A pilot is how you find out whether AI works for your business before you spend real money, on your real data, with a real number at the end. Done well, the worst case is a clear, cheap “no.” Here's how to run one without betting the business.
Pick one thing, the right one
The most common pilot mistake is choosing something exciting instead of something useful. A good first candidate is narrow, repetitive, and measurable. Score your options against four questions:
- Does it happen a lot? Frequency is where value compounds.
- Is it painful or costly today? Pick something people already complain about.
- Can we measure the result? Time, errors, cost, revenue. You need a before and an after.
- Is a wrong answer survivable? For a first pilot, choose a workflow where a human still reviews the output. Save the zero-tolerance cases for later.
Boring usually wins. “Draft the first version of routine quotes” beats “reinvent how we sell.”
Set the bar before you build
Decide what success looks like up front, in writing, so you can't move the goalposts later. A simple frame:
- Baseline: how long / how costly / how error-prone is it today? Measure it now.
- Target: what would make this clearly worth scaling? 30% faster? Half the errors?
- Walk-away line: below what result would you stop? Naming this in advance is what keeps a pilot honest.
Keep it small and time-boxed
A first pilot should be measured in weeks, not quarters, and run on a slice of real data rather than your entire operation. Small scope means a fast answer, a low bill, and a contained blast radius if it doesn't work. You're buying information, not a finished system.
Plan for all three outcomes
Before you start, agree what happens in each case, so the result leads somewhere instead of sitting in a slide deck:
- It clears the bar → scale it deliberately, with the numbers to justify the spend.
- It's close → adjust one or two things and re-test, or keep it as a smaller win.
- It misses → stop, and bank what you learned. You should still keep the working prototype and the report. Even a “no” leaves you better informed and out only a small, known cost.
The point is the number
A good pilot ends with a plain-English result and a real figure (hours saved, errors cut, revenue moved) plus an honest recommendation on whether to expand. That number is what turns “AI sounds interesting” into a decision you can defend.
If you have a candidate problem in mind, the Pilot Project is built to do exactly this, or take the readiness scorecard first if you're not yet sure a pilot is your right next step.