AI & Machine Learning
How we run a six-week AI pilot
The plan we use to test an AI tool on one real workflow before anyone commits a big budget to it.
5 November 2026 · 2 min read
Most AI projects that fail don't fail because the model wasn't clever enough. They fail because nobody agreed what job the tool was meant to do, how anyone would tell if it was doing it well, or who would look after it once the demo was over.
A short pilot with a fixed scope fixes most of that. This is roughly how we run one.
Week 1: pick one job
We look for a task that comes round every week, involves a lot of reading or typing, and that nobody on the team enjoys. Sorting inbound support emails is a good one. So is pulling the same six fields out of every supplier invoice, or writing the first draft of the Monday sales report.
We steer away from anything where a wrong answer is expensive and nobody would notice it. The best first pilots are tasks a person already checks.
By the end of the week there's a one-page brief: the task, who does it now, how long it takes them, and what a good result looks like.
Week 2: write down how a person does it
Before building anything we find out where the inputs live (an inbox, the CRM, a shared drive) and how messy they are. Then we sit with whoever does the task today and write down the decisions they make, step by step.
That list turns into the tool's instructions, and later into the checklist we test it against.
Weeks 3 and 4: build the smallest version that helps
The first version shows what it's doing while it works, gives the answer before the detail, and links to the documents it used. Anything that would leave the building, like an email or a CRM update, waits for a person to approve it.
It also lives where the team already works. If they spend the day in Outlook and HubSpot, that's where it shows up.
Week 5: test it on real history
We pull 50 to 100 real past examples and run the tool against them, then compare its output with what your team actually did. You see how often it's right, the kinds of mistakes it makes and how much time it would have saved. You don't get a cherry-picked demo.
Week 6: decide
The pilot ends with a short report and a recommendation: roll it out, change it, or stop. Stopping is a perfectly good result. You found out in six weeks instead of six months.
What it needs from you
Someone who owns the workflow and can give us an hour or two a week, access to a sample of real data, and a few people willing to try it. That's about it.
Where pilots go wrong
The most common mistake is starting with a general chatbot that's meant to answer everything. It's hard to test and people rarely keep using it. Skipping the testing week is a close second, because without numbers you can't defend the project to anyone. And tools without an owner after launch quietly get worse as the business changes around them.
If you have a task in mind, book a call and we'll tell you honestly whether it's a good first candidate.