Choosing Your First AI Automation Project

Almost every UK business we speak to wants to "do something with AI" this year. The ones that succeed have one thing in common: they picked the right first project. The ones that stall picked a project that was too broad, too visible, or too far from real revenue.

Here is the framework we walk clients through on day one.

The four questions to ask about any candidate

For every workflow you are considering automating, answer these honestly:

Does it happen at least weekly? If not, put it aside. Automation of rare workflows rarely pays back.
Is the input reasonably structured? Emails, PDFs, spreadsheets and form submissions are fine — modern models handle them well. Voicemails, handwritten notes and one-off exceptions are much harder.
Is there a clear "done" state? "Draft this proposal" is clear. "Improve our marketing" is not.
Is the cost of a wrong answer bounded? If the automation makes a mistake, can it be caught and reversed cheaply? A wrong draft email is fine. A wrong wire transfer is not.

Workflows that score yes on all four are your shortlist.

The shape of a good first project

The best first project we have ever shipped, across dozens of clients, tends to look like this:

One team owns it.
It saves that team 5–15 hours a week.
The build takes 3–6 weeks.
The technology is boring and well-understood.
There is a clear human-in-the-loop checkpoint.

Notice what is not on this list: "revolutionary", "customer-facing", "strategic differentiator". Those come later, once the organisation has seen it work.

Common wrong first projects

Avoid these, at least until you have one success behind you:

A customer-facing chatbot on your marketing site. These are visible, so failures embarrass you. They are also surprisingly hard to make consistently good.
A company-wide AI assistant. Too broad. Nobody owns the definition of success.
Anything that requires perfect accuracy in a regulated context. Get the model working on lower-stakes tasks first, then earn your way into regulated ones.
A "let's replace [expensive tool] with our own AI version" project. Almost always more expensive than expected, almost never a differentiator.

How to measure honestly

The two numbers we track from day one on every automation:

Hours saved per week, measured by the team, not the vendor.
Error rate, measured against the ground truth of what a human would have done.

If hours saved is trending up and error rate is trending down after 60 days, keep going. If not, shut it down and pick a different workflow. Do not let sunk cost decide.

The right mindset

Treat your first AI project as a portfolio bet, not a bet-the-company move. You are not trying to transform the business in one shot. You are trying to prove that your organisation can identify a workflow, ship an automation, measure the result and iterate — because that capability, once you have it, applies to every future project.

FAQ

How long does a first AI automation take to ship?

A well-scoped first project should be live in production within 6 weeks. If you are being quoted 6 months, the scope is wrong.

How much should a UK SME budget for a first AI automation project?

Realistically £5,000–£15,000 for a well-scoped internal workflow, including build, deployment and 30 days of tuning. Payback within a quarter is normal.

Should we build in-house or use an agency?

For the first one, an experienced partner is almost always faster and cheaper. Once you have one live automation as a template, in-housing subsequent ones becomes much more realistic.

The first AI project is not the important one. What is important is that it succeeds, and that your team learns from it. Everything valuable comes from the second, third and tenth project — which only happen if the first one worked. Choose it with that in mind.