Measure the cost of unnecessary AI clarification on a US maritime desk

A question should resolve something that matters

Clarification is useful when an answer changes the task, prevents a consequential mistake, or establishes missing authority. It is wasteful when the information is already available, the question concerns a minor reversible choice, or the assistant repeatedly asks for an instruction the user has already given. A US maritime desk should evaluate both unnecessary questions and failures to ask necessary ones.

Microsoft's HAX guidance on uncertainty supports disambiguation or a narrower service when user goals are unclear. Anthropic's research on agent autonomy discusses the role of user intervention and agent-initiated clarification. These sources provide interaction context, not a recommended number of questions per voyage. The commercial framework below is an original method for deciding whether interruptions help the work.

The objective is neither maximum autonomy nor maximum questioning. It is appropriate progress. An assistant should use the available evidence, continue work that is clearly authorized, and ask a precise question when the missing answer is material to the next dependent action.

Distinguish three kinds of uncertainty

The first kind is missing information that may already exist in the supplied material. If a cargo weight appears in an attachment, the assistant should inspect that attachment before asking the broker to repeat it. A question caused by incomplete retrieval is not the same as a genuinely missing fact.

The second kind is an unresolved commercial choice. The customer may need to choose a delivery endpoint or decide whether a partial shipment is acceptable. The assistant can explain the options and their consequences, but it should not invent a preference. This is a useful point for clarification because the answer changes the work.

The third kind is a routine presentation choice. The assistant may need to choose a sensible table layout or concise wording for an internal summary. If the user has not specified a preference and the choice is easily reversible, repeatedly asking can create more effort than it saves. The system should make a reasonable choice and allow correction.

A fictional enquiry review

Imagine a hypothetical US exporter supplies a cargo spreadsheet, an email with the destination terminal, and an instruction to prepare an internal comparison. The assistant asks for the cargo weight, asks whether to use a table, and asks whether it may read the attached quotation. All three interruptions are unnecessary in this scenario: the weight is supplied, the format is a routine choice, and reading the quotation is inherent in the authorized task.

Later, the assistant discovers that one proposal ends at the terminal while another includes an inland leg. Asking which endpoint should govern the comparison may now be useful if the customer's intended endpoint is not established. The question changes the basis of the analysis. It should explain that consequence rather than simply asking the user to clarify the shipment.

While waiting, the assistant can still extract the quoted components and identify exclusions. It should not halt every independent part of the task. When the user answers, the comparison can proceed from useful prepared work instead of restarting. The missing decision should block only the work that actually depends on it.

The test therefore evaluates more than whether the assistant asks. It evaluates whether the question is necessary, well timed, and narrow enough to answer. A system that asks the right question after doing the relevant preparation can save attention even when the total number of questions remains unchanged.

Count interruption cost realistically

The cost of a question includes reading the request, recalling the context, locating any needed information, answering, and returning to the original task. A short message can impose a substantial context switch. Measuring only the time spent typing the answer understates the burden.

At the same time, a necessary question can prevent much greater rework. Do not penalize clarification simply because it takes time. Record what decision it resolves and what error it avoids. The evaluation should distinguish productive interruption from repeated or poorly formed requests.

A useful review record can classify questions as necessary, answer already available, routine choice, duplicate, or unclear. These are proposed categories, not a formal standard. Reviewers should provide a brief reason for disputed cases so the product team can improve the workflow instead of optimizing toward an unexplained question count.

Improve the question before suppressing it

A broad question often becomes useful when made specific. Instead of asking for more details, identify the missing cargo group, unit, destination boundary, or authority. Explain why the answer matters to the next step. This helps the user respond quickly and reduces the chance of another clarification round.

Bundle related questions when they can be answered together, but avoid a long questionnaire containing issues that may become irrelevant. If one decision determines which later questions matter, ask that first and continue independent work. The best sequence follows the structure of the commercial problem, not the assistant's desire to fill every field immediately.

Preserve answers across the task. A user who has already authorized an internal comparison should not need to repeat that instruction after every new attachment. If the scope changes materially, identify the change and ask about that specific expansion. This keeps clarification connected to real decisions rather than becoming a ritual approval step.

Test both over-questioning and under-questioning

An evaluation set should include a complete enquiry, an enquiry with a genuinely missing decision, and a changed instruction that affects only part of the work. The assistant should proceed in the first case, ask a focused question in the second, and preserve unaffected progress in the third. These cases reveal whether it understands dependency rather than merely following a rule to ask whenever uncertain.

Zeaclub can explore this interaction problem with US desks using redacted tasks and observations of review effort. The proposed aim is fewer unnecessary interruptions while preserving material questions, not unrestricted autonomous action. The founder's stated YC inspiration supports learning from the user's actual friction instead of assuming that more automation is always better.

Choose one recent task in which an assistant asked too much or too little. Identify the information it already had, the decision it lacked, and the work it could have continued. That analysis produces a concrete improvement target: the right question at the point where its answer matters.

Sources

Related articles