Where to start automating your business with AI: a guide to the first pilot
Published on 3 min read
Most automation projects that fail don't fail because of the technology: they fail because they started too big. Six months of work, five processes at once, and by the third month nobody remembers which problem they set out to solve. The alternative that works is boring to describe but effective: one process, a short pilot, numbers before and after, and only then the next step.
This guide is that method, step by step, the way we apply it with our clients.
Step 1: pick the right process (not the flashiest one)
List the tasks your team repeats every week and score each one on four criteria: how many hours it takes, how often it happens, whether it follows clear rules and how costly a mistake would be. The ideal first project scores high on the first three and low on the last.
With that filter the same candidates almost always win: answering repetitive enquiries, confirming and reminding appointments, moving data from forms and emails into the CRM, and chasing unanswered quotes. And the flashy ones almost always lose: the AI-generated video for social media can wait; the three hours a day of copying data can't.
If you want to put euros on that list before deciding, our savings calculator does the maths in a minute.
Step 2: define what 'it works' means before you start
Before building anything, write down three numbers from today: how many hours a week the process takes, how long it takes to reply to or complete each case, and how many cases are lost or left undone. Without that starting picture, two months from now you'll be arguing about impressions.
Also set the success threshold: 'if the pilot saves at least X hours a week and doesn't create more errors than the manual process, we carry on'. Writing it down first avoids the trap of moving the goalposts afterwards.
Step 3: build a small pilot with a safety net
An AI automation pilot doesn't need months: four to six weeks is enough to connect the tools, configure the flow and get it running with real cases. Two rules for that phase: the scope doesn't grow mid-pilot (new ideas go on a list for later), and every action with external consequences (an email to a customer, a price, an appointment) goes through human approval.
As for tools: for simple flows, platforms such as Make or n8n connect your apps without custom development; when the process touches several systems or needs to understand language (emails, WhatsApp, documents), AI comes in, and it's worth having someone experienced design the limits: what it does on its own, when it asks, and what it logs.
Step 4: measure, decide and only then expand
At the end of the pilot, compare the three numbers with your starting picture. If it didn't reach the threshold, you've still gained something: you know why (messy data, a less regular process than it seemed, the wrong tool) and you fix it or drop it having invested weeks, not a year.
If it worked, expand in the same order: first remove human approval from the steps that have run for weeks without failing, then add the next process on the list. Each expansion follows the same method: numbers before, pilot, numbers after.
That's how a small business ends up with a system that runs on its own without ever making a risky bet. If you'd rather not walk this path alone, in the free audit we'll tell you which process we'd pick in your case and what savings to expect from the pilot.