Using an AI agent in a workshop, starting sensibly.

Hand over the repetitive, keep the judgement

Agents are strong where work is high volume, rule-driven and tedious, and weak where it needs commercial judgement or accountability.

That splits a workshop week fairly cleanly. Checking supplier availability across several portals, applying a pricing rule to a new list, chasing quotes that went unanswered, sending service reminders and drafting routine replies are all repetitive work somebody currently does badly because there is never time.

Deciding where you sit in your local market, whether to write off a comeback, what to say to an unhappy customer and whether a fault is worth pursuing are not. Those need someone accountable, and handing them over is how using an AI agent gets a bad name.

Start with something reversible

The first task should be one where a mistake is visible and cheap, which rules out most of the obvious candidates.

• Drafting replies for a person to send, rather than sending them.
• Preparing a supplier order for approval, rather than placing it.
• Flagging quotes that need following up, rather than following them up.
• Producing a list of customers overdue a service, rather than contacting them.
• Summarising what happened this week, where being wrong costs nothing.

That pattern of proposing rather than acting is a reasonable place to stay for weeks. It builds the judgement about where it is reliable, which is not something anyone can tell you in advance about your own data.

It is only as good as what it can see

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

A workshop with customer data in three systems, a stock record nobody trusts and job times that were never recorded will get poor results, and the problem is the data rather than the agent. That is worth knowing before concluding the technology does not work.

It cuts the other way too. The workshops that get the most from this are usually the ones that had already tidied their records for entirely unrelated reasons, which is a reasonable argument for doing that first.

Keep a person in front of the customer

There is a strong temptation to point an agent at customer communication, because that is where the volume is.

Some of it is fine: confirmations, reminders, and answers to genuinely routine questions. What should not be automated without a person is anything where a customer is unhappy, anything involving a price that has not been agreed, and anything where being wrong damages trust rather than merely wasting time.

The test worth applying is whether you would be comfortable with the customer knowing exactly how the message was produced. Where the answer is yes, automate it. Where it is no, that is a signal rather than a detail.

Who is accountable when it is wrong

This question gets skipped and it is the one that decides whether an AI agent is usable in a business at all.

If an agent orders the wrong part, quotes a price that was never agreed or sends a message that annoys a customer, the answer cannot be that the software did it. Somebody has to own the output, which in practice means somebody has to be reviewing enough of it to know whether it is reliable.

That is why the proposing-rather-than-acting stage matters beyond caution: it is where the person who will be accountable learns what the agent is actually good at. Skipping it means the first serious mistake is also the first time anyone looked closely, which is the worst possible order.

Measure whether it actually saved anything

Automation has a way of moving work rather than removing it, and nobody notices because the new work is different.

The number worth watching is the time the task used to take against the time now spent setting it up, checking it and correcting it. A task that took forty minutes and now takes thirty-five minutes of supervision has not been automated, it has been rearranged.

Where it works properly the difference is large rather than marginal, and it shows up as a task disappearing from someone day rather than shrinking. If that is not what happened, the honest response is to change the task rather than to persevere with it.