How to Scope a Predictive Maintenance Pilot That Converts to Paid

For a physical AI startup, the pilot is the product's first real exam. And most pilots fail it — not because the technology doesn't work, but because nobody agreed in advance what "working" meant.

Pilots that convert to paid contracts tend to be scoped differently from the start. Here's how.

1. Agree on one decision, not a platform

A pilot that tries to show everything proves nothing. Pick one decision the customer makes today and help them make it better: which machine to service first, which part to stock, which alert to act on.

"Predict failures across the plant" is a vision. "Tell the maintenance lead every Monday which five pumps need attention this week" is a pilot.

2. Name the outcome — in the customer's numbers

Before the pilot starts, agree with the customer what success looks like in terms they already track: unplanned downtime, emergency call-outs, parts spend, technician hours.

Write it down. Agree how it will be measured, and against what baseline. If you can't agree on this, the pilot won't convert regardless of how good your model is — because no one will be able to say it worked.

3. Check the data before you sign

Ask for a sample of the customer's sensor and maintenance data before the pilot is agreed. You're looking for:

• Enough history to learn from, including some past failures
• Maintenance records you can actually join to machines
• Sensors that are installed, working and reachable

If the data isn't there, either scope the pilot to collect it first, or pick a different site. Discovering this in week six is how pilots quietly die. (More on this in Why Your Physical AI Demo Fails in the Field.)

4. Make it short and bounded

Set a fixed duration and a fixed set of machines. Long, open-ended pilots lose momentum and sponsors. A clear end date forces a clear decision.

5. Find the person who will buy — and the person who will use

Pilots are often sponsored by an innovation team and used by a maintenance team. Both need to want it. Put the end users — technicians, operators, planners — in the room early, and design around how they work today. A brilliant prediction nobody looks at doesn't convert.

6. Agree on what happens if it works

The most overlooked step. Before the pilot starts, ask: if we hit the agreed outcome, what happens next? Who signs, what budget, how many sites, at what price range?

You're not asking for a commitment. You're making sure a path to "yes" exists — so that success leads to a contract, not to another pilot.

A one-page pilot plan

Before kick-off, you should have one page that both sides agree on:

Item	Example
Decision supported	Weekly list of pumps needing attention
Scope	40 pumps at one site, 10 weeks
Outcome measure	Unplanned pump downtime vs. last year's same period
Data required	Vibration, temperature, 2 years of work orders
Users	Maintenance lead and 3 technicians
If successful	Roll-out to 3 more sites under an annual contract

(Illustrative example — every customer's numbers will be different.)

Why this matters

In the industrial AI product we helped build, the team went from idea to a live MVP in five months, with three customer pilots running and one already converted to paid. Pilots that convert don't happen by accident — they're designed that way. You can read how that product was built in the case study.