Applies toIT and business teams collecting GenAI ideas; works with any platform or provider
Describe each idea the same way
| Field | Example |
|---|---|
| Problem | L1 agents spend 6 minutes finding the right runbook per ticket |
| Users | 40 service desk agents |
| Today's process and cost | Manual search across wiki and old tickets |
| Proposed AI help | Suggest runbook and draft first response |
| Data needed | Runbooks, resolved tickets |
| Success measure | Handling time, first-contact resolution |
| Risk notes | Wrong advice → agent still reviews |
Score
| Dimension | 1 (low) | 5 (high) |
|---|---|---|
| Value | Nice to have | Large time/cost/quality gain, many users |
| Feasibility | Data missing, complex integration | Data ready, simple integration |
| Data readiness | Scattered, outdated, permissions unclear | Curated, current, permissions clear |
| Risk (inverted) | Customer-facing decisions, regulated data | Internal, human reviews output |
Weight value and feasibility highest. Plot value vs feasibility; start in the high/high corner.
Common good first cases
- Internal knowledge Q&A on curated content with citations.
- Summarising long tickets, incidents or documents for humans.
- Drafting responses that humans edit.
- Classification and routing with human fallback.
Common traps
| Trap | Why it fails |
|---|---|
| Customer-facing chatbot first | Highest risk, highest expectations |
| "AI on all our data" | No clear user or measure; data not ready |
| No baseline | Can't prove value |
| Skipping the process change | Tool exists but nobody's workflow uses it |
Pilot with exit criteria
- Define success measures and baseline before building.
- Pilot with a small user group for a fixed period.
- Review: value achieved, quality, cost, user adoption.
- Decide: scale, adjust or stop. Stopping is a valid outcome.
Feed approved cases into the risk tiering process (see the responsible AI governance post) and the shared platform (see the multi-team platform post).
Tip
Ask for the success measure before discussing models or tools. If nobody can say how they'd know it worked, the idea isn't ready.
Checklist
- Ideas captured in a standard template
- Value, feasibility, data readiness and risk scored
- First cases chosen from high value/high feasibility
- Baselines and success measures set
- Pilot with exit criteria
- Scale/adjust/stop decisions recorded