How to use AI — practically, as an SME

Start with a task, not with the technology

Most businesses approach this from the wrong end. Someone reads that AI is transforming business, a budget appears, a workshop gets booked, and six months later there is a pilot nobody uses and an invoice nobody enjoyed paying.

Turn it around. Do not ask “what can we do with AI?” Ask: “what do we do every week that eats time and mostly involves reading, writing or mindless clicking?” That question has answers in every business — the quote rewritten from scratch each time, the contract nobody has read properly, the report that takes a morning to explain, the procedure that lives only in one person’s head, the same fields updated by hand across two systems.

Two quick tests tell you whether a task is worth handing over:

The intern test. If you catch yourself thinking “I’d give this to an intern” — or better, “I’d give this to an intern if they stayed longer than a few months” — you can usually explain it to AI once and have it done forever. These are the repetitive jobs where people slip up because they are bored or tired, not because the work is hard.

The smile test. A good candidate leaves you smiling for one of three reasons: you are glad you never have to do it again; you are glad you no longer have to babysit the sub-par output you used to get; or you finally finished the project you were 85% equipped for but never got round to the last 15%.

Start there, because you can measure it. You know roughly what these cost you today, so you will know whether anything actually improved.

What it is good at, and what it is not

It helps enormously to know what you are holding. Today’s assistants are, at heart, extremely capable language machines — and, increasingly, capable of writing and running code. That makes them very good at a specific set of things:

• Turning long into short — a forty-page supplier contract into the five clauses that actually affect you.
• Turning short into long — a few bullet points into a first draft of a letter, procedure or offer.
• Turning messy into structured — a rambling voice note or a page of meeting scribbles into a clear list of decisions and owners.
• Turning specialist into plain — a financial concept explained to your operations manager in language they will actually use.
• Turning one language into another.
• Turning manual into automated — pulling data out of the tools you already pay for, wiring two systems together, or scripting away an afternoon of repetitive clicking.

And there are two things it is genuinely bad at, which matter more than the marketing suggests. It does not reliably do arithmetic. And it does not know when it is wrong — you get the same confident, well-written tone whether the answer is correct or invented. A human who is unsure sounds unsure. This does not.

That combination is manageable once you know about it. It is expensive if you do not.

Where it earns its keep in a small business

Concretely, in a company of five to fifty people, this is where the time actually comes back:

The document you have been meaning to read. Paste in the lease, the supplier agreement, the insurance policy. Ask what the notice period is, how the price indexation works, what happens if you terminate early. Then read those clauses yourself in the original. Ten minutes instead of an hour, and you check the two paragraphs that matter rather than forty pages.

The process that lives in one person’s head. Talk it through out loud, have the assistant turn it into a draft procedure, then correct the parts it guessed wrong. Getting from a blank page to a rough draft is the hard part, and this removes it entirely.

The writing nobody enjoys. The offer you rewrite from scratch every time. The awkward email to a late-paying customer that you have drafted four times. The job posting. The polite reminder in German.

The marketing and sales you never get to. When the extra campaign, the follow-up email or the social post still hinges on you as founder, AI turns a few bullet points into a decent first draft and multiplies how much you can put out — so the growth work stops waiting for the one free evening you never actually get.

The report that needs explaining. Give it the figures and your notes and let it draft the commentary — the paragraph that says what happened and why. You correct the interpretation, which is quick, instead of writing from nothing, which is not.

The three lines of code. The code you cannot write yourself, but would never have called a programmer for — let alone paid €150 an hour for. AI can write the API calls for you, or even set up an automation to extract the data. Many SMEs have useful data sitting somewhere that is not feasible to get out by hand — and up until a few months ago, nobody wanted to instruct and pay a programmer to write one API call.

Notice what these have in common: a human still decides. The machine removes the blank page and the donkey work, not the judgment.

Let AI guide you past the roadblock

Sometimes the value is not doing the whole task — it is getting past the one step that stops you. You can do seven of the eight things a job needs, but one is a wall: a tool you have never set up, a file format you do not know, an integration you have never touched. That is how good projects stall at 85%.

Ask AI to be the guide. Tell it plainly what you are trying to do, ask whether it is even feasible, and have it walk you through step by step. Crucially, tell it to work in steps and check with you at each one — that keeps a plausibility check running the whole way, so you catch a wrong turn early instead of at the end. Used like this, an assistant is less a magic button and more a patient colleague who has done this particular thing before, and it is how the “last 15%” projects finally get finished.

The rule for anything with a number in it

This is the part we care most about, because it is where an SME can genuinely get hurt.

Use AI for language about numbers. Do not use it as a calculator. Ask it to explain why a margin moved, to draft the note that accompanies a report, to structure a comparison — yes. Ask it to add up a column, apply a VAT rate, or work out a loan schedule and take the answer on trust — no.

The rule we give clients is short: no number reaches the tax office, the bank, an investor or a customer without a human having checked it. Not because the tool is useless, but because “the software said so” has never once been an accepted explanation for a wrong return.

Newer tools handle this better than they used to — some will write and run actual code, or work directly in a spreadsheet, rather than guessing, which is far more reliable. Prefer those for anything numerical. Then still check the result.

Where not to paste your data — and how to wire it up safely

The convenient thing about these tools is that you can paste anything into them. That is also the problem.

Free consumer accounts may use what you type to improve the underlying models. That is fine for a birthday message and entirely inappropriate for a personnel file, a customer list, a salary overview or a ledger export. Paid business and team plans generally do not train on your data and give you proper user administration — for roughly the price of a lunch per user per month, which is not a serious obstacle. Under the GDPR your company remains responsible for personal data regardless of which tool an employee happened to use, so set one plain rule everybody can remember: company work goes in the company account, and if you would not email it to a stranger, do not paste it into a free chatbot.

The same care applies when you go a step further and let AI act on your systems directly, through their APIs. Give it read-only access wherever reading is enough — a dashboard only needs to look. Where it genuinely has to write back — updating records, changing master data — go read-first, write-later: have it show you exactly what it intends to change, and keep a human approving the write step. An automation that can only look is a small risk; one that can alter your data needs you in the loop.

Garbage in, confident nonsense out

Here is the uncomfortable part, and the reason this guide sits on a finance firm’s website.

AI amplifies whatever you already have. Feed it well-structured monthly figures — revenue split by the streams you actually manage, costs grouped the way you think about them — and it will help you interpret them, spot patterns and explain them to your team. Feed it a chart of accounts where two thirds of your spending lands in “general costs” and it will still give you an answer. It will be articulate, plausible and worth nothing, because the information required to answer properly was never in the data.

This is why we are unromantic about the order of operations. Get your bookkeeping structured around how the business actually runs first. Then the clever tools have something real to work with. Skipping that step does not save you the work; it just moves the error somewhere harder to see.

How to start: one task, one person, two weeks

You do not need a strategy, a consultant or a committee. You need one honest experiment.

• Pick one task that happens weekly and involves reading, writing or repetitive clicking. Not the most important one — the most repetitive one.
• Pick one person who is mildly curious rather than the most senior person available.
• Buy one business seat. Give it two weeks of real use on that one task.
• Write down, roughly, how long the task took before and how long it takes after. Rough is fine; nobody needs a stopwatch.

At the end you will know something specific rather than something fashionable. If it saved real time, widen it to the next task or the next person. If it did not, you have spent about the cost of a client lunch to find out — and you can drop it without a sunk-cost argument. One condition throughout: nothing leaves the building unchecked. No email sent, no figure filed, no procedure adopted without a human reading it first.

Examples

A few that sit close to the finance and admin work we do every day:

A live dashboard from your bookkeeping. Have an assistant help you pull your accounting data into a dashboard you actually read — then enrich it with figures from your sales system, so revenue, margin and pipeline sit on one screen instead of in three exports.

Cleaning up master data. Updating the boring-but-important fields — supplier and customer payment terms, say, after you have tightened your cash flow — across a whole system, without an afternoon of mindless clicking.

A single prompt can get you a surprisingly long way. Something like:

“Read the online documentation for my bookkeeping program [X], find its API documentation, and guide me step by step to set it up. The result should be an Excel of every vendor and their payment terms. Let me edit that Excel, then you write the changes back into [X].”

Hours of clicking and saving become one quick update — with you approving the write step, exactly as above.

Worked example: how this website was built

The best example is the one you are reading. This site — the pages, the guides, this article — was built by Thorsten Schmidt together with Claude, the AI assistant. It is a fair, unglamorous picture of what practical AI actually looks like.

Step by step, page by page. There was no grand build and no big-bang launch. Each page — the home page, the contact page, every guide — was drafted, reviewed, programmed and put live one at a time, exactly the “one task at a time” approach this guide argues for.

AI and human: drafted and reviewed. Every guide was drafted by a human and an AI, then consolidated, read, corrected and re-shaped several times by a human who knows the subject. The AI removed the blank page and did the donkey work; the human supplied the judgment, the real numbers and the truth about the business. Neither would have produced this alone.

Programmed and launched — Claude where possible, a human where needed. Claude wrote the code, the structured data and the layout, and handled the fiddly technical work. The human made the calls a machine cannot and should not: what is actually true, what goes live, and the account-and-domain steps that require a real person signing in.

The part that genuinely impressed. Over successive guides, Claude learned the subjects, what matters to SMEs, and the tone of voice from the edits it was given, and carried it into the next piece. Mistakes flagged once did not come back the next time. That is the real story of practical AI: not a magic button, but a collaborator that gets measurably more useful the more you correct it — exactly what you would want from a sharp junior colleague.

The honest footnote: this is a marketing website, not the general ledger. The same discipline we preach still applied — a human checked everything before it went live, and no client’s numbers went anywhere near a chatbot to make it.

Frequently asked questions

Let’s actually start

None of this is difficult. You could pick the task, buy the seat, run the two weeks and draw your own conclusions. But sometimes starting is the hardest part — especially with nobody to bounce ideas off, and plenty of “specialists” claiming to automate your whole business in no time, for a fee of course.

If you want help getting started, and an outside perspective on how to start profiting from AI in your company tomorrow — practically, one step at a time — emailing us is a good place to start.