AI Policy Creation: From Evidence to Implementation
Participant workbook | StarApple AI in collaboration with the Caribbean AI Association
Four-hour working session. Use redacted or example inputs. Do not upload personal, confidential or restricted material to an AI tool without permission. Your draft needs the organisation’s normal legal and operational approval before use.
1. Define the decision (45 minutes)
Spend 10 minutes selecting one workflow, 15 minutes checking evidence, 10 minutes recording a baseline and 10 minutes sharing the scope.
- Organisation and policy owner:
- Business, agency or regional remit:
- Workflow and affected people:
- Problem to solve and current evidence:
- Actions within scope; actions outside scope:
- Baseline measure, unit, period and data owner:
- Decision required and who has authority to make it:
For every source, record its exact title, issuer, URL, publication date, jurisdiction, source type, claim, denominator and limitations. Separate binding requirements from voluntary guidance. Check whether a headline figure measures businesses, employees, tasks or opinions. Record unknowns instead of filling them with guesses.
Evidence exercise: U.S. Census BTOS reported 19.8% of businesses using AI as of 3 May 2026. The population is U.S. employer businesses covered by the survey. Explain why this is neither a Caribbean adoption estimate nor evidence that your organisation will earn a particular return. Source: https://www.census.gov/library/stories/2026/05/ai-use-businesses.html . Check the survey wording and collection period before reusing the figure.
2. Local context and sovereignty (45 minutes)
Spend 15 minutes mapping data flows, 15 minutes comparing choices and 15 minutes challenging your preferred option.
Answer each question with evidence or an action to obtain it:
1. Which national laws, sector rules and internal policies need legal review? Who verifies applicability?
2. Which languages, accessibility needs and cultural assumptions must the system handle?
3. Where is data stored? Who can access it? What are the vendor’s terms for this exact account and product?
4. Can the organisation export its records and change suppliers? What time and cost would that require?
5. What happens if connectivity fails, pricing rises or the vendor withdraws the service?
6. Which skills can the team maintain locally? Who funds training and ongoing checks?
7. Which decisions can a regional body make, and which remain with each participating country?
Compare at least two procurement or operating options. Record cost, control over data, human capacity, evidence gaps and exit plan. Local hosting alone does not settle these questions.
3. Draft rules people can follow (60 minutes)
Spend 15 minutes drafting scope and permitted uses, 20 minutes setting approval gates, 15 minutes writing incident rules and 10 minutes applying the 4D check.
Policy canvas
- Purpose and measurable objective:
- Scope and exclusions:
- Accountable policy owner and review date:
- Approved tools, account types and uses:
- Data permitted, restricted and prohibited:
- Uses that need approval, approver and evidence required:
- Decisions reserved for authorised people:
- Output checks and release owner:
- Records retained, access permissions and retention rule:
- Incident contact, stop action and restart authority:
- Training requirement and exception process:
Example clause to adapt: Staff may use approved tools for the uses listed in the register. The workflow owner must approve any new use involving restricted information or decisions affecting people before work begins. The output reviewer must check the result against the source evidence before release. If a check fails, stop release, record the issue and escalate to the named owner. Legal and data-protection teams must confirm the final wording for the organisation.
4D review of one AI-assisted task
- Delegation: Which task can AI perform? Which decision remains with a person? Set limits on tool access and actions.
- Description: State the input, expected output, context, constraints and success test. Give the tool only authorised material.
- Discernment: Check claims against sources. Test relevant failure cases. Record what would change your decision.
- Diligence: Name the responsible reviewer, record the checks, protect information and disclose AI use where required by the organisation or applicable rules.
Approval matrix
For each use case record: purpose; data classification; people affected; severity if wrong; likelihood basis; proposed controls; reviewer; approval decision; conditions; next review. Do not turn an unsupported probability into a precise risk score. Test severe low-probability failures separately.
4. Break (10 minutes)
5. Implement and challenge (50 minutes)
Spend 15 minutes on cost, 15 minutes on a failure exercise and 20 minutes assigning the rollout.
Cost worksheet
Annual operating cost = licences + training time cost + output checking time cost + monitoring + support + other operating costs.
Year-one total cost = one-off setup + annual operating cost.
Net quantified value = evidenced incremental benefit minus year-one total cost.
ROI = net quantified value / year-one total cost × 100%, only when costs are positive and the benefit basis is defensible.
Illustrative calculation, USD: 10 people save 1 hour weekly across 40 working weeks at a loaded cost of $25/hour: $10,000 of potential capacity. Checking takes 0.25 hours per person weekly: $2,500. Licences cost $3,000; training and setup cost $2,000. Net capacity-equivalent value is $2,500 against $7,500 cost, or 33.3%. This is not cash profit unless the time is actually redeployed or costs avoided. Do not count the same time saving twice. Test lower adoption, greater checking time and an adverse incident. Model revenue, avoided losses and capacity separately.
Tabletop incident: A staff member uploads a customer record to an unapproved tool. Its output affects a decision. Identify who stops the workflow, preserves evidence, assesses exposure and contacts the appropriate internal teams. Ask who decides whether notification is required and who can authorise restart. Do not assume deleting a chat removes vendor-held data. Record the evidence required before resuming.
90-day plan
- Days 1–30: Confirm scope, inventory uses, verify applicable obligations, establish baseline and assign owners.
- Days 31–60: Train a pilot group, run evaluation cases, record exceptions and test incident handling.
- Days 61–90: Compare results with baseline, review failures and decide whether to expand, revise or stop.
For each action: owner, due date, dependency, budget, success measure, evidence, approval gate. Treat these time bands as a planning exercise, not a promised rollout deadline.
6. Review and revise (30 minutes)
Spend 10 minutes preparing an approval brief, 10 minutes on peer review and 10 minutes revising.
Three-minute brief: Recommendation first. Explain the problem and strongest evidence. State what the policy permits and restricts. Name the costs, remaining risks, accountable owner and decision requested.
Release gates: all must pass
- Every important factual claim has a checked source or is clearly marked as an assumption.
- Scope and legal authority are explicit; legal questions have an owner.
- Every operational rule has a person responsible and a feasible process.
- Sensitive information and decisions affecting people have appropriate controls.
- Cost and benefit calculations show units, time periods and assumptions.
- A failure scenario produces a clear stop, escalation and restart process.
- The implementation plan has resources, dates and measurable review criteria.
If a gate fails, identify the revision and owner. Do not declare the policy approved because the workshop is complete.
Reference shelf
- NIST AI Risk Management Framework, a voluntary framework: https://www.nist.gov/itl/ai-risk-management-framework
- OECD AI Principles, updated May 2024: https://oecd.ai/en/ai-principles
- U.S. Census evidence exercise: https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
Recheck sources and applicable rules when you use the workbook. Policy development is an ongoing organisational responsibility.