How to hire an AI builder: a five-step process for business owners

Hire an AI builder in five steps: write the one task and the number that would show it worked; shortlist two or three builders who show proof of a similar build; hold one thirty-minute call each with the questions below; run a two-week pilot in your own accounts with "done" written down; then decide between continuing, a support plan, or stopping with what you have. The proof and the pilot do most of the work; the call is for finding out how the builder handles what goes wrong.

Step 1: define one task

Before you look at anyone, write one sentence: "Reply to website leads within a minute, log them in the CRM, and tell me when one is urgent." Add the number: "median reply time, now nine hours." A builder can quote and pilot that. "Help us with AI" gets you a proposal for whatever the builder likes building.

Step 2: shortlist by proof

On the marketplace or anywhere else, look for builders who show a finished build close to your task: a screenshot of the actual workflow, a recording of it running, a measured before-and-after, and a client who agreed to be named. Ignore claims without evidence. Two or three candidates who pass are enough. Location matters less than whether they've worked with your channel (WhatsApp, in most Israeli small businesses) and your tools.

Step 3: the thirty-minute call

Ask, and listen for the shape of the answer:

• "Show me something similar you've built." (They open it live, or they don't.)
• "What went wrong on it, and what did you change?" (Every real project has an answer.)
• "What happens when a message doesn't fit the rules?" (There should be a path to a person.)
• "Where does my data go, and who can see it?" (A clear answer with minimisation and access per person; the Privacy Protection Authority's guidance is the reference.)
• "What's 'done' for the pilot, and what's excluded?" (Written, before you pay.)
• "What does it cost to run per month, and what does support look like?" (All the running costs, not just the build.)
• "Who owns the accounts and the workflows?" (You do, in your own tools.)

Answers that should worry you: a demo they can't open, "it never fails", a vague answer about data, a promise to automate "everything", and a price with no scope.

Step 4: the two-week pilot

One task, fixed price, in your accounts, tested on your real data, with a hand-over page your team can follow. Give the builder admin access and a decision-maker who answers within a day; slow answers are the main reason pilots overrun. At the end you should see the number, the logs, and the exceptions that went to a person.

Step 5: decide

Three good outcomes: continue to a second adjacent automation, take a support plan and let it run, or stop with a working pilot and a write-up. One bad outcome: nothing measured, nothing handed over. The pilot's structure is what prevents that, which is why it's worth insisting on before the first shekel.

Best fit and not a good fit

Best fit: owners hiring for a first or second automation. Not a good fit: businesses needing a full system or a team on day one; hire an agency or a lead builder with a track record of larger projects, and still start with a pilot.

What to do next

Write the sentence and the number. Then open the marketplace, filter for builders with a similar project and visible proof, and book the first call.