Reader question: How can a small team use an AI data tool without turning an ambiguous metric into a confident business decision?
On September 10, OpenAI announced a Data agent in ChatGPT Work that can connect to approved company data, investigate changes, and build interactive dashboards. On September 9, Google described Workspace capabilities that can create spreadsheets from Drive context, turn email threads into briefs, and convert documents into presentations. OpenAI’s Data agent announcement Google Workspace’s agentic capabilities announcement
These are vendor descriptions, not independent tests. They do mark a useful change: more people can ask for an analysis in ordinary language and receive a first chart or dashboard without waiting for a specialist.
The control a small team needs is not a prettier dashboard. It is a short metric contract written before the question is run. The contract says what the number means, which source can support it, who owns the definition, and what the result is allowed to change.
A chart is not a definition
“Show me whether retention improved” leaves several decisions open. Which customers count? What event starts the clock? What period is compared? Are pauses, refunds, reactivations, or internal test accounts included? A model can choose an answer silently and still produce a legible chart.
Write the contract in plain language:
| Field | Record before analysis |
|---|---|
| Decision | What decision could this analysis inform, and who owns it? |
| Metric | Numerator, denominator, unit, inclusion rules, exclusions, and formula. |
| Time | Date range, comparison, time zone, and data cut-off. |
| Source | Dataset or report, version, refresh time, and access owner. |
| Context | Approved business definition, known breaks, and unanswered questions. |
| Boundary | Draft, recommendation, or approved action; what the result must not do. |
The last field matters. “Prepare a budget discussion” is different from “move budget.” “Create a dashboard preview” is different from “publish a dashboard to the whole company.” Keep the proposed action separate from the analysis that may inform it.
OpenAI’s current Data guidance makes the same dependency explicit. It recommends a semantic layer with authoritative definitions and queries, asks users to name the source, metric, time period, and comparison, and tells them to check the source, filters, and metric definition before relying on a result. OpenAI’s Data plugin guidance
Run the question in four passes
1. Freeze the decision. Start with one question and one owner. “Why did sales change?” is an investigation, not a decision. Add the boundary: “Prepare three explanations for the weekly review; do not recommend a pricing change.” This keeps exploration from becoming an instruction.
2. Bind the source. Name the approved dataset, report, or semantic definition in the prompt. Record the date and refresh time. If two reports use “active customer” differently, keep both definitions visible and ask for a comparison. Do not merge them because the labels look similar.
Access is part of the source record. OpenAI says Data queries use existing table, row, and column permissions, and that installing a plugin does not grant access to an underlying app. A successful connection does not prove that the dataset is complete. OpenAI’s Data access guidance
3. Ask for evidence before prose. Request the value, formula, source, time window, filters, freshness, and limitations before asking for an explanation. Ask which evidence supports each chart. If the path from number to source is missing, label the result a lead or draft, not a verified readout.
For a retention question, a useful first response might say: “The rate is 42% for the stated cohort and period; reactivated accounts are excluded; the source refreshed at 08:00; causation is not established.” It gives the next person something to check.
4. Build a preview, not a conclusion. Create a dashboard only after the contract and evidence pass exist. Every prominent number should carry its period, definition, and source path. Keep an “unknowns” area for missing data and conflicting definitions. Compare the preview with an existing report; differences are a review task, not proof that the older report is wrong.
The dashboard can be useful before the answer is ready. Its job is to make the question easier to inspect, not to make uncertainty disappear.
Give the handoff four owners
Name the roles in the packet:
- Metric owner: maintains the business meaning and inclusion rules.
- Data owner: confirms source access, refresh, quality, and known gaps.
- Analysis owner: runs the question, preserves the prompt or query, and records the evidence.
- Decision owner: decides whether any operational change should follow.
One person can hold several roles. The handoff contains the contract, source and freshness, output, evidence notes, unresolved questions, and status: draft, recommendation, or approved action. The decision owner reviews the underlying source, not only the summary.
Sharing deserves its own check. OpenAI says a dashboard published through Sites copies the analysis data into the published site. Check the audience before sharing a preview, even when the original connection was restricted. OpenAI’s dashboard-sharing guidance
Stop when the number cannot carry the decision
Stop and mark the result unresolved when the definition differs from the team’s agreed meaning, the time window or time zone is missing, the source is stale or incomplete, the tool cannot show supporting evidence, a restriction may hide relevant rows, or the chart implies causation that the analysis did not test. Also stop when a preview is about to become a published dashboard, an email, a changed record, or another external action without a named approval.
An unknown is a status, not a zero. A dashboard that exposes the missing definition is doing useful work. A dashboard that hides it is only making the next mistake easier to distribute.
The practical sequence is simple: agree on the number, bind it to a source, ask for the evidence, preview the result, and hand the decision to a named owner. AI can make the first chart cheap. The metric contract is what keeps that chart from becoming an accidental policy.
Sources and limitations
- OpenAI, “Now everyone can put data to work” — checked September 14, 2026; supports the September 10 announcement, Data agent capabilities, connected sources, semantic context, dashboards, and connected-account permissions. Limitation: this is a vendor product announcement containing company and partner claims; it is not an independent accuracy or adoption study.
- OpenAI Help Center, “Using the Data plugin in ChatGPT Work and Codex” — checked September 14, 2026; supports the need to specify source, metric, time period, and comparison; use of semantic definitions; checking filters and definitions; dashboard refresh and sharing; and the warning that published dashboards copy analysis data. Limitation: availability and actions depend on plan, workspace, connector, account permissions, and product changes.
- Google Workspace, “Less switching, more flow: 5 new agentic capabilities across Google Workspace apps” — checked September 14, 2026; supports the September 9 announcement’s examples of generating spreadsheets, briefs, presentations, and cross-app tasks. Limitation: this is Google’s product description; rollout, plan eligibility, source selection, and output quality vary.