Your first AI agent should chase your invoices

Most AI projects that stall do so because they started with the impressive thing. A chatbot for the website, a tool that writes proposals, something to show the board. Those are fine third projects. A first agent needs four properties: the work is frequent, it follows rules you can write down, a mistake is cheap to catch, and the value is measured in pounds you can see in the bank. Credit control has all four. It happens every week, the sequence is the same every time, a badly worded reminder is embarrassing rather than fatal, and the return shows up as cash.

There is a second reason, and it comes from the people who have not adopted AI. When Novuna Business Finance asked 1,000 small business owners this summer, 59 per cent of those not using AI said their business is built on human relationships, and they did not want a machine in the middle of them. That is a good instinct. It is also an argument for this agent in particular, because chasing money is the part of a relationship owners hate most. It is the phone call you put off and the email you soften until it says nothing. An agent that is polite, consistent and never irritated takes the awkwardness out of the relationship and leaves the human parts to the humans.

Start with what it reads. Every accounts package, whether Xero, Sage, QuickBooks or FreeAgent, produces an aged receivables report: who owes what, and how long it has been outstanding. Your bookkeeper looks at it on a Monday. The agent looks at it every morning, through the same connection the package already exposes, and it also knows each customer's agreed terms, because those live in the package too.

Then it acts, on a sequence you have written down. The day before an invoice falls due, a short note with the invoice attached, on the grounds that most late payment is forgetting rather than refusing. A week after the due date, a polite reminder. At fourteen days, a firmer note with a statement of account. At thirty days, it drafts an escalation and hands it to a person, because the decision to get serious belongs to a human. Every message goes out under a real name, and every one is logged back against the invoice so the history lives in one place.

It also knows when to stop. A customer who replies with a query or a dispute is taken out of the sequence that minute and flagged to whoever handles it. An account marked on hold is never touched. Your three biggest customers can be excluded entirely if you would rather chase them yourself. And on a Friday afternoon it sends the owner one message: what is outstanding, what has been promised and when, and who has gone quiet.

An agent handling money conversations needs boundaries, and they are easy to state. It never invents terms, discounts or deadlines; it uses what is on the invoice. It never threatens. It can calculate what you would be entitled to under the Late Payment of Commercial Debts (Interest) Act 1998, which is statutory interest at eight per cent over the Bank of England base rate plus a fixed sum of £40, £70 or £100 depending on the size of the debt, and it can mention that entitlement in the fourteen-day note. Whether you actually claim it is a decision for you, not for the agent.

It never has access to the bank. It reads invoices and contacts and writes emails and notes; it cannot move a penny, and it should not be able to. There is a switch that stops it, and a log of everything it sent. Those constraints are not hedges. They are what makes an owner comfortable letting it run, and comfort is the difference between an agent that ships and a pilot that gets quietly turned off.

The tools that already exist deserve a fair mention. Your accounts package almost certainly has automatic reminders built in, and dedicated credit-control tools cost between £40 and £400 a month, as we set out in what AI agents cost a UK business. If you invoice from one entity, your terms are the same for everyone and disputes are rare, start there and see how far it gets you.

A custom agent earns its keep when the rules are yours rather than the tool's: when terms differ by customer, when a job must be signed off in your job system before anyone chases, when a dispute needs to reach the right engineer, or when you run more than one company through the same team. We price it the way our AI agent cost calculator prices any agent: a fixed £3,500 discovery to write the rules down and prove them against your real ledger, then a monthly fee from £750 that covers hosting, model usage, monitoring and the tuning that comes from your team feeding back.

The return is arithmetic. If you typically have £60,000 outstanding and it is paid, on average, 45 days after invoice, then every one of those days is about £1,300 of your cash sitting in somebody else's account. Bring the average down by ten days and roughly £13,000 comes back for good. Against the fees above, that is the first year paid for before anyone has mentioned the hours your bookkeeper gets back.

Pull the aged receivables report today and count what is over thirty days. Write down the chase sequence you would like to run; most owners discover they do not have one, which is the real reason invoices drift. Decide who handles disputes. Then choose. If your package's reminders can run that sequence, switch them on this afternoon. If they cannot, the calculator will give you a number in two minutes, and a conversation with us will tell you whether the agent is worth building.