The best first AI automation is rarely the flashiest idea. It is usually a recurring task with clear inputs, an understandable decision path, and an output a person can review.
A useful automation removes friction while preserving judgment. The goal is not to eliminate people from a process. The goal is to stop asking people to perform the same mechanical steps every day.
/ 01
Start with evidence, not a tool
Observe five to ten real examples before designing anything. Write down what happens from trigger to completion: where information arrives, who touches it, what gets copied, which decisions require judgment, which exceptions appear, and where evidence of completion lives. NIST's AI RMF calls this mapping: the task, context, people, impacts, and system boundaries must be understood before risk can be measured or managed.
Capture a baseline you can compare later. If a task occurs 80 times a month, takes nine minutes, and needs correction 12% of the time, the baseline is 12 labor hours plus correction time—not a vague claim that the process is 'slow.'
- Trigger: what observable event starts the work?
- Inputs: which fields or documents are required, and who owns them?
- Decision: which rule or judgment changes the next step?
- Action: what is drafted, created, sent, or updated?
- Evidence: where can a reviewer prove the right action occurred?
- Exceptions: which real examples do not follow the normal path?
/ 02
Choose a narrow first version
Do not begin by automating an entire department. Choose one trigger and one useful outcome. A defensible first version might turn a completed intake form into a structured summary and a draft follow-up, then stop for human review. It should not simultaneously qualify the lead, change pricing, send the message, update five systems, and close the record.
Use a boundary statement: 'When X arrives, the system will produce Y for Z to review; it will not perform A, B, or C.' This sentence prevents a prototype from quietly becoming an uncontrolled production system.
/ 03
Build review into the workflow
AI output can be incomplete, unsupported, or wrong. Separate preparation from authority: the model may extract, summarize, classify, or draft, while a named person approves consequential messages, money movement, access changes, employment decisions, contractual language, and production releases.
A useful review screen shows the source, proposed output, confidence or validation failures, what will happen after approval, and how to reject or edit it. A generic 'Approve' button without that context is not meaningful human control.
/ 04
Test normal cases, boundaries, and failure paths
Create a test sheet before launch. Include ordinary examples, incomplete inputs, duplicates, stale information, contradictory instructions, unsupported file types, provider timeouts, revoked permissions, and low-confidence output. For each case record the expected result, actual result, reviewer, and change required.
Retries deserve special attention. If a timeout occurs after a CRM update but before the workflow receives confirmation, a retry must not create a second record or send a second email. Use a stable request ID and idempotent writes.
- Five representative normal cases
- Two incomplete or malformed inputs
- One duplicate submission
- One provider outage or timeout
- One permission failure
- One case that must be handed to a person
/ 05
Measure the process after launch
Track median cycle time, manual touches, correction rate, exception rate, adoption, and completed outcomes. Compare the same measurement window before and after release. If the automation saves eight minutes upfront but creates ten minutes of cleanup, it has moved the work rather than removed it.
Review results after enough cycles to see variation. Record which failures came from code, model behavior, data quality, unclear policy, or user training; each category needs a different fix.
/ Practical takeaway
Bring one recurring task with a clear owner and outcome. Map it, automate the mechanical steps, preserve human approval, and improve it from real use.