The ChatGPT Work Data agent turns company questions into dashboards
Original publication: · Sergey Zhadov
This is a translated archive article. The date and pricing discussion refer to the original publication, not a new announcement today.
ChatGPT Work now has a Data agent. It can connect to approved company warehouses and documents, investigate a business question and build an interactive dashboard. A user does not have to start with SQL, but nobody gets to skip metric and access checks.
Start with one job, not the whole corporate junk drawer
Pick a familiar report such as regional sales, return reasons or support volume. The team already knows the correct result, so it can quickly spot a bad date range, filter or metric definition.
Connect one source and only the documents the pilot needs. The agent follows the linked account's table, row and column restrictions. It cannot magically repair permissions that an administrator set up badly.
Ask a question that has a checkable answer
A request like 'show performance' is vague rubbish. State the metric, period, segment, exclusions and desired output. Ask the agent to list the datasets, filters and assumptions it used.
Compare the first dashboard with a closed reporting period. Check the total, several rows and the extreme values. When a number differs, find the cause before the chart reaches a management meeting.
An interactive dashboard is not a finished decision
The agent can continue the investigation, change a slice and prepare work for supported BI tools. A manager can ask a follow-up immediately instead of waiting for another spreadsheet tomorrow.
The trap is equally simple. A confident chart looks more trustworthy than a rough table, so an error becomes easier to miss. Keep human approval for material decisions and block automatic actions without confirmation.
Score the pilot with three boring but honest numbers
Measure time to the first useful answer, the number of manual corrections and the share of important figures with traceable origins. Log the questions where the agent confidently produced nonsense or hit an access wall.
Expand the pilot if reporting becomes faster without turning review into a separate nightmare. If the team spends longer fixing answers than it once spent building the sheet, repair the data, permissions and review rules first.
Translation reviewed:
