A GTM agent needs more than CRM access
OpenAI's internal data agent layers annotations, code, memory and live data around the model instead of asking it to guess.
What happened
OpenAI described the internal data agent it built for its own data platform in January 2026. Teams across its Go-To-Market, engineering, finance and research functions use the tool for high-impact data questions.
The agent does more than turn a natural-language question into a query. It explores data, runs queries, synthesises findings and carries corrections between turns.
What the source says
OpenAI says high-quality answers depend on accurate context, because strong models still misread internal terms or select the wrong data without it. Its agent combines six layers: table usage, human annotations, code-derived enrichment, institutional knowledge, memory and live runtime context.
Human annotations capture business meaning and known caveats that schemas cannot show. The agent also retrieves documents with metadata and permissions, so access control follows the user rather than the conversation.
OpenAI evaluates the system against curated question-and-answer pairs with manually authored golden queries. It compares the generated result with the expected result, then uses those checks as regression canaries during development.
This is a description of one internal OpenAI system rather than an independent benchmark. The source reports its own implementation, so its design lessons travel better than any implied performance claim.
What we think
Giving a GTM agent access to a CRM is not the same as giving it context. A contact record can tell it who somebody is, but it rarely explains why an account mattered, which evidence changed a score or what a previous reply meant.
The useful pattern is not six layers copied exactly. It is the separation between raw records, human definitions, workflow rules, corrections and live signals, because each answers a different question the agent would otherwise guess.
This also gives the owner a reason to keep a correction log, because a rejected prospect, false trigger or bad message gives the next run context it needs.
Giving a GTM agent access to a CRM is not the same as giving it context.
What to do about it
Start with the definitions that change an outbound decision. Write down what makes an account qualified, which signals count, what disqualifies a contact and when research goes stale.
Keep corrections where the workflow can retrieve them. If a person changes a score or rejects a draft, record why in a structured field instead of leaving the reasoning in a chat thread.
Test the system on a small set of accounts you know well. Compare its research, score and draft with the expected answer before you widen its access or volume.
Getting those definitions out of your head and into the workflow is part of what the Diagnostic produces.
Sources
How we choose and label these: our sourcing standard.
-
Inside OpenAI's in-house data agent →
OpenAI's own engineering account of an internal-only data agent. It documents the system's design and usage, not an independently measured benchmark or a product available to customers.