Describe the process in one page
Write down what starts the process, where its inputs come from, who owns the result and how completion is observed. If different people give different answers, reconcile those definitions before connecting tools. Automation makes ambiguity happen faster.
Use a fictional invoice-routing process as an example. An invoice arrives, its supplier and amount are extracted, a responsible person checks it and an approval record is created. Payment is a separate action. This separation lets the first pilot prepare work without moving money.
Put exceptions beside the normal path
List missing fields, duplicate invoices, unknown suppliers, unavailable systems and conflicting approvals. Decide which cases stop the process and which can be retried. Give each stopped case an owner and a place to inspect it.
| Event | Safe first response | Evidence of completion |
|---|---|---|
| Valid new invoice | Prepare an approval task | Task ID linked to source invoice |
| Same invoice received twice | Reuse the existing record | Duplicate event linked to original |
| Missing supplier or amount | Stop and request review | Named owner and reason visible |
| Destination unavailable | Retry within a defined limit | Attempt log and final status |
This is an original illustrative workflow, dated September 10, 2026. It has not been deployed as a measured case study. For security responsibilities around connected systems, consult NIST’s small-business resources.
Make the first version observable and reversible
Assign a stable identifier to the business event. Check for that identifier before creating an external effect. Keep the input reference, decision and destination record together so that a person can reconcile what happened.
Use restricted access for each connection and avoid putting credentials in shared spreadsheets or task descriptions. During the pilot, send output to a review queue. Do not silently convert a draft-producing automation into an automatic sender or payer when volume grows.
If AI extracts or classifies information, record the quality criteria and the conditions requiring human review. The NIST AI Risk Management Framework is a voluntary resource for identifying and managing AI-related risks; it does not certify this workflow.
Measure the whole workload
Record baseline handling time on a defined sample. After the change, include review, corrections, failures and upkeep. If preparing a draft takes less time but checking it takes longer, the total matters. State the sample size and period before claiming a productivity improvement.
Stop the pilot if records cannot be reconciled, permissions are broader than required or an external action happens twice. Preserve the logs and return to the manual process. For uncertain text tasks, read the AI workflow guide. For inconsistent sales handoffs, first clarify the shared customer record.