methods with the full arithmetic in our post on value based pricing for AI automation. After five or ten repeats, the build that took weeks takes days, because you have seen the same patterns already.

How to get AI automation clients to say yes

A stranger will not hand you their whole business. Puru's tactics for lowering the risk on their side:

  1. Start with one small, visible fix. A first phase they can watch working.
  2. Lead with the measured problem. The contact form result from step 2.
  3. Reverse the risk. If the promised result doesn't show, you carry the cost; for lead generation, they pay per meeting booked.
  4. Promise less than you expect to deliver. On one project his team projected a 73% first-year return; it came in at 123%.
  5. Say no to the wrong request. His team won a recruiting job while being the most expensive bidder, partly by refusing a scraping plan that would have got the client's account banned.

He is also honest about failure. One clinic agent started double booking appointments because a setting skipped the availability check. The fix was a real check before every booking plus a switch letting staff take over any chat. Learn the tool well enough to judge its work and see why it is wrong; that is the level of technical skill this path needs.

Worked example: how to start an AI automation business with dental practices

Here is how a new operator might run the five steps in one quarter.

  1. One kind of business: independent dental practices within driving distance, because the operator's sister manages one and can open doors.
  2. One measured problem: the operator submits the new patient form on 20 practice websites on a Tuesday evening. Median first reply: 19 hours. Six never reply. That list becomes the outreach.
  3. Write it down: with the first practice that agrees to talk, one conversation each with the office manager, the front desk lead and the hygiene coordinator. The folder holds rules.md (how new patients are greeted and booked), skills/ (insurance questions, reschedules) and connections.md (the booking calendar, the inbox, nothing else).
  4. One fix, measured: an assistant that replies to web forms within two minutes, answers the top ten questions from the rules file and offers three open slots. Pass mark agreed up front: answer at least 90 of 100 past inquiries correctly in testing. Before: 19 hours to first reply. After, at day 30: under two minutes, with booked new patient visits counted weekly.
  5. Sell it again: the operator takes the same build, the same test and the before and after numbers to the other 19 practices. The price is set against the extra booked visits, not the hours it took.

Common mistakes

  • Learning every new tool instead of one industry.
  • Pitching automation in general instead of one measured leak.
  • Building before writing down how the business runs.
  • Skipping the before number, so nobody can prove the fix worked.
  • Pricing by the hour, then earning less as you get faster.

Rather have an operator do this for you?

Run a business rather than want to start one? We do these five steps on your side. We measure where leads and hours leak, write down how your team actually works, build one fix and report the before and after numbers. You approve the pass mark before anything is built and see the result before you commit to more, so you aren't paying for a build nobody uses. Why now: most of your competitors are still in the "we use it sometimes" stage.

What you get Where
A measured leak, a documented workflow, one fix with before and after numbers AI automation

Free tools

One free, MIT-licensed skill for Claude, Codex and Grok comes with this guide: Operator Kit (checks whether your week is a tourist's, a builder's or an operator's; sorts owner requests into time, cost, revenue and edge; runs the contact form response time test and writes each opening line; prices a fix by cost and by capacity, plus the retainer; scores a build against its pass mark; and checks the business folder before you build). Download the skill (free zip)

FAQ

Do I need to code to do this?

No, but you need to understand the tool well enough to judge its output and spot why it went wrong. Puru is not a developer.

What is a contact form response time test?

You fill in a business's website inquiry form as a normal customer would and time how long a reply takes. A slow reply is a measured problem you can bring to the first conversation.

How long until I know an industry well?

Puru puts it at a year or two without prior experience. Staying in one niche is what makes that time count.

Should I charge by the hour?

He advises against it. Price on the cost the fix removes or the capacity it frees, plus a monthly retainer for upkeep.

Which AI tool should I learn first?

Whichever one lets you build and check your first fix. The step that matters is understanding one business, because the tools will keep changing.

Sources

Ready to sell the fix, not the tool?

Book a call and we'll find the one leak in your business that is worth fixing first.

Book a Search Strategy Call