Start with work that people must take over if the agent stops.
Imagine an agent prepares your customer quotes. Your fallback says that someone in sales will prepare them if the agent is unavailable. That sounds straightforward until an order arrives: which price list applies, who can approve the discount, and where does the person find those records? Our proposed work sample begins with that practical dependency.
Choose one manageable request and name the person who would handle it during an outage. Ask a qualified colleague to review the result. The aim is to find out whether this quote can be prepared under agreed conditions and what support is missing. A failed attempt alone cannot tell you whether AI caused someone to lose a skill.[S2]
The work sample produces a new quote without AI help.
Give the person the request, current price list and discount rule. Agree the permitted tools, a reasonable time allowance and the expected output before they begin. In our invented example, the documents and a calculation tool are available. AI help and the agent’s finished quote are excluded during the attempt.
The person prepares a new draft and explains the important decisions. Nothing is sent to a real customer. If a quantity is missing or a discount needs approval, a justified clarification or request for approval belongs in the result. Memorising prices is not the task; being able to use the right records matters.
A qualified person checks the result and its reasoning.
The reviewer uses the criteria agreed beforehand. The wording need not match the agent’s draft. What matters is whether the quote follows the request and whether its price and discount can be traced to the documents. We would write down four checks:
| What to check | What the reviewer looks for |
|---|---|
| Work and quantity | The items match the request; unclear details are recorded as questions. |
| Price and calculation | The price list is identified and the calculation can be explained. |
| Discount and authority | The rule permits the discount; any required approval is named. |
| Missing information | The person asks for the missing record instead of inventing an answer. |
Keep the draft, documents, reviewer’s finding and agreed follow-up together. They show what was actually attempted. A “passed” checkbox without the task and criteria leaves the next person guessing.[S2]
If a gap appears, arrange qualified cover for the ongoing work.
Suppose the person cannot justify the discount. Find out whether they need knowledge, access, a missing document or a clearer task. Until that is resolved, agree who can prepare and review real quotes. An unverified fallback should not be the only cover the process relies on.
Where the agent’s authority depends on human review, the responsible person also needs to decide what may proceed while that review is unresolved. Agree suitable support and another attempt on a comparable request. This is a proposed process response, not a result we have measured. Our answer about responsibility for agent outputs explains how to make that ownership visible.
Recognising a wrong quote is a separate task from preparing one.
A person might recognise that an agent used the wrong discount while still needing help to prepare a complete quote. In a separate judgement check, we would give them a draft with a deliberate discount error and ask them to explain the discrepancy using the rule. Producing a new quote tests the fallback capability described on this page.[S2]
Our current public account of the Comprehension Obligation makes this distinction and corrects earlier descriptions of manual reproduction as the definition. It also corrects claims that the rehearsals were already running: no completed results are reported, and costs remain estimates or modelled figures.[S3] Time and repetition must suit the task; the earlier 30-minute quarterly parameters are not a proven threshold.
One successful attempt applies to this task under these conditions.
A small experiment helps explain the learning concern. 52 Python developers learned the unfamiliar Trio library. In an immediate quiz, the AI-assisted group averaged 50 percent and the comparison group 67 percent; the completion-time difference was not statistically significant.[S1] That is evidence about short-term learning in coding. It does not establish long-term skill loss in sales or validate our proposed work sample.
Evidence and limits
- The quotation task, deliberate error and review questions are invented MING examples, not a client case or completed trial.
- One attempt shows what the person could do with those records, tools and criteria. It does not establish lasting competence or general judgement.
- We report no pass rate, measured efficacy, supervision cost or current company-wide rollout. The public author account is not an independent audit of private records.
Choose one quote the team would actually need to prepare during an outage. Put the task, permitted records and tools, criteria, reviewer and date on one page. That gives the next attempt a purpose and leaves a usable record of what needs attention.
