ChatGPT Work Data agent
Vintage: 2026-09. Primary evidence is OpenAI’s Now everyone can put data to work (dated September 10, 2026) plus official @ChatGPT X in 2026-09-10-openai-data-agent-in-chatgpt-work-connects-warehouses-and-bi (
method: grok-bot;x_video: true). Issuer post is grain. VentureBeat is secondary (no published accuracy benchmark; competitive framing is discourse). Whisperbaseon the announce video produced unintelligible tokens (likely music/SFX) — transcript is not capability grain.
ChatGPT Work Data agent
One-line summary: On 10 Sep 2026 OpenAI launched a Data plugin for ChatGPT Work that (issuer) connects approved company warehouses/files/BI, answers in natural language, investigates what changed, and builds shareable interactive dashboards without writing SQL. No published external retrieval/accuracy benchmark.
What it is
A dated ChatGPT Work plugin — listed as Data in the Plugins directory; users start with @Data. Distinct from the overnight SMB “16 plugins” video (flagged, not grain) on gpt-6-astra. Not a model SKU and not a ChatGPT Work rewrite.
Why it matters to this thread
Enterprise knowledge-work agents and lab product surfaces are in-scope. This is a domain-specific data agent (warehouse / BI / files), adjacent to chatgpt-super-assistant-vision’s “actions + artifacts” framing — not evidence that general-purpose agents hit escape velocity.
Key facts (from 2026-09-10-openai-data-agent-in-chatgpt-work-connects-warehouses-and-bi)
Issuer page (grain)
- From 2026-09-10-openai-data-agent-in-chatgpt-work-connects-warehouses-and-bi (Now everyone can put data to work, dated September 10, 2026): Data agent in ChatGPT Work “connects to your company data, investigates what changed, and builds interactive dashboards you can share,” directed in one conversation without writing queries.
- From the same source (same issuer page): named data / platform connections — Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, “and more”; files from Google Drive and SharePoint. Semantic/context partners cited: Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and BI dashboards.
- From the same source (same issuer page): named BI build/interact targets — Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot.
- From the same source (same issuer page): Enterprise admins choose which connections and roles are available; “Queries enforce the connected account’s existing permissions, including table, row, and column restrictions.”
- From the same source (same issuer page): listed as Data in the Plugins directory in ChatGPT Work; admins can make it available or install via Workspace settings → Plugins; users start with
@Data. - From the same source (same issuer page; issuer claim, not independently verified): “Nearly all of our product team and over two-thirds of our GTM organization use data agents in ChatGPT Work.” Alpha / customer quotes on the same page (NTT DATA, Thermo Fisher, ServiceTitan, Zipline, Empower, and others) are marketing testimonials, not measured outcomes.
Official X (grain for announce; video not grain)
- From the same source (@ChatGPT, 2026-09-10 ~15:05 UTC): product-account announce of the Data agent with the same issuer URL and an attached video. Discourse reposts are secondary.
- From the same source (
## Video transcripts; Whisperbase, 1m 11s, detected English, uploader @ChatGPT): transcript body is unintelligible (awia/Uık/ repeateddeeply). Do not treat demo visuals or this transcript as verified capability.
VentureBeat (secondary — not grain)
- From the same source (VentureBeat, Sean Michael Kerner, 2026-09-10): OpenAI did not publish a retrieval accuracy / correctness benchmark for the external Data agent (internal comparison only). Quotes GM Arpan Shah that the agent stitches partner contexts “in their state” rather than building a persistent customer context layer. Competitive framing vs Anthropic Cowork and Databricks instructed-retriever claims is VB analysis — discourse, not issuer fact.
What this source does not establish
- No published external accuracy / faithfulness benchmark. Issuer page emphasizes connectors and workflow, not measured answer quality. VB’s “skipped the benchmark” line is secondary restatement of that gap.
- Whisper transcript is not grain. Likely music/SFX demo; do not invent spoken claims.
- Internal adoption and customer quotes are issuer marketing, not independently verified outcomes.
- Not a rewrite of gpt-6-astra (chat-picker vs Work/Codex stays open there), gpt-tv, overnight SMB 16 plugins, or openai-defense-factory.
- Not the afternoon chatgpt-for-financial-services offering. Same-day Work surface, different product: this page is the Data plugin (warehouses/BI/
@Data); FS is a tailored eligible-institution build with hosted market data + firm templates. Do not merge. - Not proof the Data agent outcompetes warehouse-native agents (Databricks Genie, Snowflake CoCo, etc.). Issuer partners with those vendors; competitive boundary left open.
- Did not create ChatGPT Work, Arpan Shah, connector-vendor, or BI-tool pages.
- No ticker. No invented prices or dates.
- Encyclopedia is not a source.
Contradictions / tensions
- Issuer workflow vs measured quality. Connectors, permissions, and
@Datainstall path are dated issuer text; answer correctness is unpublished. - Orchestrator vs warehouse-native agent. Named partners include Databricks Genie Ontology and Snowflake Horizon — does this mainly stitch their context “in their state” (VB/Shah, secondary) or replace those agents? Unresolved.
- VB competitive frame vs issuer partnership. VB’s Anthropic Cowork / Databricks instructed-retriever contrast is analysis, not an OpenAI claim.
Open questions
- Will OpenAI publish an external retrieval / faithfulness benchmark for the Data agent?
- Does “ChatGPT Work Data agent” outcompete warehouse-native agents, or mainly orchestrate across them?
- What does Whisper fail to capture in the @ChatGPT demo video — on-screen UI only, or a voiceover the
basemodel missed?