Six stages
- Discover: Is there a material problem with an accountable owner?
- Assess: What is the baseline, data boundary, risk and plausible opportunity?
- Design: Which option is proportionate, including a non-AI option?
- Prototype: Does a contained version meet pre-agreed quality and safety checks?
- Deploy: Can named people operate, stop and recover the workflow?
- Improve: Does measured production evidence justify keeping or extending it?
Northstar example
The first opportunity may be quote preparation. A prototype could draft a quote summary from an approved product catalogue, but it must not price, promise stock, or send an email. A sales employee reviews the draft. The pilot passes only if it reduces preparation time without increasing quotation errors.
Define the stop condition early
Write down what would make you stop. For example: missing product data, insufficient sample cases, a legal restriction, higher error rate, or no measurable benefit. A stop decision is a successful consulting result when it prevents a bad investment.
Prototype gate Input: approved catalogue + 30 historical, anonymised quote requests Output: draft only; no sending or pricing authority Pass: reviewer accepts at least the agreed quality threshold Fail: any unsafe claim, missing citation, or no time improvement Owner: Sales operations lead Rollback: disable the drafting endpoint and retain the prior manual process
Practical mission
Choose one Northstar process and write a one-line gate for discovery, assessment and prototype. Each gate must end with a decision, not merely an activity.