What AI must do when operational data is missing or contradictory

Operational data is rarely clean enough for silent inference

Incident, change, project, vendor and risk records are created by different teams for different purposes. Updates arrive at different times, terminology varies and the authoritative source for a statement may not be obvious. Contradiction is therefore a normal operating condition—not an exceptional edge case.

An AI system that smooths these differences into a single polished narrative can increase risk. Fluency may hide the fact that the business impact was never recorded, the change outcome is disputed or the named owner has not accepted accountability.

Use explicit evidence states

Every material statement should carry a visible evidence state. The labels do not need to be complicated, but they must distinguish authority from interpretation.

• Verified — supported by an identified authoritative record
• Owner reported — supplied by the accountable person but not independently verified
• Conflict — credible sources disagree or cannot be reconciled
• Missing — required information is absent or too stale to use
• AI observation — a machine-generated pattern, question or draft that requires human review

What AI may do

AI can locate related records, identify missing fields, compare timestamps, flag contradictory statements, summarise verified evidence and draft a question for the responsible owner. It can also assemble a management brief from approved inputs while preserving links back to source records.

These activities reduce coordination effort without changing decision authority. The value comes from making the evidence path easier to follow and the unresolved work easier to assign.

What AI must not invent

AI must not fabricate business impact, root cause, test completion, approval status, financial benefit or owner acceptance. It must not treat the newest update as automatically authoritative, and it must not silently choose between conflicting sources simply because one version produces a cleaner narrative.

When the required fact is unavailable, the correct output may be a question, a blocked status or a documented no-go. That is a governance feature, not a system failure.

Keep human authority explicit

A named accountable person should review, edit, approve or reject management outputs before publication or action. The workflow should retain who approved the output, what evidence was available at that time and which material edits were made.

Human-in-the-loop should not mean an undefined person clicks an approval button. It should mean that authority, responsibility and escalation routes are designed into the workflow before the AI component is introduced.

A safe pilot boundary

Begin with read-only exports or approved samples, one recurring management workflow, named data sources and a defined approver. Measure whether the pilot improves evidence coverage, reduces reconciliation effort, shortens the time to identify gaps and produces decisions that remain traceable.

Do not begin with autonomous operational changes. First prove that the reporting workflow can represent facts, uncertainty and authority correctly.