All posts

November 12, 2026 · Jermaine Barker

Why Your AI Pilot Needs an Exit Plan Before It Needs a Budget

Most mid-market AI pilots have a launch plan and no plan for what happens when results are ambiguous. Here is how to define success, failure, and shutdown criteria before you spend a dollar.

Every pilot I've been asked to review has a kickoff date. Almost none have a stop date.

That is a quiet reason so many pilots stall. About 80% of AI pilots never reach production. Not because they fail loudly. Because they never get a clear verdict. They drift. A pilot with no exit criteria becomes a permanent experiment, and permanent experiments eat budget, attention, and trust.

The ambiguous middle is where pilots die

A pilot that clearly fails is fine. You learn something and move on. A pilot that clearly succeeds is fine too. You scale it.

The dangerous one is the pilot that sort of works. Users like it. Accuracy is decent. Nobody can say if it saved money. So it gets another quarter. Then another.

In every engagement where I've seen this, the root cause was the same. Nobody wrote down, before starting, what "good enough to ship" looked like. So every result could be argued both ways.

Write three thresholds before you start

Before a pilot gets budget, I ask the sponsor to put three things on one page.

1. The ship threshold. What measurable result means we move to production? Be specific. "Reduces intake processing time from 40 minutes to 15, with error rates at or below the current manual process." Not "improves efficiency."

2. The stop threshold. What result means we shut it down? For example: "If after 60 days the tool does not reduce handling time by at least 20%, we stop." This is the line most teams refuse to write. Write it anyway.

3. The risk threshold. What event pauses the pilot immediately, regardless of performance? A data exposure. A wrong output reaching a customer or patient. A compliance flag. This is governance doing its job. Guardrails are what let leadership say yes, because they know the downside is bounded.

Start with the workflow, not the model

Thresholds only work if you know the workflow you are changing. If you cannot describe the current process step by step, you cannot measure improvement.

I've watched teams choose a model first, then hunt for a place to use it. They end up measuring model quality instead of business outcomes. A model can score well on a benchmark and still make no difference to the person doing the work.

Map the workflow. Time each step. Identify who touches it and where errors happen. That baseline is your comparison point. Without it, every pilot result is an opinion.

Set a 90-day clock

A pilot should answer a question, not run forever. My rule is one workflow, one owner, 90 days.

Days 1 to 15: baseline the current process and agree on thresholds. Days 16 to 60: run the pilot with real users and real data, inside defined guardrails. Days 61 to 90: measure against the thresholds and make the call.

The call is one of three: ship, adjust once, or stop. "Adjust once" matters. You get one structured revision, not an open-ended series of tweaks.

Name the decision maker

Pilots also stall because no one owns the verdict. A committee reviews results and defers. Name one person who decides at day 90. Give them the one-page thresholds. Their job is to compare results to the page, not to rebuild the case from scratch.

If that person cannot be named, you are not ready to pilot. That is a useful thing to learn early, and cheap.

Stopping is a success

I tell leadership teams this directly: a pilot you shut down on schedule is a win. You spent a fixed amount, got a clear answer, and freed the team for the next use case. The expensive outcome is the pilot nobody will end.

Organizations that build this habit get faster over time. Each pilot sharpens how they judge the next one. Discipline beats enthusiasm every time.

Where to start

If you are weighing a pilot right now, take an hour and write the three thresholds. If you cannot, that tells you where the gaps are. Our free AI Readiness Assessment walks through workflow clarity, data readiness, governance, and ownership so you can see those gaps before you commit budget.

And if you want to see how we structure the path from first pilot to production, the ASCEND framework lays out the 90-day approach in detail.

Ship one thing. Know how you will judge it. Know when you will stop.

Wondering where AI fits in your organization?

The free 5-minute AI Readiness Assessment shows you exactly where to start.

Take the Free Assessment