Writing · Practical AI
Your first AI project should be embarrassingly small
The instinct is to plan the whole system. The businesses that get somewhere ship one small thing, live with it a fortnight, and let it change the plan.
Ricky Bell · 6 August 2026 · 4 min read

When a business finally decides to do something with AI, the instinct is almost always the same: work out the whole system first. Map the processes. List everything it should handle. Get it right before spending any money.
I understand the instinct — it is how you would approach a new van or a new hire, and it is a sensible instinct in both those cases. Here it is the thing that kills the project. Not because planning is bad, but because you are planning against assumptions you have no way of testing yet, and roughly half of them are wrong.
One shipped thing beats ten planned ones
The pattern I would follow, and the one I used building our own audit tool, is to pick the single core piece and build only that. Not the version with the reporting and the dashboard and the integrations. The one page that proves the idea works, running in the real business, being used by real people who will tell you honestly when it is annoying.
Because what you learn from a fortnight of live use is not a refinement of your plan. It is usually a correction to it. The enquiries that arrive are not the ones you imagined. The thing everyone actually wanted turns out to be the small feature you nearly cut. You could not have discovered either by thinking harder, and neither could I.
Pick by frequency, not by glamour
The right first project is the job you repeat most often, not the one that would be most impressive to describe. Frequency is where the hours hide, and it is also where you find out fastest whether the thing works. Something that runs forty times a week gives you a verdict by Friday. Something that runs monthly leaves you guessing until October.
So: first responses to enquiries. Sorting what lands in the inbox. Answering the same five questions. Chasing the same reminders. Unglamorous, high-volume, obvious when it is wrong. Perfect.
Good enough is a legitimate standard
I am comfortable with good enough rather than great, as long as the thing is not mission critical. That is not laziness — it is where the standard belongs. Anything that touches money, promises a date, or goes to a customer with your name on it gets the careful treatment. A tool that sorts your inbox does not need to be beautiful. It needs to be running.
Firstly, ship the rough version. Secondly, live with it for a fortnight and write down every time it irritates you. Thirdly, fix only what actually irritated you, rather than everything you once imagined it might need.
Why our offers are shaped this way
This is the reasoning behind the AI Quick Win Build being one tool, 5–10 days, for a fixed fee, rather than a transformation programme. It is small on purpose. You find out quickly and cheaply whether this helps your business, and you keep the thing either way. Same reason we fix the scope and the fee upfront instead of quoting: a small, defined promise is easier to hold ourselves to.
Overall: the businesses that get somewhere with this are not the ones with the best plan. They are the ones with something running. If you are not sure which small thing to start with, that is precisely what our free audit is for — it names the candidates and puts rough hours against each, so the first step picks itself.
Frequently asked
What should my first AI project be?
The job you or your team repeat most often that does not need judgement — first responses to enquiries, sorting the inbox, answering the same five questions, chasing the same reminders. Frequency matters more than glamour, because frequency is where the hours are and it means you will find out quickly whether it works.
How long should a first AI project take?
Days, not months. If your first attempt at anything has a timeline measured in quarters, you will spend most of it planning against assumptions that turn out to be wrong. Ours are deliberately scoped at 5–10 days for one tool.
What if my first AI project is not very good?
That is the expected outcome and the reason to keep it small. A rough version running in your actual business teaches you more in a fortnight than any amount of specification. Keep the standard high for anything mission critical; accept good enough everywhere else.