Most AI-in-operations programs don't fail at the algorithm. They fail at line selection. The pilot gets placed where it's politically convenient, and a year later there's a demo nobody can replicate and a rollout nobody believes in.
When we scope a lighthouse with a client, five criteria do most of the deciding.
1. A problem the P&L can feel
Not "interesting data", but a loss someone is already accountable for: scrap on a specific step, drift that eats yield, changeovers that blow the schedule. If the line's biggest problem is vague ("efficiency"), keep looking. The lighthouse has to produce a result the COO can point at without a footnote.
2. Data that exists, even if imperfect
You need timestamps, lot or batch identifiers, and basic process parameters that can be joined to outcomes. You do not need a data lake. We've seen more pilots drown in a two-year data-platform prerequisite than in missing sensors. If 70 percent of the relevant data is reachable, start; treat the gaps as improvement work, not blockers.
3. A daily routine to land in
AI output is a signal. Signals die without a meeting that looks at them, an owner who acts on them and a standard that captures what worked. If the line has no functioning daily management, build that first, or accept that you're funding a dashboard, not a lighthouse.
4. A leader who wants it
Not tolerates. Wants. The line leader will have to defend odd hypotheses, run controlled tests, and change standards. A skeptic who was assigned the pilot will quietly starve it. One genuinely curious shift leader outweighs a steering committee.
5. Repeatability beyond the line
The point of a lighthouse is the second line. Prefer a process family that exists elsewhere in the network (same machine class, similar product mix) so that what you learn travels. A heroic one-off on your most exotic line proves nothing transferable.
What we deliberately ignore
Line size, prestige, and who asked loudest. The flagship line with the works-council spotlight is usually the worst place to learn: too many eyes, too little patience for the messy middle weeks.
Score your candidate lines against these five honestly and the choice usually makes itself. If no line passes, that's not a verdict against AI. It's a to-do list, and cheaper to discover before the pilot than after.