How to Build an AI Agent for Your Small Business

Before you start: pick the right job

Do not begin with the tool. Begin with a job that meets four tests:

• It happens at least weekly.
• It costs you or your team more than an hour a week.
• You could explain how to do it to a new employee in five minutes.
• Getting it slightly wrong is annoying, not expensive.

Chasing quiet leads, sorting the shared inbox, chasing overdue invoices and writing the recurring report all pass. Signing off on a contract does not.

If you cannot name the job in one sentence, you are not ready to build yet. That is not a failure - writing the sentence is the actual work.

Step 1: Write the brief

The brief is a paragraph. It has four parts, and the fourth is the one people skip.

What the agent does. "Watch the sales inbox and follow up with enquiries that have had no reply for three working days."

What it can see and touch. "Read the sales inbox. Draft replies. Log activity against the contact in the CRM. Nothing else."

How it should sound and behave. "Short, direct, first person, no exclamation marks. Reference what the person originally asked about. Never quote a price."

What it must not handle. "If the enquiry mentions a complaint, a refund, legal terms, or a discount, do not reply - summarise it and pass it to me."

That last part is what turns a demo into something you can leave running. Write it before anything else if you find it easier.

Here is the whole thing as one paragraph, which is how you would actually type it:

Watch the sales inbox for enquiries with no reply after three working days. Draft a short follow-up in my voice referencing what they asked about, and log the activity against their contact record. Do not mention prices. If the message involves a complaint, refund, discount or legal question, do not reply - summarise it and send it to me instead.

Step 2: Give it the least access that works

Connect only the systems named in your brief. For the example above: the sales inbox and the CRM. Not the accounting system, not the shared drive, not your personal calendar.

Two rules worth holding to:

• One agent, one job. An agent with three jobs and access to everything is harder to trust and harder to fix. Build the second agent later.
• Connect as the right person. The agent should act within someone's real permissions, so the record shows who it acted for. Avoid a single shared admin login doing everything.

Step 3: Turn approvals on before you turn the agent on

Set every outgoing action - emails, messages, record changes - to require your sign-off. Reading and drafting can run freely from the start.

You will relax this per category later, once you have watched a particular kind of action be right twenty times in a row. Some categories never get relaxed, and that is a legitimate permanent design, not a sign the thing is unfinished.

Step 4: Test on ten real cases

Not made-up ones. Ten real items from the last month, chosen to include the awkward ones: the vague enquiry, the angry one, the duplicate, the one in a language you do not speak, the one that mentions a competitor.

For each, check three things:

1. Was the answer correct and grounded? Did it use your actual information, or invent something plausible?
2. Did it escalate what it should have escalated? The complaint and the discount request should have come to you untouched.
3. Would you have sent this? If you would rewrite more than a sentence, the brief needs tightening, not the message.

Fix the brief, not the individual outputs. If it went off-tone, add a line about tone. If it answered something it should have passed on, name that case in the exclusions. Two or three rounds of this usually gets it to where you would sign off nine out of ten drafts.

Step 5: Run it live, small

Turn it on for one narrow slice - one inbox, one part of the pipeline, one week. Keep approvals on.

Then check the activity log every day for the first week. You are looking for three patterns: things it handled that you did not expect it to see, things it escalated that it could have handled, and anything that took an unusually long time. All three tell you what to change.

Step 6: Count the hours

At the end of two weeks, write down two numbers: how many items it handled, and how many minutes each of those would have taken a person. Multiply. Subtract the time you spent reviewing and fixing.

That number is your answer. If it is positive and meaningful, widen the agent's slice or build the next one. If it is marginal, either the job was wrong or the brief is too vague - both are cheap to correct at this stage, which is exactly why you started small.

What goes wrong, and what to do about it

The brief is too vague. "Handle customer emails" gives you unpredictable behaviour, because you have not said what "handle" means or where the edges are. Name the cases.

Too much access. It is tempting to connect everything on day one. Every extra connection is extra surface for a mistake and extra noise when you debug. Add access when a job needs it.

Nobody reads the log. The first fortnight of activity is the most useful information you will ever have about this agent. Fifteen minutes a day for two weeks, then weekly.

Treating setup as one-off. Expect to update the brief twice in the first month as real cases surface. On a platform where that takes two minutes, this is fine. On one where it takes a ticket, this is why the project dies.

Automating judgement. If the reason the task is hard is that it needs your read on a relationship, an agent will not fix that. Automate the admin around it instead.

The realistic timeline

• Afternoon one: write the brief, connect two systems, test on ten cases, tune twice.
• Week one: live on a narrow slice with approvals on, log checked daily.
• Week two: count the hours, tighten the brief, decide whether to widen.
• Month two: release approvals on the categories that have earned it. Build the second agent for a different job.

That is it. No strategy document, no engineering ticket, no six-week implementation. One boring, expensive, repetitive job, described properly and watched for a fortnight.

Describe your first agent free - or try the preview on the home page first, with no signup and nothing saved.