Scope an enterprise AI pilot with stop-go criteria

A pilot is a decision instrument, not a small production launch. Its job is to reduce the largest uncertainties cheaply enough that a named sponsor can continue, change direction or stop. If the scope says only “build a chatbot in six weeks,” a good demo can pass while the business case, data access and operating risk remain unknown.

Build the decision artifact

Start with the current workflow: weekly volume, cycle time, exception rate, error cost and who makes the decision today. Select one bounded user journey and keep irreversible production writes outside the pilot. Then define gates for data sufficiency, task quality, operational safety and value. Each gate needs a measure, a threshold, an evidence source, an owner and a stop condition agreed before results are visible.

The enterprise delivery lab makes this reasoning executable with synthetic data:

from enterprise_cases import pilot_scope_case

result = pilot_scope_case()
assert result["weeks"] == 6
assert len(result["gates"]) == 4
assert result["production_write"] is False
assert result["decision_owner"] == "pilot sponsor"

The fixture does not merely set success targets. It records a lower boundary that stops the work, gives the sponsor ownership of the decision and keeps production writes out of scope. The four gates also prevent a strong model score from hiding weak data, duplicate side effects or negligible time savings.

Test the failure path

A pilot becomes theatre when the team changes the test set after seeing results, reports averages that hide high-cost failures, or treats a sponsor’s enthusiasm as acceptance. Freeze the evaluation protocol, label exceptions separately and preserve the baseline. A failed gate should produce a decision record, not a quietly revised threshold.

Save evidence that another reviewer can inspect:

• signed pilot charter and baseline extract
• versioned evaluation set with inclusion rules
• gate report with pass, change or stop recommendation
• risk and assumption log with decision owners

This exercise aligns with the enterprise delivery, secure integration, operations and leadership work in the FDE for Professionals course. The course link describes the pathway; this local draft and its synthetic verifier do not claim that a client deployment, security approval or production result occurred.

Practice task

Write four gates for one workflow you know. Add the evidence that will be inspected, the person who can decide and one result that would make stopping the responsible choice.

Continue learning

This article is part of the Enterprise AI delivery and architecture sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Continue with Choose rules, classical ML or an LLM for a client workflow.
• Then apply it in Map an enterprise data boundary before building retrieval.

Reference: GOV.UK Service Manual: discovery phase.