When AI Should Recommend and When People Should Decide
DBR77 Team7 min read
On this page
Let AI act on its own only where a wrong action is cheap, quick to reverse and inside written limits. Let it recommend, with a named role approving, where a decision changes the plan, the product, cost or safety. Everything else starts at "observe": AI watches and flags, people act. Set the level per type of decision, not per AI system, and move a decision up a level only with evidence.
A plant has to settle which decisions to trust AI with, under which rules, and who answers for the result.
Why "AI or human" is the wrong question
A plant makes hundreds of decisions a day. Most are small: which task first, which order next. A few are large: releasing a lot, changing a recipe, stopping a line. Putting all of them under one rule makes no sense. If every AI suggestion needs a manager's signature, nothing moves. If none does, the plant loses control of the decisions that matter.
Toyota solved a version of this long before AI. Its idea of jidoka, "automation with a human touch", lets a machine detect the abnormality and stop automatically. The machine handles the routine and the stop. People respond to the abnormality, and, as Toyota puts it, only humans can implement kaizen. Toyota designed that split of work between machine and person, and AI in operations needs the same deliberate design.
AI can observe, advise or act, and observe is the default for anything new.
Three levels: observe, advise, act
| Level | What AI does | What the person does | Typical decisions | Approval rule |
|---|---|---|---|---|
| Observe | Watches data, flags drift, summarizes | Decides whether to act and what to do | Safety observations, new lines, new AI uses | No approval step; the person owns the action |
| Advise | Recommends an action with reasons and a confidence level | Approves, changes or rejects it | Schedule changes, work order priority, likely root cause, lot disposition | Named role approves within a time limit, else escalation |
| Act | Carries out the action inside written limits and logs it | Reviews afterward and can reverse it | Draft tasks, work orders from a threshold, routine reorders under a set value | Allowed only for cheap, reversible, bounded actions |
Most plants put almost everything at "advise" at first, and that is a sound start. "Act" should be rare and narrow. "Observe" is the default for anything new.
Five questions that set the level
For each type of decision, answer the following:
- What does a wrong action cost? Scrap, a missed shipment, a safety risk, or ten minutes of rework?
- Can it be reversed? A task can be reassigned. A shipped lot cannot be recalled cheaply.
- How fast must it happen? If the value is in seconds, a person in the loop may be too slow, so set tight limits instead.
- How good is the evidence? Is the data clean, and has the AI been right on this decision before?
- Is it regulated or safety-related? If yes, a person decides, whatever the other answers.
Illustrative example: a plastics molding plant reviews three decisions. Swapping two orders on the same press within a shift is cheap and reversible, so it sits at "advise", and the planner approves with one click. Releasing a lot on quality hold after a failed check is costly and hard to reverse, so it stays at "advise" with the quality lead as the only approver, and never moves to "act". Opening a work order when cooling-water temperature passes a set limit for 10 minutes is cheap and bounded, so it sits at "act": the work order is created, a technician owns it, and the shift lead sees it at handover.
How to write approval rules
An approval rule is short and answers six things:
- Who approves: a role, not a person, so the rule works on every shift.
- The time limit: how long the recommendation can wait.
- What happens without an answer: escalation to the next role. By default, nothing changes on the line.
- What the approver sees: the data used, the reasoning and the confidence level.
- What is logged: the recommendation, the decision, who made it and, for an override, why.
- How to stop: who can move the decision back to "observe" at once.
Law points the same way for high-risk AI. Article 14 of the EU AI Act requires that the people overseeing such systems can correctly interpret the high-risk AI system's output, can disregard, override or reverse it, and can interrupt the system through a "stop" button or a similar procedure. Whether a given plant use counts as high-risk is a legal question, but the list makes a good checklist for any approval rule.
The NIST AI Risk Management Framework, a voluntary guide to better manage risks from AI to people and organizations, is a useful reference when a plant writes its own AI policy.
When to move a decision up or down
Review the levels monthly, for example in the tier 3 meeting. For each decision type, look at how often recommendations are accepted, how often they are changed and why, and how often an accepted recommendation turned out wrong.
Move a decision up one level only when acceptance is high, overrides are rare and well explained, and errors are cheap. Move it down at once after a costly error, a new product or a change in the data. For clear owners across shifts, see How to Keep Accountability Clear Across Shifts; for the records behind each decision, see How to Keep Audit-Ready Records of Factory Decisions.
How this works in DBR77 IRIS
In DBR77 IRIS, Data & AI recommendations wait for a person to apply, approve, reject or defer them, and each one shows the basis for its score: the rule, the formula and the inputs. The score is a rule priority, so nobody mistakes it for a model's certainty. The IRIS assistant only reads data and documents, and the Compliance screen keeps a register of AI uses and high-risk decisions.
An approved recommendation becomes a task with an owner, and then a work order, a schedule change or a parameter adjustment. In IRIS APS, plan versions and approval workflows make sure the right people sign off before a plan reaches the floor. The log holds each recommendation, the decision on it and the outcome, and the outcome feeds back to improve the next suggestion. What AI Can Really Do in Factory Operations Today suggests which uses to start with.
FAQ
Can AI ever act without approval?
Yes, for narrow actions that are cheap, reversible and inside written limits, such as opening a work order from a threshold. The action is logged and a person owns it.
Who is accountable when an approved recommendation is wrong?
The role that approved it, as with any decision. The log shows what the approver saw, so the review can find whether the rule or the data needs fixing.
Won't approvals slow the shift down?
Not if the rules are short and set by risk. Low-risk approvals take one click. Only a few decision types need a senior approver.
Conclusion
List the decisions your planners, technicians and quality team make every week, run each through the five questions, and give it a level with a written approval rule, a time limit, a log and a way to stop. Start most at "advise" and keep safety and quality release with people. Review the levels monthly, and move a decision up only when acceptance, override and error rates say it is safe.
Sources
- Toyota Motor Corporation, Toyota Production System
- European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14: Human oversight, 2024
- NIST, AI Risk Management Framework, 2023