Decide Where AI, Rules, and People Belong
Assign workflow steps to process repair, ordinary automation, AI assistance, or human judgment using clear decision criteria.
Choose the mechanism step by step
A workflow does not need one answer. One step may need a required field, another a fixed rule, another AI assistance, and another a human decision. Selecting the mechanism at the step level produces a simpler and safer first release. It also gives employees a clear way to inspect and challenge the proposed design.
| Label | Use it when | Example |
|---|---|---|
| Repair | The policy, data, field, template, or ownership is unclear. | Require the requested service before routing. |
| Automate | The trigger, rule, and output are predictable. | Create a review task when a form arrives. |
| Assist with AI | The input varies and interpretation helps, but a person can review the output. | Summarize the request and extract stated needs. |
| Keep human | The step requires accountability, sensitive judgment, negotiation, or relationship context. | Decide how to respond to a strategic customer. |
Separate deterministic rules from probabilistic output
Ordinary automation follows defined conditions. When the CRM already identifies the account owner, a rule can assign the task. AI produces an interpretation or generated output. It can extract details from a varied email or draft a concise brief, but the output may change and may be wrong.
Do not ask AI to rediscover information the business already stores in a reliable field. Do not replace a missing policy with a model prompt. If two teams disagree about what qualifies, the first task is to agree on the decision rule or define a review path.
Place human control where the consequence changes
A person should usually retain decisions involving customer commitments, unusual exceptions, sensitive information, commercial criteria, and outcomes that are expensive or hard to reverse. The human checkpoint should have a purpose. “Human in the loop” is incomplete unless you name the reviewer, evidence shown, available actions, and expected response time.
Design the review experience:
- Show the source material and the system’s recommendation.
- Expose uncertainty, missing fields, and conflicting evidence.
- Let the reviewer approve, correct, redirect, or escalate.
- Record the action and reason so the team can improve the rules.
Worked example: a human-led enquiry system
The company uses a rule to create the intake task and check for an existing account. AI extracts the stated problem and prepares a short company brief from approved sources. Written rules recommend a queue. A coordinator reviews incomplete, conflicting, high-value, or sensitive requests. The salesperson decides the response and makes any commitment.
This design supports the employee who owns the outcome. It removes searching and preparation while preserving judgment, relationship context, and accountability.
Your turn: label the map
Return to your Module 2 map. Label each step Repair, Automate, Assist, or Human. For every Assist step, write the input, expected output, reviewer, and consequence of a wrong answer. For every Human step, name the context the person needs.
The completed allocation and review rules become Section 4 of the Automation Decision Pack.
Quick Check
1. Two teams disagree about qualification criteria. What should happen first?
2. What should route a record when its owner is stored in a reliable CRM field?
3. What should a human review screen show?
What you have produced
Your responsibility map and human-control plan are Section 4 of the Decision Pack. Module 5 turns this design into a small, testable first release.