You want to try AI properly, not just play with a chatbot, but you do not want to bet the business on it. What should the first project be? A good first project is small, boring and measurable: one routine task, done behind the scenes, with a before-and-after comparison and a decision date agreed at the start. This guide is about choosing and sizing that project before it starts. Once it is running, our guide to why AI pilots stall covers how to carry it through to a clear decision.
Key takeaways
- Pick one routine, repetitive task with clear inputs, where an occasional mistake is easy to catch.
- Start behind the scenes rather than in front of customers.
- Measure how the task performs today before anything changes.
- Agree in advance what result means "carry on" and what means "stop".
What makes a task a good first project?
The best first projects are dull. They are the jobs that eat hours every week, follow a pattern, and that nobody would miss doing. Think of a 12-person accountancy firm answering the same client questions every week, or a trades business typing details from supplier invoices into its accounts.
Use this table to sort your ideas.
| Good signs | Warning signs |
|---|---|
| The task happens often and follows a pattern | It happens rarely, or every case is different |
| The information it needs is already written down or stored digitally | The knowledge lives in someone's head |
| You can tell a right answer from a wrong one | "Better" is a matter of taste or judgement |
| A mistake is easy to spot and fix before it does harm | A mistake could cost a customer, money or your reputation |
| It happens behind the scenes | Customers see the result directly |
| The people who do it now are willing to help test | The project is being done to them, not with them |
A task with most of the good signs is a strong candidate. A task with several warning signs may still be worth doing later, once you have learned from something safer.
Be wary of the showy option. A customer-facing chatbot can feel like the obvious first project, but it puts mistakes in front of the people you most want to impress. Sorting your inbox or matching invoices is less exciting and far easier to recover from.
Why measure before I start?
Because otherwise you will never know whether it worked. If you do not record how long a task takes and how often it goes wrong today, any result afterwards is a matter of opinion.
Before anything changes, write down:
- how long the task takes, per item or per week;
- how often it goes wrong, and what that costs to put right;
- who does it, and what else they could be doing with the time.
Then ask a second question: how would you fix this without AI? Sometimes a tidier spreadsheet, a template or a simple rule in software you already own gets you most of the way. That simpler option is the bar the AI has to beat. If it only does slightly better, at much higher cost, it is not worth it.
How big and how long should it be?
Small enough to finish, and short enough that you still remember why you started. A first project should cover one task, involve a small group of people, and have an end date a few weeks away, not a few months. No published source gives a typical length or cost for a small-business first project, so set yours by what you can afford to learn from.
Use real work and real data from the start. A project that only ever runs on tidy examples will look good and teach you little. The awkward cases are where you find out whether it helps.
If you are working with a consultant, ask for the first project to be paid and time-limited. That keeps both sides focused and gives you a clean point to decide what happens next. See how much an AI consultant costs for sourced UK figures.
Decide now what would make you stop
Before the project starts, write one sentence for each outcome: what result means "roll it out" and what result means "stop". Afterwards, any result can be explained away, so the time to agree is now. Running the project to that decision is the subject of why AI pilots stall.
Who needs to be involved?
Three people, even in a tiny business:
- An owner who cares about the result and can make the stop-or-scale decision.
- The people who do the task now, who know the exceptions and will tell you if it is making their day harder.
- Someone to set it up, which may be you, a member of staff or outside help.
The second group matters most. If they are not involved, they may quietly keep doing the task the old way alongside the new one, and your results will mean nothing.
What should I have at the end?
- A clear answer: carry on, change course, or stop.
- Before-and-after measures for time, errors and checking effort.
- A short note of what went wrong and how it was caught.
- Someone on your team who understands how it works and can adjust it.
- A copy of any settings, instructions or examples used to make it work.
- An idea of the next task to try, if this one went well.
What if the first project does not work?
Then it did its job. A small project that stops on schedule costs you a little and teaches you a lot about your data, your processes and what AI can do for you. If an earlier trial has already drifted, why AI pilots stall explains how to rescue it or close it cleanly.
If you want help scoping a first project, start with how to write an AI project brief and how to choose an AI consultant.