Where to start with AI in a 20-person business

Most AI advice written for SMEs is really advice for large companies, shrunk. It assumes a data team, a transformation budget and eighteen months. If you run a twenty-person business in Leicester or Wolverhampton, none of that describes your week.

Here is a sequence that fits the business you actually have.

Weeks 1–3: find the expensive problem

Do not start with the technology. Start by finding the thing that costs you most in time and is most repetitive, because that combination is where current AI tools are genuinely strong.

Ask three questions of every department:

• What do you do every week that feels mechanical?
• What do you wait on, and who are you waiting for?
• What do we get wrong often enough that we have built a check for it?

The answers are usually unglamorous: quoting, goods-in paperwork, chasing purchase orders, rewriting the same specification for the fifth customer, finding out which drawing revision is current. That is fine. Unglamorous problems have measurable costs, which makes them the right place to start.

Weeks 4–5: measure the baseline

This is the step almost everyone skips, and skipping it is why so many SMEs cannot tell you whether their AI spend worked.

Pick the one problem you will tackle and measure it for a fortnight. How many hours? How many errors? How long from enquiry to quote? Write the numbers down before anything changes. They do not need to be precise to three decimal places — they need to exist.

Without a baseline, every later conversation collapses into opinion.

Weeks 6–10: build one narrow thing

One problem. One team. One measurable outcome. Resist every temptation to widen the scope, because scope is what kills SME pilots — not technology.

Narrow looks like: "the estimating team drafts quotes from the enquiry email and the drawing, and a human checks every one before it goes out." It does not look like: "we are rolling out AI across the commercial function."

Keep a person in the loop on anything that reaches a customer. Not because the technology cannot do it, but because a wrong quote sent automatically costs you more than the automation saves.

Weeks 11–12: compare, then decide

Put the new numbers next to the baseline. Then make one of three decisions, explicitly:

• Keep and extend — it worked, roll it wider
• Keep as is — it pays for itself where it stands, leave it
• Stop — it did not pay, switch it off and say so

The third option is the one that builds credibility. A business that has visibly killed one AI project is far more likely to get support for the next one than a business with four half-live pilots nobody will pronounce on.

What to avoid in the first 90 days

Do not buy a platform. Platform decisions made before you understand your own workflows are expensive to unwind. You can pilot almost anything on a monthly subscription.

Do not hire for it yet. A twenty-person firm rarely needs a full-time AI person before it has proved there is a job to do.

Do not start with the customer-facing thing. It is tempting, because it is visible. It is also where mistakes are most expensive and where trust is hardest to rebuild.

Do not let it become a side project for someone already at capacity. If nobody has time allocated, it will lose to whatever is on fire that week, every week.

The honest expectation

Ninety days will not transform your business. It should give you one working improvement, a number that proves whether it paid, and an internal understanding of what these tools are actually good at — which is worth considerably more than a strategy document nobody reads.