Most "AI strategies" we see in the wild are 40-slide consultancy decks that confuse activity with progress. They describe vision but rarely commit to a measurable outcome. They name 20 use cases but pilot none.
The framework below is what we actually do. It's stolen from no methodology and works because it's deliberately small.
Days 0–14: Discovery
We spend the first two weeks doing structured interviews with 10–15 people across the business. Not a survey — actual conversations, 45 minutes each. Output: 30 candidate AI use cases.
The interview prompt is simple: "In your job, what's a task that's repetitive, language-heavy, or pattern-recognition-heavy, that you wish a robot could do?" The answers cluster fast. Document extraction, email drafting, support classification, data summarisation, lead qualification.
Days 14–21: Scoring
We score the 30 use cases on a 2-dimensional matrix: business value (revenue uplift, cost saved, hours reclaimed) × feasibility (data available? technically possible today? governance OK?). The top-right quadrant — high value, high feasibility — usually has 3–5 use cases. We pick the top 2.
Days 21–60: Pilot
Two parallel pilots, 6 weeks each. Each pilot has:
- A single accountable business owner (not IT).
- A defined success metric. ("Reduce invoice-processing time from 4 minutes to 1 minute, measured on 100 invoices.")
- A weekly demo every Friday.
- A kill criterion. (If by week 4 we can't see a path to success, we stop.)
The pilot stack is deliberately boring. Azure OpenAI for the model layer. Power Automate or Python for the orchestration. SharePoint or Excel for the data layer. We don't introduce new tools unless absolutely needed.
Days 60–75: Governance & Production
The pilot proved value. Now it has to survive in production. We document: data flowing in and out, access controls, prompt versioning, evaluation methodology, fallback behaviour when the model is wrong, monitoring dashboards.
This phase kills more "successful pilots" than the pilots themselves. A pilot that works on 20 invoices may not work on 20,000. A pilot owned by a single enthusiast doesn't survive when she goes on leave. Production is harder than pilot, and we plan for that explicitly.
Days 75–90: Adoption
The last two weeks are training, change management, and measurement. Power users identified. Documentation written in plain English. KPI dashboard showing actual adoption (not just licenses issued).
At day 90, we have: two production AI capabilities, measured impact, a governance pattern that scales, and a backlog of 5+ scored next use cases. The next 90-day cycle starts.
What we don't do
We don't run "AI strategy workshops" that produce a vision deck and no production code. We don't build custom LLMs from scratch. We don't try to be model-cutting-edge — we use what's stable, governed, and supported. We don't promise that AI will replace headcount; we measure whether it actually saves time.
Want to run this sequence at your company?
We do 1-day AI Discovery workshops as a paid kickoff. Output: a scored use-case map and a Day-0 pilot plan.
Book a Discovery Workshop