ChatGPT Data Can Cut Weekly Reporting Handoffs

OpenAI added Data to ChatGPT Work and Codex on September 10, 2026. The business change is not another chart button: a team can move a weekly report from repeated analyst handoffs into one connected, refreshable workflow, while keeping metric definitions and human sign-off in the loop.

What ChatGPT Data actually is

Data is a plugin for analysis work inside ChatGPT Work and Codex. You give it a business question, it works across data sources your account is allowed to reach, and it can turn the result into an interactive dashboard or report. If those surfaces are new to you, the ChatGPT Work review explains where Work ends and Codex begins.

It is not a warehouse, a new copy of your database, or a replacement for every BI tool. OpenAI's Data setup guide recommends connecting a warehouse, supplying a semantic layer, and using ChatGPT Sites for the finished dashboard. A semantic layer is the approved meaning of the business: what counts as revenue, when a customer becomes active, which refunds are excluded, and how tables relate.

That definition layer is the hinge. Without it, Data can write a clean explanation of the wrong number.

The old reporting loop usually looks like this. A leader asks a question. An analyst finds the tables, rebuilds the query, exports the result, updates a chart, and answers follow-ups. Each turn creates another handoff.

With Data, the leader and analyst can stay in one thread. Data can query an approved source, use the team's metric definitions, show the filters and evidence, revise the analysis, and build the dashboard. The analyst's job moves from carrying numbers between tools to defining the metric, checking the query, and approving the result.

This is also different from Deep Research in ChatGPT Work. Deep Research is built for gathering and citing evidence across the web, files, and supported apps. Data is aimed at governed business analysis over connected operational data. One produces a sourced research brief. The other can become the recurring metric report your team opens every week.

The workflow moved, not the source of truth

The useful shift is fewer reporting handoffs. The risky shift is that a conversational request can now travel all the way from a warehouse to a shared dashboard.

Data can work with sources including Amazon Redshift, ClickHouse, Databricks, Google BigQuery, MongoDB, and Snowflake. It can also use business context from Google Drive and SharePoint, then work alongside BI tools such as Power BI, Tableau, Sigma, and ThoughtSpot when those connections are available. The exact actions depend on each tool and the user's access.

None of that makes the model the system of record. OpenAI tells users to check the source, time period, filters, and metric definition before relying on a result. If a Data answer disagrees with an existing report, those are the first four things to compare.

Teams already paying a handoff tax on finance packs, pipeline reviews, client dashboards, retention reports, and operating reviews benefit most. No approved source, stable definitions, or Data access means no recurring workflow. A one-off CSV may save clicks, but recurring reports carry the business case.

Four jobs that can change this week

A finance lead at a software company

The finance lead can ask Data to refresh a weekly revenue and retention pack from an approved warehouse view. The prompt names the reporting period, comparison period, and finance-owned definitions. Data drafts the variance explanation and shows what supports each chart.

The payoff is not automatic close. It is fewer trips between the analyst, spreadsheet, slide deck, and finance lead. Finance still signs off on the figures before the pack reaches leadership.

A revenue-operations manager

The RevOps manager can investigate why pipeline moved by region, using the approved pipeline definition and the CRM or warehouse data their account can access. Follow-up questions can separate a real sales change from a filter change, a late sync, or unmapped records.

The payoff is a shorter path from “the number moved” to “this is the segment and source row that moved it.” Sales operations still owns stage definitions and the decision to change a forecast.

An agency account lead

The account lead can build a client performance dashboard from approved advertising and commerce data, then ask Data to explain the largest movement before the weekly call. Brand rules can shape the dashboard without rebuilding the layout by hand.

The payoff is less recurring slide assembly. The agency must still check whether the client is allowed to see every dimension copied into the finished dashboard. Source access and dashboard access are not the same control.

A product-operations lead

The product-operations lead can trace a retention change across product analytics and the document that defines an active user. Data can compare cohorts, expose the time window and filters, and turn the checked analysis into a report for product and leadership.

The payoff is fewer rounds between the product manager and analyst. If the activation definition is disputed, Data will not settle the dispute. The team has to settle it in the semantic layer first.

How to pilot one weekly report

Do not begin with the executive dashboard. Start with a report whose owner, source, definitions, and current output are already known.

1. 1

   Choose the report and the decision

   Pick one recurring report and name the decision it supports. “Weekly revenue pack” is still too vague. “Decide whether retention needs an intervention before the operating review” gives the analysis a finish line.

2. 2

   Confirm access before prompting

   An administrator checks Workspace settings > Plugins > Data, chooses who can install it, and enables the required source connection. An individual user installs Data from the plugin directory. In Codex, open Sources, choose Use plugins, and select the installed plugin. The source app may still need its own setup and authorization.

3. 3

   Write the metric contract

   Give Data the approved source, metric definition, reporting period, comparison, exclusions, and owner. Point it to the semantic view or trusted dashboard that wins when definitions conflict. If the team cannot write this down, the workflow is not ready to automate.

4. 4

   Run it privately once

   Start with @Data so there is no ambiguity about which plugin handled the request. Ask for the dashboard, the calculations behind it, source freshness, filters, unmatched records, and any assumption that could change the answer. Tell it to keep the result private.

5. 5

   Check the evidence

   Compare the decision-carrying figures with the current report. Confirm the source, period, filters, definition, and excluded records. Ask Data to explain any mismatch before you accept a cleaner-looking chart.

6. 6

   Measure, then schedule

   Record hands-on review time, Work or Codex usage, warehouse activity, and any BI cost. Schedule a cloud refresh only after the manual run passes. The scheduled ChatGPT Work task checks the source and updates the dashboard. The Site itself does not run the schedule.

Here is a starting request you can adapt:

@Data, build our weekly revenue and retention report from the approved finance semantic view. Compare the last complete week with the prior complete week. Use the finance team's definitions, show the source freshness, filters, calculations, and unmatched records, and explain the largest changes. Keep the dashboard private. Do not publish, share, message anyone, or change source data. Ask me before using a different definition or source.

The first run should reveal weak source and definition boundaries, not hide them behind a finished dashboard.

The cost math that actually matters

OpenAI has not published a separate Data-plugin price. That does not make the workflow free. The bill can land in Work or Codex usage, warehouse compute, a BI license you keep, and the human review that remains.

For a concrete planning model, assume a weekly report currently takes 3.0 person-hours at a blended loaded cost of $80 per hour. That is $12,480 a year across 52 runs. If Data reduces assembly and review to 0.75 person-hours per run, the labor line becomes $3,120 and the modeled capacity released is $9,360.

Those are assumptions, not an OpenAI benchmark. Replace them with your own timer after the pilot.

The ChatGPT line uses an explicitly hypothetical refresh with 50,000 input tokens, 200,000 cached input tokens, and 10,000 output tokens on GPT-5.6 Sol. OpenAI's current new-Enterprise rate card charges $4.00, $0.40, and $20.00 per million tokens respectively. That makes the example $0.48 per refresh and $24.96 a year.

Business and credit-based Enterprise workspaces use a different unit. The current credit rate card prices GPT-5.6 Sol at 100 credits per million input tokens, 10 per million cached input tokens, and 500 per million output tokens in Work and Codex. Standard Business seats use included limits first, then eligible work can draw from purchased workspace credits.

The Snowflake row assumes one X-Small warehouse start per weekly refresh and a run no longer than its 60-second minimum. Snowflake bills X-Small at one credit per continuous hour, charges at least 60 seconds on every start, then bills per second. Data may issue more queries, keep the warehouse awake longer, or hit a larger warehouse. Use the warehouse history, not this floor, for the real bill.

Power BI is optional in OpenAI's recommended setup. Microsoft currently lists Power BI Pro at $14 per user per month when paid yearly, but an existing license may be a sunk cost and checkout pricing varies. Do not add a BI seat just because Data can connect to it.

The decision formula is straightforward:

Annual value = reporting runs × hours removed × loaded labor cost, minus incremental Work usage, warehouse compute, new BI licenses, and remaining review.

If this workflow causes you to buy a ChatGPT seat, include the seat. If your team already pays for the seat and the BI tool, use their marginal cost instead. The pilot should replace every assumption with an observed number.

The permission trap is after the query

Data queries use the connected account's existing permissions, including table, row, and column restrictions. Installing the Data plugin does not grant the warehouse connection, and connecting the warehouse does not grant access beyond the authorized role.

Publishing creates a second permission boundary. OpenAI says the data used in the analysis is copied into the published ChatGPT Site. The Site then has its own audience setting, which can include selected people, a workspace, named external viewers, or the public when the account and administrator allow it.

The practical consequence is that someone who could not run the source query may still be in the audience for data copied into the Site. Review the dashboard's fields and audience as a new publication, not as a live window that automatically inherits warehouse permissions.

A new Site starts limited to its owner and workspace administrators until access changes. Enterprise public publishing is off by default, but that is only one control. Named external sharing and workspace-wide sharing can still widen the audience when enabled.

The honest limits

Plan availability is not clean enough for a universal promise. The plugin directory exists across ChatGPT plans, but each plugin depends on the plan, workspace, role, region, surface, rollout, and included capabilities. The Data guide does not publish an exact eligible-plan list. Check whether Data appears in the account and workspace you intend to use before buying around it.

Sites has another open price question. It is in public beta for ChatGPT workspaces, Plus, and Pro accounts, subject to rollout and controls. Usage is included only up to plan-specific beta limits, and OpenAI says full pricing information is still coming.

Automation consumes usage too. Supported Work and Codex automations are metered with the model activity under the applicable rate card. OpenAI's dashboard workflow says a cloud task can continue when your laptop is off, while a desktop task needs the computer on and the app running. Choose the runtime deliberately, then watch both usage records.

Data also cannot repair weak business logic. A missing semantic layer, disputed KPI, incomplete account access, stale source, or unsupported tool action will still produce a gap. The dashboard may make that gap prettier. It does not make it smaller.

The Monday move

Act this week if your team rebuilds the same report, can name its owner and approved source, and has stable metric definitions. Wait if Data is absent from the intended workspace, the source connection has not passed review, or two teams still calculate the KPI differently. You are largely unaffected if an existing BI refresh already gives the right people a checked report with little manual handoff.

On Monday, take one report that is already due and run it privately through Data. Record the current assembly time first. Give Data the approved view and metric contract, then compare every decision-carrying figure with the existing report. Record Work usage, warehouse activity, and review time. Do not schedule or publish until the result passes that check and the Site audience has a named owner.

For more plain-English operating guides when AI changes the workflow math, subscribe to the newsletter.