An AI plan for a small insurance agency: the first 90 days

A small insurance agency (the UK word is brokerage) can start AI in 90 days with six steps: choose one job, measure how it runs today, set data rules, run a small trial with a person checking, decide whether to keep it, and decide who keeps it running. One job, one owner, one number to beat.

What is a sensible AI plan for a small agency?

A sensible plan for a small agency covers one job, not the whole office, over about 90 days. "AI" here means software that reads and writes text. The plan below puts the stages in order, with a rough time for each. The times are a guide, not a rule.

How do I choose the first job?

Choose the job that scores highest on four questions: how often it happens, how long it takes, how much it hurts, and how risky a mistake is. Give each job a score from 1 (low) to 5 (high) on the first three. Score the fourth the other way round, so 5 means a mistake is cheap and easy to catch.

1. How often? A task done every day beats one done once a quarter, because you learn faster.
2. How long? Minutes per case multiplied by cases per month gives hours. Hours are what you get back.
3. How much does it hurt? The task people put off, or the one that causes weekend work, scores high.
4. How risky is a mistake? A draft that a person checks before it leaves the office is low risk. Anything that goes straight to a client or an insurer without a check is high risk.

Add the four scores. The highest total is your first job. Ties go to the lower-risk job. Everyday writing is low risk but often low pain too, so it may not win. Renewal preparation, statement matching and policy changes often do. Our guides on preparing renewals, matching insurer statements and processing policy changes show what each job looks like.

How do I measure the job before AI touches it?

Measure by timing the job by hand for a few weeks and counting its mistakes. You need three numbers before the trial starts: time per case, cases per month, and how often a case comes back wrong. Without them, you cannot say afterwards whether the trial helped.

Ask the person who does the job to note the start and finish of each case for two weeks. Add a tally of what had to be redone. Write two sentences describing what "done correctly" looks like. A system cannot do better than the description it is given, and a person checking its work needs the same description.

What data rules do we need before a trial?

You need a one-page rule that says what staff may put into an AI tool, which tool, and on which plan. Write it before the trial, because the trial is where client data first meets the tool. The full reasoning is in our guide on whether ChatGPT is safe for client data.

Makers publish what their business plans do with your data, and the terms differ between consumer and business plans. Anthropic's privacy policy (effective 10 September 2026), for instance, says it may use consumer conversations to train its models unless you opt out, and that the policy does not apply to content processed for its business customers, which is governed by customer agreements. Read the plan you will actually use, and keep a copy of what it said on the day you chose it.

If you are regulated in the UK, the FCA says on its page on its approach to AI that it will "rely on existing frameworks, which mitigate many of the risks associated with AI." The page lists accountability and governance among them. Our suggestion, not an FCA rule: put a named senior person behind the data rule.

How do I run a small trial?

Run the trial on real cases for six weeks or so, with AI preparing each draft and a named person checking every one before it goes anywhere. Keep the old way available. Time each case and tally each correction, the same way you did before.

Keep the check real. If the person approves everything within seconds, they are not checking. Ask them to note each correction and why. Those notes are the most useful output of the trial, because they show where the system goes wrong and whether the rules need to change.

How do I decide whether to keep it?

Keep it if the trial numbers beat today's numbers on time and on mistakes, and the person checking says the drafts are worth starting from. Compare the three numbers from stage 2 with the trial. If time fell but mistakes rose, do not keep it yet. If neither moved, drop it and try the second job on your list.

A made-up worked example

This example is made up, to show the arithmetic. A fictional office of twelve people matches commission statements each month. The owner scores the job: often 3, long 5, painful 5, cheap to catch mistakes 4, total 17. Renewal letters score 14, so statements go first.

Today, one person spends 10 hours a month on it and finds a handful of mismatches that need a second look. In the 6-week trial, the system prepares the match, the same person checks it, and the check takes 3 hours. The person caught 2 lines the system had matched wrongly, both from a statement with an unusual layout. The owner keeps the job, adds a rule that unusual layouts go to a person first, and names the office manager as owner. The numbers are invented. Your own measurement will differ, and it is the only one that counts.

Who keeps it running afterwards?

One named person owns it: they watch the results, handle corrections, and decide when to change the rules. Without an owner the job drifts. A system that worked in week 10 can fail in month 6 when an insurer changes the layout of its statements and nobody notices.

Ask three questions about ownership. Who looks at the output each time it runs? Who is told when something looks wrong, and how fast? Who decides to switch it off? If you do not want to be the one who maintains software, the owner can be an outside builder who runs the system and watches it, while your team keeps the checking and the decisions. Compare that route with the others in our guide to AI tools for insurance agents.

Back at the start of this plan? Read the overview of AI in an insurance agency or the main guide to AI for insurance brokers.

Where Navicade fits

Navicade builds and runs AI systems that do work the office does by hand. A free 30-minute call and a free operations map of where your hours go come first. Then we take on one agreed task at a time, which fits the plan above: one job, measured, checked by your people. Results arrive in your existing tools. Insurance decisions stay with your team. The system checks its work on every run and alerts us.

A beginner may not need us yet. Stages 1 to 3 cost you only time, and you can do them alone. If you reach stage 4 and want someone else to build and run the trial, ask for the free call and map.