AI in a small company: the easy half of your week

A company owner’s week splits cleanly in two. There is the half that only you can do — winning the client, hiring the right person, deciding what the company does next year. And there is the other half: re-keying a supplier invoice, chasing a VAT deadline, writing the same quote for the fourth time this month, rebuilding a commercial proposal from a blank page on a Sunday evening.

Ask where to start with AI and almost everyone reaches for the first half — the impressive use case, the one that would transform the business. Start at the other end. The second half is where the hours actually are, where a mistake is caught the same day, and where nobody will fight you over territory, because nobody wanted the task in the first place.

The test

Before the list, the test. Four criteria, and a task has to pass all four.

It repeats. Weekly at minimum. A brilliant one-off costs more to set up than it will ever give back.

A human reviews before anything leaves the building. The AI drafts, prepares, sorts, extracts. A person clicks Send, signs, files. That single rule is what makes everything below safe.

The information already exists in writing. An email, a PDF, a spreadsheet, a recording. If the answer is only in your head, no tool can reach it.

An error is visible the same day and cheap to fix. A badly summarised email is caught on reading. A badly posted accounting entry is caught by your accountant, three months later. Only the first kind belongs here.

Nine tasks pass that test in almost every small company we work with.

1. The mail that actually needs you

Clients, suppliers, the bank, the accountant, the administration. A hundred messages a week, of which maybe six need you today.

Instead of scrolling, you ask what needs a reply, and get back a short list — with a first draft under each one. You correct, you send. Forty minutes of triage become ten minutes of deciding.

The trap — automatic sending. It should not exist in the tool. A human clicks Send, every time.

2. Invoices, delivery notes and quotes, turned into rows

The most re-typed documents in any company. A supplier invoice, a delivery note, an expense receipt: someone opens the PDF and copies six fields into a spreadsheet or the accounting tool.

That extraction is one of the things AI does best, precisely because the answer is already on the page — it is not inventing it, it is finding it. You get the rows, with the source document one click away for each line.

The trap — checking the total and not the lines. Spot-check against the PDFs for the first month. If nobody is prepared to do that, do not automate it.

3. The calendar of small deadlines

VAT, payroll cut-offs, contract and insurance renewals, certifications, the lease notice period, the subscription that renews for a year if you miss the date. None of them is difficult. Missing one is expensive.

You describe the recurring obligations once, and get a quiet nudge before each deadline instead of a scramble after it.

The trap — letting the AI compute the deadline. It reminds you; your accountant sets the date.

4. Notes that survive the year

The client meeting, the site visit, the call with the supplier. Someone records a two-minute memo on the way back to the car and gets structured notes: decisions, who does what, next steps — filed and searchable a year later, when the question becomes “what did we actually agree with them?”

This one pays twice: it saves the writing, and it saves the archaeology.

5. Questions asked directly of your own files

“Which of these contacts aren’t in the CRM?” “Which lines in this supplier price list went up since the last order?” Cross-checking two or three files is an afternoon of lookups, and it is asked in one sentence instead.

The trap — asking about files nobody has kept up to date. If the spreadsheet is three versions behind, the answer will be confidently wrong. The AI does not know it is reading an old file.

6. First drafts of routine writing

The quote, the standard reply, the payment chaser, the job advertisement, the client follow-up. None of it is creative writing, and all of it takes twenty minutes to start.

Starting is the part you hand over. The blank page costs more than the editing does.

7. The presentations — and this is the big one

The commercial proposal. The offer deck. The one-page summary for a prospect. For most owners this is the largest single time sink of the month, and the one nobody thinks to delegate, because “it has to look right”.

It now can be. You give the numbers and dictate what each page should say; a draft in your own branding comes back; you spend your time correcting a document that exists instead of building one from nothing. Same for the annual report, the price list, the memo that has to be presentable.

The general models from Anthropic and OpenAI have become very good at this. What matters is elsewhere: the drafting and layout are delegated to them, while the layer deciding which documents they may read — and what they may do with them — stays yours, which is the whole subject of our previous note — the model is replaceable, the layer around it is not.

The trap — the numbers. A beautiful proposal with the wrong price is worse than an ugly one. The figures come from your files and are checked by a person, page by page, before the client sees it.

8. Watching five subjects, not the news

Every company runs an informal watch: a competitor, a key client, a public tender, a regulation coming into force, sometimes its own name. It is done by whoever remembers, which means it is done unevenly.

Named properly — five subjects, not “our industry” — it becomes a short briefing each morning, sourced, with a link under every item, and silence on the days nothing happened. The value is as much in the silence as in the briefing.

The trap — a summary with no source. Anything you cannot open and read at the origin does not go into a decision. Demand the link.

9. The short research list

“The ten largest accounting firms in Lyon, with size and a contact.” “Every company in this sector within fifty kilometres.” A list you would have built in an afternoon arrives as a spreadsheet in your drive.

Useful for a first sweep, never for the decision itself: every line still has to be verified before it reaches a client or a proposal.

What we deliberately do not start with

Three categories, and they are the same three in every company.

Anything that pays, signs or files on its own. No payment, no order, no signature, no declaration. This is not a technical limit — it is a design decision, and it should be visible in the tool rather than promised in a meeting.

Anything that lives only in someone’s head. Half of what a small company knows is written nowhere: why that supplier was dropped, which client is never to be called before eleven. AI cannot reach it. Making it available is a writing project first.

Anything whose mistakes surface late. Accounting entries, tax positions, deadlines computed rather than looked up. The error is easy to make and expensive to find three months later.

The confidentiality paragraph, in short

Your company’s whole picture sits in a handful of mailboxes and drives. One rule covers the question:

The assistant sees exactly what the person it serves sees — and loses that access when they do.

Nothing is copied into a separate index to make it searchable; each document is read at the moment of the question, with that person’s own credentials. Removing an access takes effect immediately, because there is no copy to clean up. The wall between accounting and HR is the same wall as today, enforced the same way. We set out how that works, and the four questions to put to any vendor, here.

That leaves the question every careful owner ends up asking: are my documents used to train the models? On the professional plans from Anthropic and OpenAI, no — what you send does not join the training data, and it is written into the contract. On the free consumer versions, that is not always true. Check it before the first file goes out, and expect an answer that sits in a contract clause rather than in a sales promise.

Asking well is a skill, and it is taught

None of the nine tasks above requires technical ability. They all require the same single thing: knowing how to hand over a simple task. Saying what you expect, giving the context the other side does not have, and checking what comes back.

It is the exercise you go through with a temp on their first morning. And in a company, the people who get something out of it are not the ones most at ease with computers: they are the ones who can describe precisely what they want.

The good news is that this is learnt in hours, not months. The less good news is that it is not picked up by watching someone else do it. One properly trained person per team is enough to start the others off.

Within two years, handing a task to an AI will be an ordinary skill, no different from writing a clean email or keeping a spreadsheet. Better to learn it while it is still an advantage.

Starting Monday

Not a programme. One task.

Take the one that annoys the company most, write down what it costs today in minutes per week, and run it for two weeks with a person reviewing every output. At the end of the fortnight you keep it or you drop it — and you know which, because you wrote the number down at the start.

Then take the second. A company that has been through four of these has usually given a day a week back to two people, and has learnt, at no risk, exactly how far it wants to let AI near its business.

In short

The impressive half of your week is not where AI pays first. The paperwork half is: it repeats, it is written down, it is checked the same day, and nobody defends it.

Start with what nobody wanted to do anyway.