Back to all articlesAI answer quality

Decide Whether an AI Answer Can Use an Estimate

Alactic Editorial9 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

A customer asks how much work a data import will take. The assistant knows the record count but not the file condition, preparation steps, or handling method. It supplies a precise number of hours. The arithmetic may be correct, but the answer can still mislead: the inputs do not yet describe the job.

The review question is not simply whether a number can be calculated. It is whether the available evidence supports using that number for the customer's decision. This article follows one fictional import-preparation task from an unsupported staffing request to a useful, limited estimate.

1. Define the decision and the quantity

Before checking arithmetic, write down what the customer will do with the answer. Exploring whether work deserves investigation is different from reserving staff or committing to a completion date.

  • Quantity: What is being estimated—hands-on labour, elapsed time, records processed, or something else?
  • Decision: What action would the answer support?
  • Scope: Which activities are included, and which remain unresolved?
  • Consequence: What happens if the result is materially wrong?

Do not let the quantity change silently. Person-hours are not calendar hours. Preparation effort is not successful import completion. A subtotal for routine work does not describe the whole job.

A provisional decision still has consequences. Reserving staff consumes capacity even if the reservation can later change. If the evidence supports only further investigation, the answer should help organise that investigation rather than imply that staffing is ready to plan.

2. Classify the answer with one decision table

Use this table before drafting the customer response. Classification depends on the evidence and intended use, not on whether the result is a single number or a range.

Answer typeWhen it fitsWhat to disclose
CalculationThe inputs, units, scope, and rule are supplied or established.Input origins and the rule used. Correct arithmetic does not validate supplied inputs.
Conditional estimateMaterial assumptions have a relevant basis, and remaining uncertainty is acceptable for the stated decision.Basis, conditions, exclusions, and the decision it can support.
Sensitivity illustrationThe reader needs to understand how a deliberately varied input changes the result.Which input was chosen, why it was varied, and what the illustration helps decide.
Clarification without a totalMissing information could change the method, scope, or suitability of the answer.The blocking evidence gap and the next useful request.

A range needs evidence for its endpoints. Two selected scenarios do not establish lower and upper bounds. Similarly, a known component can be calculated without estimating the total, but it should remain labelled as a component.

3. Keep evidence separate from assumptions

A reusable review record needs four input categories: provided values, derived values, justified assumptions, and unresolved inputs. The distinction prevents a missing value from becoming an invisible premise.

For example, an unspecified file condition must not silently become a clean-file assumption. Zero rework and zero waiting time are substantive claims, not neutral defaults.

Check whether evidence transfers to the proposed task. Was the same activity timed? Did timing include checking and correction? Were difficult cases omitted? Will the same method be used? A genuine observation can still be a poor basis for the customer's decision.

Also separate variation within a known method from uncertainty about the method itself. Routine handling time can be explored through timing observations. Unknown manual reconstruction may require a separate scope assessment before any rate is useful.

4. Filled worksheet: withhold the initial staffing estimate

Fictional worked example: Every quantity, observation, and result in this worksheet and its follow-up is hypothetical. They are not Alactic customer results or productivity benchmarks. The fields form a reusable review artifact; replace their contents with evidence from the actual task.

Decision and scope

  • Customer question: How much staff time should we reserve for import preparation?
  • Intended use: Provisional staffing.
  • Quantity: Person-hours of hands-on setup and routine preparation, not elapsed completion time.
  • Proposed included work: Setup plus routine handling of records using an agreed preparation method.
  • Unresolved scope: Duplicate resolution, missing-data reconstruction, mapping changes, and other exceptions.

Inputs and status

  • Provided: The customer reports 240 records.
  • Provided: A customer-selected sample of 12 records took 36 minutes to handle, excluding setup. The selection method is unknown.
  • Derived: The sample average is 36 ÷ 12 = 3 minutes per record.
  • Illustrative assumption: Setup takes 60 minutes. No activity breakdown or timing supports this placeholder.
  • Illustrative sensitivity input: Handling takes 5 minutes per record. This is deliberately selected, not observed.
  • Blocking unknowns: Sample coverage, setup activities, handling variation, and exception work.

Internal sensitivity calculations

Keep these calculations in the review record. Offer them separately only if the customer needs to understand the drivers—for example, whether further investigation should concentrate on setup or per-record handling.

  • Illustration A: 60 + (240 × 3) = 780 minutes, or 13 person-hours.
  • Illustration B: 60 + (240 × 5) = 1,260 minutes, or 21 person-hours.
  • Driver: At this record count, each additional minute of average handling adds 240 minutes, or 4 person-hours. Each additional setup minute adds one person-minute.

The two totals are sensitivity illustrations, not staffing estimates or bounds. The alternative rate demonstrates the arithmetic's dependence on handling time; it does not establish a likely slower case. That makes it useful for explaining why handling evidence matters, not for reserving capacity.

Review decision and evidence request

  • Decision: Withhold the staffing estimate.
  • Setup evidence: List the activities and obtain relevant timings or task-based allowances with a stated basis.
  • Handling evidence: Establish how the original sample was selected, then observe the intended method across known file variations.
  • Scope evidence: Identify exceptions and decide whether to assess them separately.
  • Owner: A designated operations reviewer reassesses suitability for staffing after the evidence is collected.

Default customer-facing answer

We cannot yet recommend a staff reservation. The current timing sample may not cover the file's variations, setup has not been scoped, and exception work is unresolved. Next, we need a setup breakdown, timing observations using the intended method, and a review of records needing non-routine work.

This response gives the customer a next step without making unsupported totals the most memorable part of the answer.

5. Follow-up worksheet: approve a limited conditional estimate

Continue the same hypothetical task. The reviewer obtains additional evidence and narrows the decision. The question is now whether to provisionally allocate routine preparation capacity, while keeping exceptions in a separate queue.

Revised evidence

  • Count and scope: An authorised review confirms 240 records. A documented file review separates 216 routine records from 24 exceptions needing individual assessment.
  • Routine definition: Routine records have the required fields and use the agreed mapping. Duplicate decisions and missing-data repair are outside this work package.
  • Setup basis: A rehearsal using the same file structure and method takes 20 minutes for mapping, 15 for validation-rule checks, and 10 for preparing the working file: 45 minutes in total.
  • Handling basis: After setup, 36 routine records covering each identified routine layout take 108 minutes, including the agreed per-record checks. Those records are complete.
  • Remaining routine count: 216 − 36 = 180 records.
  • Transfer assumption: The remaining routine records use the same layouts, method, and checking steps as the observed work. The reviewer accepts this basis for provisional routine-capacity planning, not a completion commitment.

Calculation and approved use

The observed routine average is 108 ÷ 36 = 3 minutes per record. Estimated remaining routine handling is therefore 180 × 3 = 540 minutes, or 9 person-hours. Setup and the observed batch are already complete, so they are not added again to the remaining-work estimate.

For reconciliation, the routine package comprises 45 minutes of completed setup, 108 minutes of completed handling, and 540 minutes of estimated remaining handling: 693 minutes, or 11 hours 33 minutes. That mixed actual-and-estimated total belongs in the worksheet; the immediate staffing decision concerns the remaining work.

  • Approved use: Provisionally allocate about 9 person-hours to remaining routine preparation, with a checkpoint after the next batch.
  • Exclusions: The 24 exception records, changed mapping, import execution, waiting for answers, and subsequent rework.
  • Revision triggers: A new layout, changed checking requirements, or records that no longer meet the routine definition.
  • Unresolved: Total project effort and elapsed completion time.

Customer-facing answer

For the remaining 180 routine records, allow about 9 person-hours of hands-on preparation, conditional on the same method and record layouts. This uses the observed handling time for the completed routine batch. It excludes the 24 exception records, import execution, and waiting time. Use it for a provisional routine-work allocation, with a checkpoint after the next batch—not as a completion deadline.

The evidence now matches a narrower work package and decision. Setup has been observed, completed work is separated from remaining work, and exceptions have their own queue. This supports the limited estimate without establishing statistical confidence or guaranteeing that the remaining records will take exactly the same time.

6. Ask only for decision-changing evidence

Clarification should reduce a specific uncertainty, not become an exhaustive questionnaire. If a timing sample exists, ask how it was selected before requesting another one. If the method is unknown, establish the method before refining a rate.

In the initial worksheet, the useful requests concern setup, sample coverage, and exceptions. Asking for a preferred completion date would not repair those gaps. It might clarify urgency, but it would not justify a handling assumption.

Requesting an estimate does not authorise inspecting customer data, importing records, or starting work. Obtain the appropriate permission for proposed actions. A human reviewer should resolve scope questions that ordinary clarification cannot settle.

7. Correct unsupported numbers explicitly

If an assistant has already sent an unjustified total, quietly replacing it with a range leaves the original planning risk intact.

  1. Identify the earlier figure so the customer knows which answer is being corrected.
  2. Name the defect: explain the unsupported input or omitted scope.
  3. Withdraw the implied use: say whether the number should no longer support staffing or a commitment.
  4. Preserve valid arithmetic only where useful, with its limited status attached.
  5. State the recovery step: identify the evidence and reviewer needed for a revised answer.

A concise correction for the initial example would say that the earlier staffing figure relied on an unsupported setup allowance and an unverified extrapolation from the sample. It should not be used to reserve staff; the setup, routine handling, and exception assessment must come first.

Record a specific failure category—such as invented rate, unsupported allowance, omitted scope, unit mismatch, or scenarios presented as bounds—so the review can change future answers.

8. Check whether the rule improves real answers

Review a small batch of numerical answers after removing unnecessary customer information. For each, reconstruct the intended decision, input origins, units, assumptions, exclusions, and reason for presenting a number.

Count unsupported premises separately from arithmetic errors. Check whether customer-facing wording gives a result more authority than the worksheet permits. Also look for unnecessary withholding: an assistant should provide a useful conditional estimate when the evidence supports one.

Test missing inputs and conflicting values. Does the answer ask the question that changes the decision, or silently fill the gap? Can a reviewer trace every customer-visible number to its basis and permitted use?

The observable outcome is a clearer planning choice: use a supported estimate, inspect a limited calculation, collect targeted evidence, or defer the decision. The purpose is not to maximise the number of numerical answers. It is to make the answers safe to interpret.

AI answer qualityEstimatesCustomer communicationOperational planning

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications