Your sales software says revenue is up. Your bank balance feels tighter. Your team is busy, but you are not sure which work is actually making money.
The answers may already exist in your business. They are just spread across reports, spreadsheets and systems that nobody has time to piece together.
OpenAI’s new Data agent in ChatGPT Work, announced on 10 September 2026, is designed to help people investigate business data by asking questions in ordinary language. It can analyse connected sources, create interactive dashboards and help turn findings into approved actions.
For a small business, the useful question is not “How many dashboards can we build?” It is: Which decision could we make better if we could get a reliable answer sooner?
What has OpenAI actually released?
According to OpenAI’s announcement, the Data agent can connect to approved company data sources, investigate changes and create dashboards that teams can edit, share and refresh. You can ask follow-up questions and review the evidence behind its findings.
Named data sources include Google BigQuery, Snowflake, Databricks, Amazon Redshift, MongoDB and ClickHouse. It can also bring files and documents from Google Drive and SharePoint into an analysis. OpenAI says it can work with dashboards in tools including Power BI and Tableau.
That is more specific than “ChatGPT now connects to every business app”. The announcement does not establish a direct connection to every accounting package, online shop, booking system or CRM. Your information needs to be available through a supported, authorised route.
OpenAI also says the agent can use shared business definitions and relationships to interpret data. Think of these as the rules behind your numbers: whether “sales” includes VAT, whether refunds are deducted, and what counts as an active customer.
Those definitions matter. Faster analysis of the wrong definition is still the wrong answer.
Can a small business use it straight away?
OpenAI places the feature in ChatGPT Work, where it appears as Data in the Plugins directory. Administrators can make it available through Workspace settings, configure the relevant data-source plugins and control who can use them. Once installed and connected, users can start a conversation with @Data.
Before buying anything, check whether your workspace has access, whether your data sources are supported, and what your subscription and any connected services will cost. The announcement does not give a complete plan-by-plan pricing breakdown or establish that this is included in every ChatGPT subscription.
If your business mostly runs on disconnected spreadsheets and a basic accounting package, do not commission a large database project just to try a new feature. First establish whether a supported file or existing data connection can answer one useful question. If it cannot, the setup cost may outweigh the benefit for now.
Five small-business examples
The following are illustrative workflows, not reported customer results. Each depends on having the relevant data available through an approved connection.
1. An online shop: find out why sales are up but margins are down
A small retailer runs a successful promotion. Orders increase, but the owner suspects that discounts, delivery costs and returns have swallowed the extra revenue.
Rather than requesting another sales chart, ask a specific question:
@Data Compare the last eight weeks with the previous eight. Break down net sales, refunds, discounts and gross margin by product. Use our agreed margin definition. Identify the biggest changes, show the supporting records and flag missing cost data.
The analysis might highlight a heavily discounted product or an increase in refunds. That is a lead to investigate, not proof of the cause.
What the owner does next: check the underlying costs and sample orders, then decide whether to change the promotion, delivery threshold or product mix. Do not let the agent change prices automatically during the first trial.
2. A plumbing or electrical firm: spot where enquiries stop becoming jobs
A trades business has plenty of enquiries but inconsistent bookings. The team suspects slow follow-up; the owner thinks certain job types are harder to quote.
Useful data would include enquiry dates, quote dates, quoted values, job categories and booking outcomes.
@Data Compare quote-to-booking conversion by job type and response time over the last three months. Separate open quotes from lost quotes. Show the number of quotes in each group and identify records with missing outcomes.
This could help distinguish an actual follow-up problem from a reporting problem. A batch of quotes labelled “open” might simply need updating.
What the owner does next: review a small sample with the office manager and test one change, such as a next-day quote follow-up. Avoid drawing strong conclusions from groups containing only a handful of jobs.
3. A marketing agency: identify busy clients that are not profitable
An agency’s largest client is not necessarily its most profitable. Extra revisions, unrecorded calls and work outside the agreed scope can make a healthy retainer look better than it is.
Bring together invoicing, time records and agreed scopes where supported.
@Data Compare invoiced revenue with recorded delivery time by client for the last quarter. Apply our approved internal hourly cost assumptions. Flag work above the agreed scope, missing time records and unpaid invoices separately. Do not treat unpaid invoices as received cash.
The result could become a short monthly client-review dashboard rather than another spreadsheet someone has to assemble manually.
What the owner does next: validate the time records with the account lead before changing a contract or discussing workload. Incomplete timesheets should not become a basis for judging an employee’s performance.
4. A café: understand waste before cutting the menu
A café owner notices food waste but does not know whether the problem is purchasing, preparation or demand on particular days.
If sales, purchasing and waste records are available, the agent could help compare them.
@Data Compare recorded waste and item sales by weekday over the last twelve weeks. Flag items with repeated high waste, and note promotions, closures or missing records that affect the comparison. Distinguish observed patterns from possible explanations.
What the owner does next: check the findings with the kitchen team and trial smaller preparation batches on selected days. Do not treat a quiet week as a reliable forecast for every future week, or assume recorded waste captures everything discarded.
5. A small wholesaler: make the weekly management meeting useful
A wholesaler may already have the information it needs, but someone spends Monday morning collecting sales, stock and unpaid-invoice reports.
@Data Create a weekly management dashboard showing sales against target, overdue invoices and low-stock items. Show the last refresh time and source for each section. Explain the three changes that need attention, and suggest who should review them. Draft the update for approval; do not send it.
OpenAI says dashboards can be shared and refreshed, and approved findings can be sent through connected tools such as Slack or email. That does not mean every dashboard refresh or action is automatically scheduled; check the actual workflow available in your workspace.
What the owner does next: confirm the figures, assign each follow-up to a person and approve any message. The point is a shorter reporting process and a clearer meeting, not a prettier chart.
Start with one question, not all your company data
A sensible first trial is one recurring question you already answer manually.
- Choose the decision. For example: which product promotions should we review this week?
- Define the numbers. Agree the date range, VAT treatment, refunds and relevant cost assumptions.
- Limit access. Use the smallest necessary dataset. Leave out bank details, sensitive personnel records and customer information the task does not need.
- Compare with a known answer. Ask the agent to reproduce a recent report your team has already checked. Investigate discrepancies before trusting new findings.
- Keep actions under human review. Start with analysis and drafts, not automatic refunds, purchasing or customer contact.
- Measure the whole job. Include setup, data cleaning and checking time—not just how quickly the first answer appears.
Continue only if the result is reliable enough to use and the total effort is lower than your existing approach.
Permissions help, but they do not replace judgement
OpenAI says administrators control available connections and roles, and queries enforce the connected account’s existing table, row and column restrictions. That is useful, but a broadly privileged account can still expose too much. Give the connection appropriate access in the first place.
Check your organisation’s applicable data terms, retention settings and obligations before connecting confidential information. Do not assume one product announcement answers every privacy question.
For important findings, ask: Where did this number come from? What is missing? What else could explain the change?
An agent can identify a relationship between slow quote responses and fewer bookings. It cannot establish from that relationship alone that response time caused the lost business.
The best use is a better decision
Small businesses do not need to copy an enterprise analytics department to benefit from this release.
If you have compatible data, clear definitions and someone who can check the answer, the Data agent could make routine investigation more accessible. If your records are inconsistent or your systems cannot connect, fixing one small data problem may be the more valuable first step.
Pick the question you keep putting off because the report takes too long. Test that. A dependable answer to one real business question is worth more than a dashboard full of numbers nobody acts on.
Source: OpenAI, “Now everyone can put data to work”, 10 September 2026. Product capabilities above reflect that announcement; availability, pricing and connection requirements should be checked for your workspace.