ChatGPT Work Data Agent Enterprise Review: Should Your Data Team Care?
OpenAI's Data agent in ChatGPT Work queries Snowflake, Databricks, and BigQuery in plain language. What it does, what setup really costs, and who should pilot it.
TLDR: OpenAI’s Data agent, launched September 9 inside ChatGPT Work, lets business users query governed company data in plain language and get interactive dashboards back. It’s a credible threat to ad-hoc analyst requests and spare Tableau Viewer seats, not to your BI platform. The catch: OpenAI’s own data team admits broad internal use only worked after building shared business definitions, access rules, and safeguards for sensitive data. Pilot it if your semantic layer (dbt, Snowflake Horizon, Databricks Genie ontology) already exists; wait if your metric definitions still live in people’s heads.
Why This Review Matters Now
On September 9, 2026, OpenAI put a dashboard builder inside the chat window your employees already have open. The Data agent connects ChatGPT Work to Snowflake, Databricks, BigQuery, Redshift, and your existing BI tools, then answers questions like “diagnose why weekly active users changed last week” with an investigation plus an interactive dashboard. Eleven alpha customers including NTT Data and Thermo Fisher are already using it for sales, spend, and staffing analysis.
But you’ve seen this movie before. Every conversational analytics launch promises self-service; most of them land on the data team’s backlog as another governance problem. This review evaluates the Data agent the way an IT leader should: what it actually connects to, what the admin setup demands, where it threatens existing spend, and where warehouse-native alternatives like Snowflake Cortex and Databricks Genie beat it.
Data Agent At a Glance
| Capability | OpenAI Data Agent | Tableau / Power BI | Snowflake Cortex / Databricks Genie |
|---|---|---|---|
| Best for | Ad-hoc questions from non-analysts | Governed reporting, certified dashboards | In-warehouse governed Q&A |
| Query interface | Natural language in ChatGPT Work | Drag-and-drop + semantic models | Natural language inside the warehouse |
| Connectors | Redshift, BigQuery, Snowflake, Databricks, MongoDB, ClickHouse, Datadog, Drive, SharePoint | Hundreds of native connectors | Native to own platform (+ federation) |
| Metric definitions | Imports from dbt, Genie ontology, Horizon, existing dashboards | Semantic layer / datasets you build | Native semantic views / Unity Catalog |
| Governance | Inherits connected-account permissions (table, row, column) | Row-level security, certifications | Warehouse-native RBAC, masking, lineage |
| Dashboard output | Interactive, shareable, refreshable; can also build inside Power BI, Tableau, Sigma, ThoughtSpot, Omni, Oracle BI | Pixel-controlled, versioned, certified | Native dashboards (Snowsight, Dashboards) |
| Pricing model | ChatGPT Work seat (Business/Enterprise); no separate Data price disclosed | Creator/Explorer/Viewer tiers or capacity | Warehouse compute + platform licensing |
| Admin effort | Shared definitions + access rules + safeguards required up front | Semantic model maintenance (ongoing) | Semantic view curation (ongoing) |
What Launched on September 9
The Data agent lives in the Plugins directory inside ChatGPT Work. An administrator installs it from Workspace settings, enables the relevant source plugins (Databricks, Snowflake, and others), decides which roles can use it, and workers start conversations with @Data. There is no separate SKU. If you hold ChatGPT Work, you hold the Data agent.
Under the hood it does three things in one conversation: retrieve (query approved sources through the connected identity), interpret (apply your organization’s business terms, metric definitions, custom calculations, and table relationships), and present (build an interactive dashboard your team can edit, share, and refresh). It can also reach into the BI tools you already own - Power BI, Tableau, Sigma, ThoughtSpot, Omni, Oracle BI - and build or interact with dashboards there in plain language. That last detail matters more than the headlines suggest, and I’ll come back to it.
The semantic context comes from sources your data team already maintains: Databricks Genie ontology, dbt models, GitHub, Snowflake Horizon, and existing BI dashboards. Brand guidelines can be attached so outputs match your look and feel. After analysis, it recommends next steps, names who should be involved, shares findings over Slack or email, and carries out approved actions through connected tools - with a human approval step before anything executes.
Where the Definitions Do the Real Work
Here’s the part OpenAI deserves credit for admitting. The launch post states plainly that broad internal use required its data team to create shared business definitions, set access rules, and put safeguards in place for sensitive data first. Nearly all of OpenAI’s product team and over two-thirds of its go-to-market org now use data-agent capabilities - but only after that groundwork.
Earned insight: In every self-service BI rollout I’ve watched, the tool demo takes a day and the definitions take a quarter. One org I worked with stood up a governed semantic layer covering 40 core metrics before letting 300 sales users loose on conversational queries - and still spent the first month adjudicating “whose revenue number is right” disputes. OpenAI running its agent across 70,000 datasets and 600 petabytes internally sounds impressive until you realize it proves the opposite point: they needed enterprise-grade metadata (built on OpenMetadata, per their own case study) before the agent was safe to use. Your pilot budget should assume the same ratio - 10% tool setup, 90% definitions and access rules.
The Snowflake Angle: Partner and Rival
Snowflake’s quote sits prominently on OpenAI’s launch page: customers have “built a trusted foundation for enterprise data and context in Snowflake,” and with the Data agent, employees can tap into that data and access controls. Databricks’ Ken Wong says nearly the same thing about Genie. Both warehouses are cooperating.
But notice what’s happening. Snowflake simultaneously ships CoCo and CoWork - its own conversational experiences powered by the same frontier models - plus managed MCP servers that let any agent reach governed Snowflake data over an open standard. Databricks offers Genie with benchmark sets, generated-SQL review, manual corrections, and a monitoring queue. The warehouses are happy to let ChatGPT Work be one more query surface because the governance - and the compute billing - still happens on their side.
What this means in practice: if your data already lives in Snowflake with Horizon governance and clean semantic views, the Data agent is a nicer front door to a house Snowflake already built. If your data is scattered across Redshift, MongoDB, spreadsheets, and SharePoint with no shared definitions, the agent doesn’t fix that. It inherits your mess at chat speed.
Warning: Queries enforce the connected account’s existing permissions - table, row, and column restrictions. That sentence does a lot of quiet work. If your warehouse permissions are coarse (everyone in Sales sees all sales tables, including comp data), the agent will faithfully surface everything the connected identity can see. Audit your service-account and role permissions before enabling broad access, not after someone asks the agent for data they shouldn’t have.
Governance and Admin Effort: The Honest Cost
OpenAI lists the admin controls accurately: administrators choose which connections are available, which roles can use them, and the agent inherits source permissions. Role-gated connections plus inherited permissions are a reasonable starting point.
The problem is what the launch materials don’t disclose. There is no published accuracy benchmark, no failure-rate data from the alpha program, no latency or warehouse-query-cost figures, and no service-level commitments. Nobody outside OpenAI knows how often the agent returns a plausible-but-wrong number, how many alpha questions needed analyst rescue, or what a heavy month of warehouse compute driven by curious business users costs. These omissions don’t prove the product performs poorly. They define what you can’t verify before piloting - so your pilot must measure it.
Data Agent Governance Strengths:
- Inherits source-system permissions automatically (no parallel access model to maintain)
- Role-gated connections let admins scope who reaches which sources
- Draws on existing semantic layers (dbt, Horizon, Genie) instead of inventing its own
- Human approval step required before connected actions execute
Data Agent Governance Weaknesses:
- No published accuracy benchmarks or alpha failure rates
- No clarity on what approval (if any) gates a generated dashboard becoming a shared record
- A shared, refreshed dashboard can keep informing decisions after schemas or definitions change
- Warehouse compute costs from exploratory queries are unmetered and undisclosed
The sharpest independent critique I’ve seen puts it this way: OpenAI’s internal adoption percentages document reach, not accuracy. “Nearly all of our product team uses it” tells you nothing about how often the answer was right, what the cost per accepted analysis was, or what business results followed. Treat vendor adoption stats as distribution metrics. Your pilot needs correctness metrics: sample 50 representative questions, have analysts grade the answers, and track the intervention rate.
Tip: Before rollout, build a 30-question golden set with known-correct answers from your certified reports. Run every Data agent response against it weekly. Databricks ships benchmark sets for Genie for exactly this reason - steal the practice even if you don’t use their tool. If the agent’s pass rate on your golden set drops below 90%, it stays in pilot.
Where It Threatens Existing Spend
Three budget lines should make their owners nervous.
First, ad-hoc analyst time. The classic queue - “can you pull churn by segment for the QBR?” - is exactly what the agent compresses. Product managers at alpha customers like ServicePiston now run sales-funnel, support-ticket, and spend analysis themselves. At NTT Data, non-engineers in sales and corporate functions build and update dashboards in plain language. Every one of those requests that never becomes a ticket is analyst capacity returned.
Second, spare BI seats. Tableau Viewer and Power BI Pro licenses bought for executives who open three dashboards a quarter are vulnerable. If those users get answers in ChatGPT Work instead, renewals shrink. One alpha customer reportedly rebuilt a performance dashboard in half an hour and found errors in the original - the kind of story that travels fast in budget reviews.
Third, shadow analytics tools. Teams paying for lightweight BI point solutions because the central platform is too slow may consolidate onto the agent - provided governance approves.
But here’s the catch: none of these savings materialize without the definitions work described above. Cutting five Tableau Viewer seats saves roughly $900 a year. Standing up the semantic layer that makes the agent trustworthy costs analyst-weeks. The ROI math works at scale - hundreds of self-serve users amortizing one definitions project - and fails for a 20-person pilot that still needs the same groundwork.
Pricing Reality
OpenAI discloses no separate price for the Data agent. It ships inside ChatGPT Work, so your cost is the seat plus everything around it. Prices verified September 24, 2026.
| Cost component | What you’ll pay | Notes |
|---|---|---|
| ChatGPT Business | $20/user/month annual, $25 monthly | 2-seat minimum; includes Work agent, SSO, no-training guarantee |
| ChatGPT Enterprise | Custom, annual-only, volume-negotiated | Reported $45-75/seat range; renewal conversations increasingly include metered tokens |
| Data agent add-on | $0 disclosed | No separate SKU announced; bundled in Work |
| Definitions + access-rule setup | 4-12 analyst-weeks | Shared business definitions, role mapping, sensitive-data safeguards; OpenAI’s own prerequisite |
| Warehouse compute | Pass-through to your Snowflake/Databricks/BigQuery bill | Exploratory NL queries can be compute-hungry; set query governors before broad rollout |
| BI licenses retained | Unchanged in year one | You still need Tableau/Power BI for certified reporting; cut Viewer seats only after pilot proves substitution |
The honest total cost of ownership for a 200-person pilot: roughly $48,000-$60,000 a year in Business seats, plus a one-time definitions project, plus whatever your warehouse charges for the extra queries. Don’t let anyone present this as free because there’s no add-on SKU. The seat is the subscription; the definitions are the implementation; the compute is the meter you forgot to watch.
Earned insight: Watch the Enterprise renewal language. Multiple buyers report new and renewing ChatGPT Enterprise contracts shifting from pure per-seat subscriptions toward API-style metered tokens - with the subscription positioned as the better deal versus raw consumption. If that shift reaches the Data agent’s heavy query workloads, today’s “bundled at no extra charge” becomes tomorrow’s consumption line item. Negotiate a cap or a fixed-query allowance into your Enterprise agreement now, while the agent is still a sweetener rather than a meter.
Who Should Pilot, Who Should Wait
Good fit: You run Snowflake, Databricks, or BigQuery with a maintained semantic layer (dbt, Horizon, Genie ontology, certified datasets). Your analysts drown in ad-hoc requests. Your BI Viewer seats are underused. You have an admin team that can scope role-gated connections and audit permissions first. NTT Data’s profile - technical organization, governed warehouse, broad non-engineer population - is the template.
Not a good fit: Your metric definitions live in tribal knowledge. Your warehouse permissions are coarse. You’re a regulated shop that needs audit trails on every number shown to leadership - the agent’s evidence-review flow isn’t a substitute for certified reporting. Or you’re small enough that five analyst-weeks of definitions work exceeds a year of the BI seats you’d save.
Pilot It If:
- Semantic layer already exists and is maintained
- Ad-hoc request queue exceeds analyst capacity
- Warehouse permissions are already fine-grained
- You can staff a golden-set accuracy review weekly
Wait If:
- No shared definitions for revenue, churn, or headcount
- Coarse warehouse roles would overexpose sensitive data
- Regulated reporting requires certified, versioned outputs
- Team is too small to amortize the setup effort
Bottom Line
The Data agent is the best conversational analytics front door OpenAI has shipped, and the Snowflake/Databricks partnerships give it real governed data to stand on. But it doesn’t replace your BI platform, your semantic layer, or your analysts’ judgment - it moves the bottleneck from writing queries to defining metrics and reviewing answers. That’s a genuine productivity shift, not a headcount event.
Rating: 3.8 / 5 for governed self-service analytics. 2.4 / 5 as a BI replacement.
Your 30-day action: pick one business domain with clean definitions, stand up the Data agent for ten users against it, and grade 50 answers against certified reports before you touch a single Tableau license.
FAQ
Does the ChatGPT Work Data agent replace Tableau or Power BI?
No - and OpenAI doesn’t claim it does. The agent builds and interacts with dashboards inside Power BI, Tableau, Sigma, ThoughtSpot, Omni, and Oracle BI rather than routing around them. Think of it as a natural-language front door: it handles ad-hoc questions and first-draft dashboards, while certified reporting, pixel-controlled layouts, and versioned metrics stay in your BI platform. The realistic substitution is spare Viewer seats and ad-hoc analyst tickets, not the platform itself.
What data sources does the Data agent connect to?
At launch it connects to Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake, plus files and documents from Google Drive and SharePoint. Semantic context comes from Databricks Genie ontology, dbt models, GitHub, Snowflake Horizon, and existing BI dashboards. Administrators choose which connections are available and which roles can use them - users only see sources they’ve been granted.
Which ChatGPT plan do you need for the Data agent?
The Data agent ships inside ChatGPT Work with no separate SKU or price disclosed. ChatGPT Business runs $20 per user per month billed annually ($25 monthly, 2-seat minimum) and Enterprise is custom annual pricing. Either tier gets you the agent; the practical gating factor is admin capacity to configure connections and definitions, not the license tier. Verify current tier requirements in Workspace settings, since OpenAI hasn’t published Data-specific availability details.
How does the Data agent handle permissions and sensitive data?
Queries enforce the connected account’s existing permissions, including table, row, and column restrictions, and administrators role-gate which connections each group can use. That inheritance is the core safeguard - but it means your warehouse roles must already be fine-grained. If a role can see compensation tables, the agent will surface them. Audit service-account and role permissions before enabling broad access, and confirm sensitive-data safeguards with your security team during pilot.
How does it compare to Snowflake Cortex or Databricks Genie?
Cortex and Genie are warehouse-native: governance, RBAC, masking, and lineage happen where the data lives, with purpose-built SQL review and monitoring queues. The Data agent is warehouse-agnostic and conversational - nicer for cross-source questions spanning Snowflake plus Drive docs plus Datadog, weaker on deep warehouse governance. Many shops will end up with both: Genie or Cortex as the governed semantic engine, the Data agent as the chat front door. Snowflake and Databricks both endorse this pattern on OpenAI’s own launch page.
How accurate are the Data agent’s answers?
Unknown - OpenAI publishes no accuracy benchmarks, alpha failure rates, or cost-per-accepted-analysis figures. Internal adoption stats (“nearly all of our product team uses it”) measure reach, not correctness. Google warns its own conversational analytics can return plausible-but-wrong output, and the same caution applies here. Build a golden set of 30+ questions with known-correct answers from certified reports, grade the agent weekly, and keep it in pilot until the pass rate holds above 90%.
What does setup actually require from the data team?
OpenAI’s own words: shared business definitions, access rules, and safeguards for sensitive data - created before broad internal use. In practice that means defining core metrics (revenue, churn, active users), mapping table relationships, configuring role-gated connections in Workspace settings, auditing warehouse permissions, and building a golden-set test suite. Budget 4-12 analyst-weeks depending on how much semantic layer (dbt, Horizon, certified datasets) already exists.
Can business users share and refresh the dashboards it builds?
Yes - dashboards are interactive, editable, shareable, and refreshable, and brand guidelines can be applied. But treat sharing discipline as a governance decision: an editable draft supports exploration, while a shared chart may be treated as an official number and a refreshed dashboard keeps informing decisions after schemas change. Define which outputs count as certified versus exploratory before rollout, or you’ll inherit a shadow-BI sprawl problem wearing a ChatGPT badge.
Related Articles
Discussion