A shop owner asked me last month whether AI could tell him which of his products were actually making money. It is a good question. It is the right question. And the honest answer was no - not because the technology cannot do it, but because nowhere in his business was anyone recording what he had paid for the things he was selling.
There was no AI problem to solve. There was a bookkeeping problem wearing an AI costume.
What we filmed, and why so little of it is AI#
We published a full walkthrough of a shop running on one system - counter sales, stock, serial numbers, supplier deliveries, M-Pesa, tax invoices, and the evening reports. It runs about an hour and twenty minutes.
Almost none of it is artificial intelligence. It is a shop day being written down carefully. I am putting it on an AI website on purpose, because it is the part that everybody wants to skip and nobody can.
The four questions, and what each one needs#
Owners tend to ask AI the same handful of questions. Here is what each one quietly requires the business to have been recording all along.
| The question | What has to exist for an answer |
|---|---|
| What sold today? | Every sale captured at the moment it happens |
| Which items make money? | The buying cost of each item, not just its price |
| What is not moving? | Stock levels, with dates, per location |
| Who owes me? | Sales linked to named customers, with payments matched |
| Is this unit ours? | A serial number recorded when it arrived and when it left |
Look at the middle column. Not one line of it is AI. All of it is the boring part of running a shop, done in a place a computer can read.
A business at Stage 0 - sales in a book, stock counted by eye, the real record in the owner's head - can buy the most capable AI in the world and it will still answer the second question with a shrug or, worse, with a confident guess.
Why "we have the data in WhatsApp" does not count#
This is the most common objection and it deserves a straight answer rather than a lecture.
Yes, the information exists. Yes, it is technically digital. But data that is scattered across chat threads, photographs of delivery notes, and three different phones is not a record - it is evidence. Somebody still has to gather it, interpret it and decide what it means, and that somebody has been you, at night, after closing.
What changes at Stage 1 is not that the information appears. It is that it stops needing to be gathered, because it was already in the right place when it was created.
An assistant that has to guess will guess. It will sound just as certain as one that knows.
We have written about why AI gives you confidently wrong numbers, and the root cause is almost never the model. It is being asked a question the business never made answerable.
The part that surprises people: it pays for itself before the AI arrives#
Here is why this is not a discouraging article.
Everything you would have to put in place to make AI useful is worth doing on its own merits. A shop that records the cost of every item finds out which lines are carrying the business and which are wasting shelf space. A shop that tracks serial numbers stops arguing about warranty claims. A shop that matches M-Pesa payments to invoices stops chasing customers who already paid and starts chasing the ones who have not.
None of that requires AI. All of it saves money. And when you do add an assistant on top, it has something true to stand on.
What "adding AI" looks like once the record exists#
Once the shop day is being written down properly, the assistant stops being a novelty and becomes something narrow and genuinely useful. Asking, in plain language, which items sold below their usual margin this week. Getting the morning summary without opening a report. Asking whether a specific customer still owes anything before extending them more credit.
That is what Rai is for, and it is deliberately unexciting: it reads the records the business already keeps and answers questions about them. It cannot tell you what your margin was in a month when nobody recorded a cost, and it should not pretend otherwise.
The same applies to where that data lives. A record worth building is a record worth keeping on your own server, which is a decision best made at the point you start recording rather than after two years of history has accumulated somewhere else.
The honest summary#
If you are running a shop from a notebook and your memory, you are not behind on AI. You are at the beginning of something that has a correct first step, and the first step is not buying an AI tool.
It is making sure that tonight's sales, this morning's delivery, and the price you paid for the thing you just sold are all written down in one place. Do that for a month and two things become true: you will already be running the business better, and every AI conversation you have after that will be about something real.
Watch the walkthrough if you want to see what "written down properly" actually looks like in practice. It is eighty minutes of unglamorous work, and it is the foundation of everything else we build.
