Turn a repetitive task into a controlled workflow with human review and a safe path for anything that does not fit. This playbook is useful for operations, finance, support, procurement, and administration. Start with “which part of the work is stable, measurable, and reversible?” rather than “what can AI do?”
Expected result: a documented automation decision, an executable workflow, and clear stop and hand-back criteria.
Triage: decide what deserves automation
Score each criterion from 1 to 5:
| Criterion | 1 | 5 | Score |
|---|---|---|---|
| Repetition | rare | daily or more | |
| Rule clarity | heavy judgment | observable rules | |
| Error cost | low and reversible | high or hard to reverse | |
| Data sensitivity | public | personal, financial, or confidential | |
| Volume | few cases | many cases |
Prioritize work with high repetition, rule clarity, and volume, while keeping error cost and sensitivity low or controllable. A high score does not remove review. It only identifies where to investigate first.
Workflow canvas
- Task name:
- Purpose and outcome:
- Trigger:
- Inputs and format:
- Deterministic rules:
- AI task:
- Automated checks:
- Human reviewer and timing:
- Fallback or exception queue:
- Acceptance criteria:
- Logging and retention:
- Process owner:
AI can classify, summarize, extract, or suggest. Deterministic rules should decide amounts, permissions, statuses, and limits whenever possible.
Instructions
- Document the current process, including waits, corrections, and exceptions.
- Score repetition, rule clarity, error cost, sensitivity, and volume. Pick a narrow case.
- Separate input data, fixed rules, and interpretive work.
- Define the trigger and validate the input format. Reject incomplete or unexpected files.
- Specify the AI task with a structured output, not an open-ended instruction.
- Add checks for totals, required fields, duplicates, limits, and consistency.
- Name the human reviewer. Review must happen before sending, paying, publishing, or changing a critical record.
- Define the fallback: manual queue, error template, or return to the previous system.
- Test normal, ambiguous, and adversarial cases. Log errors and refine rules.
- Measure time per case, exception rate, corrections, and error cost during a pilot.
Worked example: processing supplier quotes
Before: Procurement receives 20 quotes by email. Someone copies prices and lead times into a spreadsheet, normalizes units, and prepares a comparison. The process takes two days, and incomplete PDFs create follow-up emails.
Applied canvas
- Trigger: a PDF arrives at
procurement@company.comwith subject “Quote”. - Inputs: PDF, supplier, received date, and project.
- Deterministic rules: accept PDFs under 10 MB; require currency, validity, total price, lead time, and tax details; flag duplicates by supplier, project, and quote number.
- AI task: extract line items, units, prices, tax, lead time, and exclusions as JSON; flag illegible text or uncertain fields.
- Checks: recalculate subtotal and tax; compare currency; verify that the total matches the document; detect blank required fields.
- Reviewer: Procurement owner reviews every record and approves the comparison table.
- Fallback: send failed documents to a manual queue and leave the original email unchanged.
- Acceptance: 98% accuracy on required fields in a 50-document sample; zero automatic payments or awards.
Proposed output
| Supplier | Total | Currency | Lead time | Validity | Status |
|---|---|---|---|---|---|
| Norte Industrial | 18,400 | EUR | 14 days | 30 days | Review |
| Delta Parts | 17,950 | EUR | 21 days | 15 days | Review |
AI does not select the supplier. It prepares a traceable comparison for Procurement to decide.
Before-and-after card
| Before | After |
|---|---|
| Manual copying from each PDF | Structured extraction |
| Back-and-forth for missing fields | Required fields and exception queue |
| Comparison takes two days | Draft table the same day |
| Risk of mistyping a total | Recalculation and human review |
| Implicit decision process | Recorded acceptance criteria |
Copy-ready decision sheet
TASK:
MEASURABLE PROBLEM:
SCORE (repetition / rules / error / sensitivity / volume):
PILOT SCOPE:
TRIGGER:
INPUTS:
FIXED RULES:
AI TASK:
CHECKS:
REVIEWER:
FALLBACK:
ACCEPTANCE CRITERIA:
METRICS AND REVIEW DATE:
DECISION: [pilot / keep manual / discard]
What not to automate or invent
Do not automate irreversible decisions, payments, terminations, sensitive-data access, or external communications without appropriate controls and approval. Do not invent missing fields, suppliers, prices, taxes, dates, or rejection reasons. Route ambiguous, unreadable, or contradictory documents to fallback. Never treat a confidence score as a substitute for review.
Next step
For a one-off task, keep this canvas and continue manually or use the prompt tester. Request a business audit when the workflow recurs, affects a team, carries meaningful volume or spend, handles sensitive data, or creates a visible bottleneck.