1. Describe the workflow as it works today
Choose a recurring task with a clear start, finish and owner. Do not begin with a company-wide transformation. Describe the actual systems and exceptions, including the steps currently handled through spreadsheets, inboxes or verbal approvals.
- Trigger: what starts the task?
- Inputs: what does a person read or collect?
- Actions: which systems are updated, and by whom?
- Exceptions: when is the normal process inappropriate?
2. Specify a business outcome
Choose an outcome that the business can observe. Pair a speed or volume measure with a quality guardrail so the system cannot look successful merely by processing more work badly.
- Current baseline: record a representative sample, not a guessed saving.
- Desired improvement: define the result worth testing.
- Quality guardrail: identify errors that must not increase.
- Decision owner: name who can decide whether the pilot helped.
3. Inventory data and permissions
List approved knowledge sources and the systems that hold them. Access to an export is not necessarily permission to use it in another service. Bring sanitized examples to discovery rather than sharing production credentials or sensitive customer records in an enquiry form.
- Identify the data owner, source freshness and conflicting records.
- Document read-only versus write access for each integration.
- Identify sensitive fields and restrictions for specialist review.
- Confirm who can approve a test environment and scoped credentials.
4. Draw the human boundary
Decide whether the pilot suggests, drafts, or acts. Begin with the narrowest useful authority. An assistant can draft an answer for review before it is trusted to send responses or make irreversible changes.
- List actions requiring a person to approve them.
- Define escalation conditions and the receiving team.
- Specify an appropriate response when information is missing.
- Decide how incorrect actions can be detected and recovered.
5. Write the pilot brief
Combine your answers into a scope that can be evaluated. Compare AI with simpler automation or a process change; the right intervention is the one that solves the problem sustainably, not the one with the most complex model.
- Workflow and users: one paragraph.
- Inputs and integrations: named systems and access constraints.
- Evaluation: representative examples with expected outcomes.
- Launch boundary: human approvals, support and rollback.
- Economics: implementation and recurring-cost assumptions kept separate.
Continue planning
B2B SEO and lead generation: a practical planning guide
AWS RAG architecture: a production planning checklist
White-label AI delivery: an agency readiness checklist
Voice AI pilot readiness: a scorecard for business calls
WhatsApp chatbot and CRM integration: a launch checklist
White-label web development: a handover checklist for agencies