When missing data should remain missing in an AI workflow

An AI system may infer a customer category, date or identifier from context. That inference can be useful as a suggestion, but should not automatically become an established fact.

Identify why the value is absent

Missing information can reflect a collection error, an inapplicable question or a customer who has not supplied the detail. These conditions should not all receive the same treatment.

Scikit-learn’s imputation guidance describes methods for handling missing values in model inputs. Statistical imputation does not authorise an application to fabricate an operational record.

An illustrative CRM assistant may suggest an industry from a company description. The interface should retain that suggestion’s status until the organisation’s approved verification process accepts it.

Preserve original and inferred states

Record the supplied value separately from a proposed completion where the design permits. Show the source or reasoning needed for review without exposing unnecessary private information.

Define which fields may be inferred and which require direct confirmation. A missing internal categorisation differs from a missing payment destination or consent record.

Validate downstream actions so a plausible completion cannot bypass a required human or customer decision.

Measure the value of the suggestion

Track accepted, corrected and rejected completions using a suitable sample. Include the reviewer effort required to establish the answer.

If most suggestions require substantial investigation, the system may be moving work rather than removing it. A request for missing information could be the more useful next step.

Record the reason a field was completed and whether confirmation occurred. Downstream users should be able to distinguish customer-supplied information from an unverified suggestion. If the system exports both as ordinary values, the original uncertainty can disappear and later users may rely on a fact nobody actually established.

GAT’s SME automation analysis emphasises practical returns. Preserving uncertainty helps ensure those returns do not depend on turning unknown information into confident-looking data.