Using an AI agent in a garage, starting sensibly.

Delegate the repetitive, retain the judgement

Agents excel in high-volume, rule-based, tedious tasks, but struggle where commercial judgement or accountability is needed.

This divides a garage week neatly. Checking supplier availability across multiple portals, applying pricing rules to new lists, chasing unanswered quotes, sending service reminders, and drafting routine replies are all repetitive tasks that someone currently does poorly due to time constraints.

Deciding your position in the local market, whether to write off a comeback, how to address an unhappy customer, and whether a fault is worth pursuing are not. These require accountability, and delegating them is how AI agents gain a bad reputation.

Begin with something reversible

The first task should be one where a mistake is clear and inexpensive, which rules out most obvious choices.

• Drafting replies for someone to send, rather than sending them.
• Preparing a supplier order for approval, rather than placing it.
• Flagging quotes that need chasing, rather than chasing them.
• Producing a list of customers overdue for a service, rather than contacting them.
• Summarising this week's events, where being wrong costs nothing.

That pattern of proposing rather than acting is a sensible place to remain for weeks. It builds judgement on reliability, which no one can tell you in advance about your own data.

It’s only as good as what it can access.

An agent working from incomplete records gives confident, incorrect answers, and confidently wrong is worse than nothing.

A garage with customer data in three systems, a stock record no one trusts, and job times never recorded will yield poor results; the issue is the data, not the agent. It’s important to know this before concluding the technology doesn’t work.

It works the other way too. The garages that benefit most are often those that have already organised their records for unrelated reasons, which is a good case for doing that first.

Keep a person in front of the client.

There's a strong urge to direct an agent at customer communication, as that's where the volume lies.

Some is fine: confirmations, reminders, and answers to routine queries. Avoid automating anything where a customer is unhappy, involves an unagreed price, or where errors harm trust rather than just waste time.

The key test is if you'd be comfortable with the customer knowing how the message was created. If yes, automate it. If no, that's a signal, not a detail.

Who is responsible when it goes wrong?

This question is often overlooked, yet it determines if an AI agent is viable for a business.

If an agent orders the wrong part, quotes an unagreed price, or sends a message that irritates a customer, the answer can't be that the software did it. Someone must take responsibility for the output, meaning they need to review enough to ensure reliability.

That's why the proposing stage is crucial: it's where the accountable person learns what the agent excels at. Skipping it means the first major mistake is the first close inspection, which is the worst order.

Assess if it truly saved anything.

Automation often shifts work rather than eliminates it, and no one notices as the new tasks differ.

The key metric is the time the task used to take versus the time now spent on setup, checking, and corrections. A task that took forty minutes and now takes thirty-five minutes of oversight hasn't been automated; it's been rearranged.

Where it works well, the difference is significant, showing as a task vanishing from someone's day rather than just reducing. If that's not the case, the honest response is to change the task, not persist with it.