Framework · free to use

How to run an AI pilot that doesn't stall.

Most AI pilots in B2B fail not because the technology doesn't work, but because the pilot was set up to fail. They lack clear scope, no real owner, no decision criteria for what "success" means, and no plan for what happens after the pilot. This is how to avoid all four.

The four rules of a non-stalling pilot

Get these right first

One workflow, not three. Pilots that try to test "AI across the sales team" stall. Pilots that test "AI for proposal drafting" ship.
One owner, not a committee. A single human, accountable for the pilot's outcomes, with calendar protected.
Pre-defined success criteria. Measurable. Time-bound. Written down BEFORE the pilot starts.
A decision rule for what happens at the end. If success is hit, what's the next step? Pre-commit so the pilot doesn't become a permanent purgatory.

The 4-week pilot structure

Week-by-week

Week 1 — Setup

Define the workflow precisely. What is the input? What is the output? Who uses it?
Identify 2-5 users who will participate. Get their commitment.
Set up the AI Project: load examples, write system prompt, test with 3 real cases.
Define success criteria: time saved per use, quality met or exceeded, voluntary adoption rate.

Week 2 — Shadow use

Users do the workflow both ways — their normal way AND with the AI tool. Compare outputs.
Daily check-in (5 min) with the AI lead. Surface friction immediately.
Refine the Project based on what's not working.

Week 3 — Primary use

Users do the workflow primarily with the AI tool. Fall back to manual if it fails.
Track time per use, quality issues, complaints.
Mid-week refinement pass.

Week 4 — Production

Users do the workflow with the AI tool, no manual fallback by default.
End-of-week review: did we hit success criteria? What's next?
Pre-committed decision: scale to whole function, refine for 2 more weeks, or kill.

Success criteria worth using

Pick 2-3

Time per use: X% reduction vs baseline (measure both).
Quality at or above baseline: reviewed by [named person].
Voluntary adoption: users prefer the AI workflow when given the choice.
Throughput: users complete N% more of the work per period.
NPS / satisfaction: end users rate the new workflow as net positive.

What kills pilots

Recognize and avoid

Open-ended pilots. No end date = no decision = no production.
Too many participants. 6+ users means slow feedback, mixed signals.
No baseline measurement. If you can't measure before, you can't claim after.
Skipping the Project setup. Generic prompting produces generic results. Always set up a Project with knowledge loaded.
No pre-committed decision rule. Successful pilots stall waiting for someone to decide "is this really worth deploying?"

When the pilot succeeds

What to do next

Pre-committed before pilot start: "If we hit our success criteria, we'll roll this out to the whole [function] within 30 days." Without pre-commitment, success leads to delay.

On day 28 of the pilot, the next 30 days of rollout work should already be calendared — not waiting for a steering committee meeting.

When the pilot fails

Diagnose honestly

Did the workflow not fit AI? (Some don't. That's fine. Move on.)
Was the Project under-built? (Often the case. Load better examples; iterate.)
Was the owner under-resourced? (Most common cause.)
Were the users unwilling? (Sometimes a signal you're forcing AI into the wrong workflow.)
Was the success criteria wrong? (Sometimes you measured the wrong thing.)

Document the failure honestly. Pick a different workflow. The pilot failing is data, not defeat — provided you actually learn from it.