A natural-language question still needs a governed definition
OpenAI’s September 10 announcement describes a Data Agent in ChatGPT Work that can connect to company data and context, answer questions, produce interactive dashboards, and support follow-up action. The product description says it uses business terms, metric definitions, custom calculations, and data relationships from semantic layers and trusted sources. It also describes reviewing evidence behind findings. Those are vendor descriptions, not independent proof of data accuracy in a buyer’s warehouse.
The practical lesson is not that every company needs an AI data agent. It is that a quick answer can only be trusted when the organization agrees on what the measure means. “Active customer,” “new lead,” “completed job,” and “weekly adoption” can each have several plausible definitions. A fluent explanation cannot resolve that disagreement unless a named owner has established the source and rule.
Put a person behind each important metric
Create a compact metric card for the questions managers ask repeatedly. Record the business definition, source table or system, included and excluded cases, time zone, reporting period, refresh timing, and accountable owner. Note any known delay or missing data. Keep the card close to the dashboard so a person can challenge the measure without searching through a separate document or guessing which analyst wrote the query.
Test the definition with edge cases. Does a canceled order count? Does a rescheduled appointment belong to the original or new week? What happens when the customer changes account? These details affect whether two dashboards tell the same story. Ask the data owner to approve the definition and the person responsible for operational action to confirm it answers the real question.
Trace the answer before allowing it to drive a task
When an AI tool produces an analysis, inspect the source and transformation before using it to set priorities. Verify filters, joins, date boundaries, and the comparison group. Ask it to show the calculation or query, then have a human who understands the data check the logic. A dashboard should show where its inputs came from and when they were refreshed. If a result cannot be traced, treat it as a prompt for investigation, not a management fact.
Keep the path from analysis to action visible. A recommendation to improve onboarding, for example, should identify the observed measure, the affected group, and the proposed owner. Separate “the data shows” from “we believe this change will help.” The second is a hypothesis until the organization measures it. This keeps a persuasive narrative from turning a noisy trend into an unsupported decision.
Make alert rules explicit before scheduling them
An alert needs an owner, threshold, comparison period, refresh schedule, and response plan. Decide how large a change matters and how many readings must persist before it triggers attention. Test with historical or synthetic cases to see whether the alert catches the scenario the business cares about. A model can help draft the rule, but a responsible person should approve what counts as a meaningful change and who receives the notice.
Start with one recurring question, one approved definition, and a dashboard that exposes its source and owner. Add automation only after the team can explain the metric and the response. A data tool may make it faster to ask questions; it does not remove the work of agreeing on what the answers mean.