Your First AI Agent: How to Choose the Right Job

Why the first job matters

An AI agent is software that reads something, decides what to do with it, and then does it inside your systems. If that idea is still new, our plain explanation of what an AI agent is covers it in a few minutes.

Most owners we speak with already have a long list of jobs they would hand to an agent tomorrow. The hard part is choosing the first one.

The first job sets the tone. If it goes well, your team trusts the next one. If it goes badly, the whole idea gets shelved, often for reasons that had nothing to do with the technology. So the first job should be useful, safe to get wrong, and easy to measure. This guide walks through how to find it.

Step 1: List the repetitive work

Spend a week writing down every task your team does more than a few times a week. Ask each person, not just the managers. The people doing the work know where the time goes.

For each task, note four things:

• What starts it (an email, a phone call, a form, a supplier invoice, the end of the month)
• Which systems the person opens to do it (Xero, MYOB, your job system, your CRM, a spreadsheet)
• What the finished result looks like (a quote sent, a bill entered, a booking made)
• Roughly how often it happens in a week

A typical trades business might end up with quoting from enquiry emails, chasing unpaid invoices, booking jobs into the calendar, and entering supplier bills. A clinic might list appointment requests, reminder calls and answering the same handful of questions about parking and fees. Write them all down. Do not judge them yet.

Step 2: Score each task

Score every task from 1 to 3 on four questions. Higher is better for a first agent.

The middle question needs care. A task that is pure rules, such as copying a paid invoice from one system to another, does not need an agent at all. Ordinary automation does that more cheaply and more predictably. Our guide on AI agent versus workflow automation explains the difference. A task that is pure judgement, such as deciding whether to take on a difficult client, should stay with a person. The sweet spot is work where the input is messy (an email, a PDF, a photo) but what to do with it follows rules you could explain to a new starter.

Step 3: Pick one job

Add up the scores and look at the top three. Then pick one, using two tie breakers.

First, choose the job where the person who does it today wants the help. Their knowledge goes into the brief, and they will be the one checking the agent work in the first month. A willing expert is worth more than a slightly higher score.

Second, choose the job with the clearest finish line. "A draft bill sits in Xero ready to approve" is easy to check. "Customers feel better looked after" is not.

Resist the urge to pick two. One job, done well, teaches you more than two done halfway.

Step 4: Write the agent brief

The brief is a one page description of the job, written as if for a new staff member. It becomes the basis for how the agent is built and tested. It has three parts.

What it does. The trigger, the steps and the finished result. For example: "When a supplier invoice arrives in the accounts inbox, read it, find the matching purchase order, and create a draft bill in Xero with the right supplier, account codes and GST (goods and services tax)."

What it must never do. Write these down plainly. Never approve or pay a bill. Never change a supplier's bank details. Never email a customer about money. Never delete anything. These limits are built into the agent's access, not just its instructions, so it cannot do them even if asked.

When it hands to a person. The situations where it stops and passes the work on, with a note explaining why. The amount does not match the purchase order. The supplier is new. The invoice is in a format it has not seen. A customer sounds upset. A good agent knows the edges of its job and says so.

Step 5: Decide what a person approves

At the start, a person should approve anything that spends money, commits the business to something, or goes out to a customer. The agent prepares the work; a person presses the button.

That approval step should be quick. A draft quote with the agent's reasoning beside it, an approve button and an edit button. If approving takes as long as doing the work, the agent has saved nothing.

Some steps can run without approval from day one, because they are internal and easy to undo: sorting incoming emails into the right folder, adding a note to a job, or tagging an enquiry by type. Decide this step by step, not for the agent as a whole.

Step 6: Measure it for a month

Run the agent on live work for a month with the approvals in place. Keep the measures simple:

• How many items it handled, and how many it handed to a person
• How often the approver changed its work before approving, and what they changed
• Mistakes caught at approval, and anything that slipped through
• Time the team spends on the task now, compared with before

The change rate is the number to watch. If the approver changes one draft in five, the brief needs work, and each change tells you which rule is missing. If they are approving almost everything untouched by week four, the agent is ready for more.

Every action the agent takes should be logged: what it read, what it decided and what it did. That log is how you answer "why did it do that?" when a question comes up, and it is part of any proper custom AI agent development project.

Step 7: Expand, one step at a time

Once the numbers are steady, expand in one direction at a time:

• Remove an approval for the low risk cases, for example bills under an amount you choose from suppliers you have used for years
• Widen the job to the cases it currently hands over, one type at a time
• Start the next job from your scored list, using what you learned

Change one thing, then measure again. If something goes wrong, you know which change caused it.

Two worked examples

These are typical situations, not client stories.

A plumbing business quoting from enquiry emails

A plumbing business with eight staff gets a steady stream of quote requests by email, many with photos. The office manager reads each one, looks up the customer in the job system, checks the price list, and writes a quote. Quoting scores high on volume, mostly follows the price list, and every quote can be checked before it goes out. The data sits in the job system and the price list, both reachable.

The brief: read the enquiry, find or create the customer, draft a quote from the standard price list, and attach the photos to the job. Never send a quote. Never discount. Hand over anything involving gas, anything outside the usual service area, and any enquiry where the photos do not show enough to price it. The office manager approves each quote for the first month, and the change rate shows which jobs the price list does not cover properly.

A wholesaler entering supplier bills

A wholesaler receives supplier invoices as PDF attachments in all sorts of layouts. Someone in accounts types each one into the accounting system and checks it against the purchase order. This scores well on every question: lots of invoices, clear matching rules, mistakes are caught before payment, and the data is in the accounting system and the inbox.

The agent reads each invoice, matches it to the purchase order, and creates a draft bill with the right codes. It flags mismatches in quantity or price and hands new suppliers to a person. It never approves a bill or touches bank details. Reading documents like this is close to what our AI document processing work covers. The agent adds the decisions and the handover on top.

Customer enquiries and bookings are also good first jobs for many service businesses, as long as the agent hands complaints, refunds and anything unusual to a person. If the job is only answering public questions on your website, a ready-made chat tool may be enough. Our article on custom AI agent or off-the-shelf covers how to tell.

Common mistakes

• Starting with the hardest job. The task everyone hates is often hated because it is full of judgement calls. That makes it a poor first job.
• Giving the agent more access than the job needs. If it only drafts bills, it should not be able to approve them. Limit access to the brief.
• Skipping the month of measuring. Without the numbers, the decision to expand is a feeling.
• Messy data underneath. An agent working from a customer list full of duplicates will create more duplicates. Our post on whether your data is ready for AI covers what to check.
• Leaving the team out. The people who do the work today should help write the brief and do the approving. They are the ones who will trust it, or not.

If you have scored your list and want a second opinion on the job at the top, our page on AI agents for business explains how we scope and build a first agent, starting with a short discovery and a fixed price after it.