An AI strategy document should fit on a few pages and answer six questions: what business objectives AI will serve, how you will judge success, which use cases you will pursue first, how they will be governed, what they will cost, and who owns each part with dates for review. Below is a template with those sections and a filled example you can adapt.
If you have not yet chosen your use cases, start with how to build an AI strategy for your company, then come back to document it.
Section 1: objectives and evaluation criteria
State the two to four business objectives AI will support, and how you will know it is working.
- Objective: the business outcome, in business language
- Measure: the number that will change
- Baseline: where it is today
- Target and date: where you expect it to be, and by when
- Guardrails: measures that must not get worse, such as quality, complaints or error rates
Section 2: use cases
For each prioritised use case, record:
- the problem it solves and who it affects
- the data it needs, and whether you have it
- the expected benefit, and how it will be measured
- the main risks
- whether a human reviews the output
Section 3: governance
- Principles: for example human oversight of customer-facing outputs, no personal data in public tools, and accuracy checks before use
- Approval: who approves new use cases, and against what criteria
- Data and security: what data can be used, where, and how it is protected
- Monitoring: how quality, bias and errors are checked after launch
- Compliance: relevant legal obligations, such as data protection law, reviewed with your legal adviser
Section 4: costs, ownership and review gates
- Costs: licences, integration, data work, training, staff time and ongoing monitoring
- Owners: one accountable owner per use case, plus an overall sponsor
- Review gates: dates when each use case is reviewed and a decision is made to scale, adjust or stop
Section 5: people and change
How roles will change, what training is planned, and how affected teams are involved.
A filled example
The example below is fictional and shortened.
| Section | Example entry |
|---|---|
| Objective | Reduce time to resolve customer queries |
| Measure and baseline | Median resolution time: 26 hours today |
| Target | 12 hours within 12 months |
| Guardrails | Customer satisfaction stays at or above 82%; complaint rate does not rise |
| Use case | AI drafts first replies to common email queries; an agent reviews every reply before sending |
| Data | Past query and reply history, anonymised before use |
| Main risks | Wrong or inappropriate answers; personal data exposure |
| Governance | Human review of all replies; weekly quality sample of 50 replies; no customer data in unapproved tools |
| Costs (year 1) | Licences, integration work, 2 days' training per agent, 0.5 FTE quality lead |
| Owner | Head of Customer Service; sponsor: COO |
| Review gates | Month 3: pilot results against baseline; month 9: decision to scale |
AI roadmap template
Add a one-page timeline at the end showing each use case, its stages (prepare, pilot, scale), owners and review gates. Keep it to 12 to 18 months; beyond that, plans change too much to be useful.
Common gaps
- Objectives written as "use AI more" rather than business outcomes
- No baseline, so nobody can show improvement
- Costs limited to software licences
- No stop criteria
- Governance written but not assigned to anyone
Related reading
- AI strategy subject guide: building, documenting, measuring and presenting an AI strategy
- AI use-case and business-case pack: a worked first-year business case with a downside case
- Common AI strategy failures and how to avoid them
- How to measure AI ROI and strategy outcomes
- How to present an AI strategy to your board
To learn to build an executive AI roadmap with feedback on your own organisation's document, see our AI Strategy for Executives course. For a wider introduction, see Strategic Application of AI in Business.
The bottom line
A good AI strategy document is short, specific and owned. Objectives with baselines, prioritised use cases, real governance, full costs and dated review gates are what make it usable.