5 November 2026 · 4 min read
How to choose your first AI project
Most companies that want to start with AI have a long list of ideas and no good way to rank them. The first project matters more than it seems. A good one earns trust inside the company and funds the next step. A poor one leaves people sceptical for a year. This is how we would choose.
Look for work that is frequent and costly
Start with the work people repeat every day. Answering the same kinds of questions, sorting incoming requests, reading documents to pull out a few facts, preparing a first draft of a reply or a report. If a task happens once a quarter, automating it will not change much, however clever the solution.
Cost matters as much as frequency. Count the hours, but also count the delays and the mistakes. A task that is only a few minutes long but blocks an order until someone finishes it is expensive in a way the timesheet does not show.
Prefer text and documents
Current AI systems are strongest where the work is made of language. Emails, contracts, invoices, tickets, product descriptions, reports, internal knowledge scattered across files. If your team spends its day reading, writing and cross-checking text, there is probably a good first project nearby. If the work depends mostly on physical judgement or on rare, high-stakes decisions, it is a worse place to begin.
Check that the data is available
An AI system is only as useful as what it can see. Before choosing, ask where the information lives, who can open it, and in what shape. Documents in a shared folder are fine. Answers that exist only in one person’s head are not yet data. Sometimes the first step of a project is simply getting the material into a place the system can reach, and that is worth knowing before you commit to a plan.
Agree on one measurable outcome
Decide before you build how you will know it worked. Fewer hours on a task, a shorter wait for a customer, fewer corrections after review. Pick one or two numbers you can read today, so that you have a baseline to compare against. Without a baseline, every result becomes a matter of opinion, and the project is judged on how people feel about it that week.
Name one person who owns the process
Every successful project we can describe has one thing in common. Someone on the client side knows the process well, cares whether it improves, and has the authority to say what correct looks like. That person reviews the early results, spots the cases the system gets wrong, and decides when it is ready. A project with no owner drifts, because nobody can answer the question of whether the output is good enough.
Start narrow
Pick one process, one team and one kind of document. A narrow scope lets you finish, learn and show a result in a short time. It also makes errors easier to understand. Once the first piece works, widening it is mostly a matter of adding cases, and you do it with evidence instead of hope.
A useful test is whether you can describe the project in two sentences, including who uses it and what they stop doing by hand. If you cannot, the scope is still too wide.
What to avoid
Some patterns repeat often enough to name.
- Starting with the technology. A project chosen because a tool looks impressive tends to look for a problem afterwards. Begin with the work that hurts.
- Choosing the most visible process. The one the board talks about is often the most complex and the most political. Choose something useful that can go well quietly.
- Aiming at full automation on day one. Let the system prepare and a person approve. You get most of the benefit, you stay in control, and you learn where the system is weak.
- Ignoring the people involved. Those who do the work today know its exceptions. Involve them from the start, and make clear that the goal is to remove tedious work, not to judge their performance.
- Treating it as a one-off. Systems that touch real work need someone to look after them once they are live. Decide who that is before launch.
A short checklist
A good first project is frequent, costly and based on text or documents. The data is reachable. There is a number to move and someone who owns the process. The scope fits in two sentences, and a person stays in the loop at first. If a candidate ticks most of these boxes, it is a sound place to begin.
If you are weighing a few candidates and want a second opinion, we are glad to talk it through. The first conversation is free and non-binding. You can reach us through the contact page.