What an AI agent actually does inside a business

Nearly every business I walk into has one person who has got good at ChatGPT. They open a tab, ask it something, read the answer, copy the useful bit into an email and get on with the rest of their day. It genuinely helps that person, and I would never tell them to stop. But it is a faster way to type. It is not the work getting done, which is why a year of it can go by and nothing about how the business actually runs has moved an inch.

Why the chat window never changes anything

The problem is not that the model is weak. Models are very good now. The problem is that a chat window is sealed off from your business. It cannot see your diary, so it has no idea whether Thursday is free. It cannot look a customer up, so it does not know that this one still owes you four hundred quid from March. It cannot send anything, so a person has to copy the words out and press the button. It forgets the whole job the second you close the tab. And it will never come back on Thursday afternoon to check whether the customer ever replied.

So every step still waits for a person with a free ten minutes. That waiting is the actual bottleneck in most small firms, and a chat window sits politely outside it.

What an agent is

An agent is software you hand a job to rather than a question. It is the same AI underneath. The difference is what it is wired into and what it is allowed to do.

• It reaches your systems, so it can read the inbox, check the diary and pull up a customer's history in the same few seconds.
• It decides using rules you have set down. What work you take on, what you charge for it, how far you will travel, what counts as urgent.
• It carries out the next step instead of suggesting it. It books the slot, writes the email, files the invoice.
• It comes back. If a quote has gone quiet for a fortnight, the agent is what notices, rather than you at half nine on a Sunday night.

That last one is the part people underestimate. Most of what leaks out of a small business leaks because nobody had a spare afternoon to follow it up.

One enquiry, from the first email to the money landing

Take a commercial heating and plumbing firm. One enquiry arrives, and here is the whole run through to a paid invoice. Nine steps. Two of them stop and wait for a human being.

How much of that is actually AI

Count the steps that genuinely need a language model. One, and a bit of another. Step two reads a badly typed message from a customer who does not know the difference between a boiler and a plant room, and works out what they are asking for. Step six writes the reminder so it sounds like a person at your firm wrote it. That is the clever AI part, and it is small.

Everything else is a lookup, a rule or a diary check. Whether this customer has an unpaid invoice. Whether Dave holds the commercial ticket. Whether there is a two hour gap on Tuesday. Whether the invoice matches the quote. Whether it has been 21 days. None of that needs intelligence. It needs access, and it needs something that does not get distracted.

This is why buying a smarter model on its own changes almost nothing. If it is not wired into your diary and your accounts, all you get is better drafts for somebody to copy out by hand.

Where people stay in it

Money and promises. Anything that quotes, prices, sends or promises a customer a date waits for a person to say yes. It is worth holding that line even when it looks like the slow part, because the moment software can promise a date on your behalf you have handed over the thing your name rests on. In practice it is a few seconds of glancing at a number that is already right, rather than the half hour that job used to cost you.

What it plugs into

The systems you already run. The same inbox, the same diary, the same accounts package, the same way your engineers write a job up on their phone. Nobody gets a new screen to learn, there is no retraining week, and nothing gets ripped out. If a build like this needs your team to change how they work on day one, the people who have to live with it will quietly kill it inside a month, and they would be right to.

What actually changes

That enquiry used to take four days to become a booked job. Almost none of those four days was work. Add up the real effort and it is about forty minutes. The rest was queueing. Sitting in an inbox until somebody looked. Sitting on a desk until somebody had time to open the diary. Waiting for the quote to get typed up on Friday afternoon.

Take the queue out and the same team, on the same systems, with the same engineers and the same van, turns it round the same day. Nobody is working harder or longer. The waiting has just stopped.

Where to start

One process. The one costing you real money that you could name right now without thinking about it. Quotes taking three days. Invoices nobody chases. The enquiries you know go cold in the inbox.

1. 1.Pick the process you can put a rough cost against, in hours lost or in work that went elsewhere.
2. 2.Get the rules out of people's heads and written down. This is the hard part, and an outside pair of eyes earns its keep, because the rules nobody has written down are the ones nobody realises they are applying.
3. 3.Build that one process the whole way through, with a person approving anything that quotes, prices, sends or promises a date.
4. 4.Run it live for a few weeks against your own old numbers and fix what the real world breaks. Something always breaks, and it is never the thing anyone predicted.
5. 5.Only then start the second one. It goes quicker, because most of what you built the first time is what everything else stands on.

Almost none of what makes that example work is clever. It is access, rules somebody bothered to write down, and something that never forgets to chase. If you have tried AI and nothing changed, the tool was probably not the problem. The more useful question is which job in your business is quietly costing you the most while it sits there waiting for somebody to have a free ten minutes.