AI in your company without launching a project

Bringing AI into a small company doesn't have to start with a big consultancy or a custom platform. I describe the three small uses I see working in real small businesses, how to get them going in a week and what return they give, with a simple way to measure whether it pays off.

I come across two kinds of company. One believes that bringing AI into a small business requires a consultancy, a three-month assessment and a platform with an English name. The other has told the team "use ChatGPT" and left it there: no rules, no idea who uses what or what data has walked out the door.

Both are wrong, and the second is also taking risks.

And between the two there's plenty of ground: a few hours from someone who knows, to set some rules, business accounts with admin controls, a limited trial with the IT supplier you already have. I'm not saying advice is unnecessary; I'm saying it doesn't have to be the first step, and the controls can be added as usage grows.

I'm speaking from what I do: I apply these tools on my own systems and on clients' systems, mostly companies of between 5 and 30 people: a renovation firm, an advisory firm, an online shop and an architecture practice. I won't give names, but I will tell you what has worked and what hasn't.

What you'll find: three concrete uses, a one-week plan you can copy and a way to know whether it was worth it. What you won't find is magic.

The central idea is simple: one specific, repetitive and measurable task, tested for a few weeks with tools that already exist. That's all.

AI in a small company usually comes in at three levels:

1. Each person's chat: someone uses Claude, ChatGPT or whichever to draft, summarise or think.
2. The tools you already have, like email, office software or your customer program, which increasingly come with AI built in.
3. Automations or custom tools, which are already a project.

Almost every small business should start with the first two. Not because the third is bad, but because without having tried the others you don't know what to ask for. And getting it wrong on a small scale is cheap: a subscription and a few hours, not a consultancy invoice nobody wants to look at afterwards.

There's another advantage that rarely gets noticed. When someone on the team has used the assistant for a few weeks on one of their own tasks, they understand what it does well and what it does badly better than any presentation could show. That person is the one who then tells you, knowing what they're talking about, whether it's worth going further. At two of the clients, the idea for the next step came neither from me nor from the manager, but from the person who'd been using it for a month.

The three uses I describe below have something in common: they repeat every week, they can be measured with a clock and, if they go wrong, the error is caught before it reaches the customer.

The use I see working most is drafting and answering emails, especially the questions that keep coming up: opening hours, lead times, order status, what a service includes.

How it's done: one person has a document with the usual good answers and gives it to the assistant along with the email that has just arrived; the assistant proposes the reply and the person corrects it and sends it. With Claude on a team account. It's used by whoever handles customer email.

The case: an online office supplies shop, four people. They received about 40 emails a day, half of them with the same five questions. The person in charge went from spending the whole morning on email to getting through it in an hour and a half.

What always needs checking: that the reply doesn't promise what the company doesn't offer. On the first day the assistant offered a discount for the inconvenience that nobody had authorised. Since then, the instruction explicitly says no compensation is offered.

Return: about 8 hours a week. And one I didn't expect: the replies come out more orderly and friendlier than before, when they were written in a hurry between orders. A couple of customers commented on it. I can't measure that with a clock, but I'm noting it.

The second is preparing quotes and proposals. I cover it in detail in Invoices and quotes with the help of AI, so here I'll get straight to the point.

What's done: the assistant receives the client's request, the price list and a previous quote as a model, and prepares a draft with line items and wording. The maths is done in a spreadsheet, not in the chat. With what: Claude and the usual spreadsheet. Who: the admin person, with the technical manager reviewing.

The case: the renovation company, eight people, about 25 quotes a month. Each one went from almost an hour to about 20-25 minutes, review included.

What to watch out for: without the price list, the model makes up "reasonable" prices that aren't yours. And what you promise in a quote commits you, whoever wrote it.

Return: about 12 hours a month and, according to them, quotes sent sooner, which has won them some jobs that used to be lost through slowness.

The third is summarising: meeting minutes, long contracts to pull out the key dates, regulations that need reading.

What's done: the meeting is recorded with the attendees' permission, transcribed, and the assistant extracts decisions, owners and dates. With what: the video call tool's transcript and Claude for the summary; in some cases NotebookLM to query several documents at once. Who: whoever used to take notes.

The case: the architecture practice, twelve people, three site meetings a week. The minutes went from being written "whenever possible" to being sent the same day.

Here I have to tell you about a failure. At the advisory firm I tried summarising client contracts and dropped it after two weeks: the summaries were good, but they had to be reviewed so thoroughly, given what was at stake, that no time was saved. And on top of that it was third-party data that required a different account and different safeguards. Not every use works everywhere.

Return where it worked: about 4 hours a week.

If you're wondering how to implement AI in your company, this is the plan I follow, day by day:

Accounts: business ones, not free personal ones. And the output is reviewed by the person who knows the task, not by whoever set up the test.

Before starting, I put five things in writing, on one sheet, not in a manual:

• An owner: a named person, usually the one who does that task most, who answers questions and decides whether to extend it.
• Approved tools: which ones can be used and with which account. Anything not on the list isn't used with company data.
• Which data goes in: three levels. Public data goes in; internal data, only on the business account; personal or client data doesn't, unless the contract and data protection are sorted.
• Training: the half-hour I describe below and a review together after two weeks.
• A way back: the previous process isn't deleted. If the trial goes badly, you go back to it the next day. That's what happened at the advisory firm with the contracts.

How I explain it to the team: in a half-hour meeting, showing the use with a real case of theirs, not with a presentation. And making it clear that it's not here to replace anyone, just the boring part of the job. If the team doesn't understand what it's for, they won't use it.

I measure four things: time per task, how many drafts are accepted as they are or with a small touch-up, errors caught in review and, when customers are involved, how long it takes to reply. The starting point is measured before changing anything. Without measuring before and after it's impossible to know whether it pays off; it stays at "it seems to be going better", which is what everyone says about everything.

The figures from the three cases:

Two things that don't fit in the table. Acceptance: with the emails, about seven out of ten drafts went out with a small touch-up or none; with the quotes, almost all needed adjusting, but in the wording, not the prices. And errors: in the first month, review caught the made-up discount I mention above and a couple of wrong delivery dates in the emails; in the quotes, the odd line item that wasn't on the price list. None reached the customer, which is what the review is for.

What it costs to keep going: the team account subscriptions and a little time each month to update reference documents and instructions. And something that doesn't show up in the tables: someone has to take ownership. When the person driving it goes on holiday, usage drops.

A note on how I measure, because it's easy to fool yourself here. The "after" time always includes review: if the assistant writes in two minutes something that then needs twenty minutes of correcting, the saving is whatever's left over, not what the chat's stopwatch says. I log it in a simple sheet, one row per task, for two weeks before and two weeks after. It's not science, but it's much better than everyone's impression, which tends to be optimistic in the first month and pessimistic in the third.

There are signs that the time has come to invest more: the small use works and falls short, there's your own data worth making use of, or a process repeats so many times that automating it pays off. Then yes, a project. But with what you've learnt.

And the usual warning: when personal data is involved, data protection comes first. When in doubt, ask someone who knows.