An AI agent that reads your till: what daily sales tell you

Your till records every sale, every voided ticket and every discount someone keys in at seven in the evening. It knows everything. The trouble is that hardly anyone looks at it, apart from cashing up and, if you're lucky, the monthly summary the software emails you and nobody opens.

This piece is for the owner of a shop, a bar, a salon or any business with a point-of-sale system who suspects there's something useful in there but has neither the time nor the appetite to build a dashboard. By the end you'll know what kinds of questions an AI agent can answer by reading your daily sales, what it needs to do that well, and when it won't tell you anything worth knowing.

Why this is different now

Until recently there were two ways to get anything out of your sales data. You did it yourself, exporting to Excel and wrestling with pivot tables, or you paid for a reporting tool full of charts you then had to learn to read. Either way, the interpreting was your job.

What's changed is that current AI models can read a table of figures and explain it in plain language with real fluency. Nobody has to design a chart for each question: the agent can look at yesterday's takings, compare them with the same day last week and send you three lines on your phone. The format stops being a panel you have to go and check and becomes a message that comes to you.

That shift matters more than it sounds. A dashboard only helps if you open it, and in a small business nobody opens it on an ordinary Tuesday. A short message at nine in the morning does get read.

Five questions it can answer every morning

The point isn't for the agent to analyse everything. It's to flag the little that deserves your attention. Some realistic examples:

1. Was yesterday a normal day? Not the total, which you already see when you cash up, but the comparison: "Tuesday was down 30% on the previous three Tuesdays, mostly in the afternoon." Now you know whether it was the rain or something else.
2. What's stopped moving? Products or services that used to sell every week and haven't gone through the till in a fortnight. In a shop that might be stock or shelf placement; in a beauty clinic, a treatment nobody is offering any more.
3. Is anything odd in the tickets? Voids or discounts above the usual, refunds bunched in one shift, cash payments suddenly dropping. This isn't about accusing anyone. It's about finding out before your accountant does at quarter end.
4. When are you losing people? Hours where you sell little with the whole team on, or peaks where the queue is long enough to dent the average ticket. Handy for planning rotas on evidence rather than gut feel.
5. Did the thing you tried actually work? If you launched a new menu or a two-for-one last week, the agent can tell you how often it sold and whether it pulled other sales along or just replaced what people usually buy.

A Monday message might be no more than this:

Week of 2–8: sales in line with recent weeks. Saturday morning was the best stretch of the month. Heads up: the 200 ml sun cream hasn't sold in 18 days and you have 14 in stock. Three voids on Thursday's afternoon shift, above normal.

Four lines. If something catches your eye, you ask, and the agent digs in.

What it needs to get this right

This is where the projects that work part company with the ones that never get past a nice demo.

Access to the data, read-only. Many cloud-based tills let you download sales as a file or connect to other software through an API (a doorway built so two apps can pass data to each other). Tills installed on a computer in the shop tend to be more closed, and sometimes a daily export is the only route. Either way, the agent should be able to read and never change or void anything. It's a simple rule, and worth writing down.

A reasonably tidy product list. If the same item appears as "latte", "LATTE" and "misc 1.80", the agent will do its best, but its conclusions will only be as good as your till menu. It doesn't have to be perfect; your important products just need a consistent name.

Context the till doesn't have. That you were shut on the 12th for building work, that you close for two weeks in August, that the December bank holiday always flattens the Tuesday. Without this, the agent will warn you every year about a "drop" you already understand. A short list of special dates, updated now and then, is enough.

A few well-chosen questions. The worst outcome is a morning message that grows into a two-screen report you stop reading within a week. Start with two or three questions that would genuinely change a decision, and add more only when you find yourself missing something.

What the agent shouldn't do

The agent reads and summarises; your official numbers still come from the till. If you need the exact quarterly turnover for your accountant, pull it from the system, not from a sentence written by an AI. A language model can get a sum wrong or misread a column, so its alerts are prompts to look, not figures to file.

Nor should it draw conclusions about people. "Thursday afternoon voids are above normal" is a fact that invites you to check. "So-and-so is stealing" is something a machine should never write. That interpretation, and whatever conversation follows, is yours.

When it isn't worth it

To be fair, not every business gains from this:

• You sell little, and always the same things. A business with twenty tickets a day and ten items on the menu already knows itself by heart. The agent will tell you what you know.
• Your till won't let the data out. If the only option is photographing the paper Z-report, sort that first. Sometimes the answer is a different till, and that decision goes well beyond AI.
• Someone already does it. If your manager goes through the sales every Monday and tells you over coffee, the system may already work. Automating it only makes sense if she'd rather spend that time on something else.
• You won't act on it. An alert that changes no decision is noise. Before building anything, ask yourself what you'd do differently if you knew a product had sat untouched for two weeks.

Starting without a big project

The sensible move is to try it by hand before automating anything. Export last month's sales, give them to an AI assistant and ask the two or three questions you care about most. If the answers surprise you, or save you a decision made in the dark, you know it's worth wiring up so it lands on its own every morning. If they tell you nothing new, you've saved yourself a project.

When it's time to connect it properly, it usually has to fit your particular till, the way your product list is set up, and what you want to receive and on which channel. That's exactly what we do with process automation: we start from your data as it is, and you decide what reaches you and what doesn't.