Your first AI project: how to pick one that pays back

The question we hear most from business owners isn't "what is AI?" It's "where would it actually help us?"

That uncertainty shows up in the numbers. UK adoption is climbing quickly: the British Chambers of Commerce found 54% of firms actively adopting AI this year, up from 35% a year earlier. But Simply Business found that only 19% of small business owners feel “very confident” using it day to day, and 39% aren’t sure how it could be useful to them. Many firms use AI for writing and general questions, and few have used it to change how work actually gets done.

The gap isn’t ambition. It’s knowing which task to point it at.

Start with the task, not the tool

The most common mistake is starting with a tool (“we should get Copilot”) and then looking for uses. It works better the other way round. Make a list of repetitive tasks, then ask which of them AI is genuinely good at.

AI is reliably useful when the work involves reading, sorting, extracting or drafting:

• Reading documents: supplier invoices, delivery notes, purchase orders and contracts, pulling out the figures and dates that matter.
• Sorting messages: classifying incoming email into enquiry, complaint, invoice or spam, and routing each to the right person.
• Extracting data: turning messy, inconsistent input into tidy fields your systems can use.
• Drafting: writing a first reply or summary for a person to check.

It’s much less reliable when the work needs exact calculation, accountability or a judgement call that only your business can make. That’s where a person should stay in charge.

A five-question filter

Run each candidate task through these questions. The ones that score well are your shortlist.

1. How often does it happen? Daily and weekly tasks add up. Annual ones rarely repay the effort.
2. How long does it take each time? Ten minutes forty times a week is a lot more than it sounds (we do that sum here).
3. Is there a clear right answer? “Extract the invoice total” has one. “Decide our pricing strategy” doesn’t.
4. What’s the cost of a mistake? If a wrong answer is cheap and easy to catch, AI is a safe fit. If it’s expensive or invisible, add a human check or leave it alone.
5. Is the input data in decent shape? AI amplifies what you feed it. Tidy data in, useful output out.

Three projects that tend to pay back

These come up most often in small and medium-sized businesses.

1. An enquiry agent. Reads incoming emails, looks up products, prices and account details, and drafts the reply for a person to approve. We built one for a timber merchant (see the case study).

2. Supplier document capture. Reads supplier invoices and delivery notes and puts the data into your accounts, so nobody keys it in by hand.

3. Weekly reporting. Pulls numbers from your systems, writes a plain-English summary and emails it out before Monday’s meeting.

None of these replaces anyone. They remove the dull middle of the job so people have time for the part that needs them.

Keep a human in the loop at first

Starting with drafts rather than fully automatic actions has three advantages. Your team sees what the system does and learns to trust it, mistakes get caught before customers see them, and you gather the evidence to decide what can safely run unattended later.

Be sensible about data

Security and privacy are the top reason non-users give for holding back, and it’s a fair concern. Before any AI project, be clear about:

• what data the system sees;
• where it’s processed and stored;
• what’s logged, so you can check what happened;
• who can see the outputs.

A good build keeps your data in accounts you own and sends the minimum necessary to the AI model.

Measure it

Agree what success looks like before you start: time saved per week, replies sent faster, errors reduced. Then measure the same thing after. If it hasn’t helped, you should know.

Not sure where to start?

That’s exactly what a discovery sprint is for. In two weeks we map your processes, cost the manual work and tell you which tasks are worth automating with AI, which are better served by simple automation, and which to leave alone. You keep the plan either way. Get in touch to talk it through.