In their review of the research on how people use automated aids, Parasuraman and Manzey (2010, Human Factors) found that when an aid is imperfect, people make two kinds of mistake: they miss problems the aid did not flag (omission errors) and they follow advice that is wrong (commission errors). They also found that this happens to experts as well as novices, and that practice alone does not remove it. A route planner who has used a routing tool for a year is not safe from it. AI route recommendation training exists to close that gap on purpose.
Picture the situation. The routing model proposes the shortest path for tomorrow's deliveries. It does not know that the council imposed a weight restriction on one bridge this morning. The planner sees a tidy plan, a green status and a deadline. The question this guide answers is how you train people to stop and ask what the model cannot see.
This guide shows logistics training managers and warehouse or dispatch supervisors how to build training around one task: challenging a routing recommendation when relevant context is missing. It covers a short diagnosis, practice with feedback, a worked case with a retry, a scored rubric and a transfer check. The case and rubric are illustrations, so ask a planning specialist to review them against your own procedures before anyone is scored.
Quick facts
| Question | Short answer |
|---|---|
| What do you train? | Checking a route recommendation against facts the model may not hold, and deciding to accept, adjust or override it. |
| What comes first? | Checking who is allowed to override, what the tool shows and how much time planners have. |
| How do you practise it? | Realistic cases where one piece of context is missing, with feedback on each decision. |
| How do you score it? | A 0 to 2 rubric for task execution, reasoning and boundaries, with critical errors defined separately. |
| How do you check transfer? | A new, unseen case, then a supervised review of a live plan. |
| What can training not fix? | Overrides that need a sign-off nobody has time to give, and tools that hide their assumptions. |
1. Describe the performance problem
Start from the job. When a recommendation arrives, a planner has to compare it with what they know: current restrictions, road works, vehicle limits, customer windows, weather and driver hours. They decide whether the recommendation holds, name the missing context, choose to accept, adjust or override, and record the reason. When they cannot decide, they escalate.
Two failures matter. One is following the model into a restriction it never knew about. The other is distrusting the model so much that every plan gets rebuilt by hand, which wastes the benefit of the tool. Training should aim at calibrated judgement, not suspicion.
2. Diagnose the learning and context gaps
Compare the task with how your planners are trained and tested. A walkthrough of the tool's screens checks that people can operate it. It does not show whether they can spot a recommendation that is wrong for reasons the tool cannot show. Analysing a plan and evaluating whether to override it are higher-order demands than operating a screen, and they need practice with cases. Our guide to matching learning activities to Bloom's taxonomy explains how to read the demand of a task.
Then check the causes training will not fix:
- Authority. Who may override, and is a sign-off needed that nobody is awake to give?
- Visibility. Does the tool show its inputs and assumptions, or only the answer?
- Data. Are restrictions and incidents fed into the model, or kept in an email thread?
- Time. Is there room in the shift to check, or are planners measured on plans released per hour?
- Practice. Has anyone rehearsed an override with feedback?
Only the last is a plain training gap. Name the others to the sponsors and agree an owner for each before you build content.
3. Design purposeful preparation and practice
Three layers, each with its own job.
Preparation: check the prerequisites. Give a short diagnostic on what the tool uses as inputs, what it leaves out and which of your own rules limit a route (axle weights, delivery windows, hours of service). A recall check confirms that prerequisites are in place and nothing more. Do not report it as capability. Self-paced study suits this layer.
Practice: cases with one missing fact. Give the learner a recommendation and a small evidence pack, such as a route map, a notice of a restriction, a vehicle record and a customer window. Ask whether the recommendation holds, what the model cannot know and what they would do. Feedback should name the choice: "You accepted the route without checking the vehicle weight against the notice" teaches more than "Try again". This is different from general disruption practice: here the trigger is a specific recommendation, so the skill is review, not recovery. Our guide to building logistics disruption simulations covers the wider skill, and our guide to designing model stress-testing labs shows a neighbouring approach for analysts.
Review and transfer: discuss one ambiguous case. Live input from a supervisor adds most where the evidence is mixed, for instance when a restriction is rumoured but unconfirmed. Asynchronous review is enough for clear checks against an agreed answer.
4. A worked case with feedback and a retry
Illustrative. A dispatch or planning specialist should validate the case and the answer criteria before use.
The case. The model recommends a route for a 26-tonne truck that crosses a bridge on a back road to save 18 minutes. Your evidence pack includes a notice, received that morning, of a 7.5-tonne limit on that bridge starting today. The model's inputs list does not include local notices.
The task. Explain the decision, cite the evidence you used and say what further information or support you need.
A first attempt. "The system picked this route, so it must have checked the roads. I would release it." The reviewer asks which input covered local notices and whether the learner opened the evidence pack. The learner cannot name one, and had not read the notice.
Feedback and retry. The reviewer points to the notice and to the model's inputs list, then asks for a new response.
A stronger attempt. The learner states that the recommendation fails: the vehicle is over the bridge limit and the model does not hold local notices. They override, choose the main road, note the extra 18 minutes and tell the customer's window owner. They record the reason, flag the gap to whoever owns the data feed and say what they still need: confirmation of the limit from the council notice and whether the alternative route has its own restriction.
Why each response is defensible or incomplete. Judge against the approved criteria, not the reviewer's taste. The first attempt trusts the tool without checking and sends a vehicle over a limit, so it counts as a critical error. The second names the missing context, overrides within procedure and escalates the data gap. Add a variant where the notice is unconfirmed and ask the group what a defensible choice looks like.
5. Assess capability and plan transfer
Score each criterion from 0 to 2. Adapt the criteria with your specialist: this rubric is a proposed starting point, not a validated readiness standard.
| Criterion | 0: absent or unsupported | 1: partial, with omissions | 2: complete and justified |
|---|---|---|---|
| Task execution | Accepts the recommendation without checking. | Checks some context but not all of it. | Challenges the recommendation with evidence and records the action and reason. |
| Reasoning | Overrides or accepts with no reason. | Gives a reason but ignores alternatives. | Explains the constraints, compares options and states what is uncertain. |
| Boundaries | Overrides outside approved procedure. | Refers to procedure inconsistently. | Uses approved procedures and escalates when information or authority is missing. |
Define critical errors separately, for example releasing a route that breaks a vehicle limit, or overriding with no record. An overall score must never hide a critical failure. Check that two assessors score the same answer the same way before you rely on the scores. Our guide to assessing judgement and problem solving in training covers anchors and assessor calibration.
A practice score tells you what a learner did in practice. To learn whether it carries over, give a new case after a short delay, then add a workplace follow-up: a supervised review of a live plan scored against the same rubric. Two measures fit this design:
- Verification rate: the share of reviewed plans where the planner checked the missing-context items before releasing.
- Override appropriateness: the share of overrides that a reviewer judges justified, with the denominator and timing set before you start.
Compare a baseline with an unseen follow-up task, and look for concurrent changes, such as a new routing release, a new region or a seasonal peak, before you credit the training. Keep the claim modest: a better practice score is evidence about reasoning in practice, not a promise of fewer failed routes.
Where interactive delivery helps
Interactive formats help where people should commit before they see what others chose. In a live session, a facilitator can show the recommendation and the evidence pack, ask each planner to choose accept, adjust or override and type a one-line reason, then reveal the spread before discussion. That makes over-trust and over-correction visible side by side. Self-paced delivery suits the preparation and the unseen follow-up. AhaSlides, an all-in-one engagement platform for teams and classrooms, supports interactive learning and both live and self-paced delivery. Check how you would move this case between the two against your own workflow before you rely on it.
What training will not fix
- Authority. A sharp rubric cannot create an override right that policy withholds.
- Hidden assumptions. If the tool does not show its inputs, no case will make them visible.
- Time and incentives. If speed alone is rewarded, planners will keep accepting.
- Evidence limits. One review of the research describes how people use automation. It does not show that any training method works for route planners. Measure your own population.
Frequently asked questions
What is AI route recommendation training?
Training that teaches planners to check a route recommendation against context the model may lack, and to accept, adjust or override it with a recorded reason.
When should a planner override a route?
When the evidence shows the recommendation breaks a rule or a fact the model cannot see, and within your own approval rules.
How do you assess override skill?
Use a scored rubric on a realistic case, with critical errors defined separately, then a new case and a supervised review of a live plan. A tool walkthrough only confirms prerequisites.
What can training not fix?
Missing override authority, tools that hide their inputs, no time to check and incentives that reward speed. Check these first and name an owner for each.
Who should review the case and rubric?
A dispatch or planning specialist who knows your tool and procedures, before anyone is scored on them.
Where to start
Pick one recommendation your planners should have challenged, write one case with an evidence pack, two defensible responses and a clear critical error, and adapt the three-criterion rubric with a planning specialist. Run it with a small group, then score a new case and a supervised live review before you scale it.
To see how a case like this could run live or self-paced, explore AhaSlides.
