A small business may have the same privacy, accuracy and customer-trust obligations as a larger organisation but far fewer specialist resources. That makes scope discipline a competitive advantage. Start with work that is reversible, uses public or synthetic data and has a visible review step.
The framework here links four decisions: define the unit of work, describe the minimum context, discern whether the output passes, and designate an accountable approver. It is an operating habit, not a claim that every task should be automated.
Evidence the course or workflow should provide
Define the work unit
Break a broad ambition into a repeatable input, transformation and output. 'Improve marketing' is not testable; 'draft three subject lines from an approved brief' is.
Observable evidence: One real sample and its current time cost.Describe context and limits
Provide voice guidance and approved examples while excluding customer identifiers, confidential contracts and credentials unless an authorised system and purpose exist.
Observable evidence: A minimum context pack with a data classification.Discern quality
Write pass criteria before generating. Include a hard-fail rule for unsupported facts, privacy leakage or policy violations.
Observable evidence: A rubric and ten representative cases.Designate accountability
Name who operates the tool, who reviews the work and who can stop the workflow. Accountability cannot be delegated to the model.
Observable evidence: A lightweight RACI or owner-reviewer table.Measure net effect
Compare total cycle time, including review and repair, against the current process. More output is not automatically more value.
Observable evidence: A two-week before-and-after log without invented ROI.Four-decision pilot canvas
Complete this before selecting a vendor. An unanswered cell is a scope risk, not a reason to guess.
| Decision | Question | Evidence | Accountable role |
|---|---|---|---|
| Define | What exact unit repeats? | Sample and baseline | Process owner |
| Describe | What context and limits apply? | Approved context pack | AI operator |
| Discern | What passes or fails? | Rubric and test set | Reviewer |
| Designate | Who accepts the consequence? | Approval and log | Business owner |
How to apply this guide
- Choose reversible workInternal drafts or summaries of public material are safer pilots than autonomous customer replies.
- Classify the dataSeparate public, internal, confidential and restricted information; use the minimum necessary data.
- Run ten casesInclude edge cases and record review time, not just generation speed.
- Make a documented decisionContinue, revise or stop based on the baseline and failure log rather than enthusiasm.
Failure modes to avoid
- Tool-first buyingA subscription cannot clarify an undefined process.
- Invisible approvalIf nobody owns the final decision, the workflow is not controlled.
- ROI theatreUse measured time and defect data; do not manufacture savings projections.
Official and primary sources
Practise the framework on a bounded use case
Review the AI Fluency for Small Businesses course and compare its stated outcomes with this pilot canvas.
View the related course