OpenAI and ServiceNow Push Agents Into IT Workflows — Most of It Ships Later

OpenAI and ServiceNow said on October 28 that they are widening a partnership first announced in January, with plans to embed OpenAI models and agents across ServiceNow products and to pipe ServiceNow business context and actions into ChatGPT and Codex. ServiceNow also plans to give every employee access to ChatGPT Work and Codex, expanding past its engineering organization. Read the availability section closely, though: one component is live today, and the more ambitious pieces are scheduled across 2027.

What is actually shipping now

According to the announcement, the ServiceNow MCP plugin is available in the OpenAI marketplace today, with what the companies describe as enhanced capabilities expected by the end of 2026. Everything else carries a future date. Voice capabilities inside ServiceNow's Super Intelligence Control Tower are expected in the first quarter of 2027. The jointly built security capabilities are slated for early access in January 2027 and general availability in the second quarter of 2027. Codex inside ServiceNow's AI Workflow Factory is expected sometime in 2027.

That distribution matters for anyone trying to plan against this. The near-term deliverable is a connector built on Model Context Protocol, the now-common plumbing for letting a model reach into an external system. The differentiated work — security triage, agentic application development, voice-driven intake — is roadmap.

ChatGPT as the front door to the service desk

The most concrete described experience is a ServiceNow integration that surfaces business context and workflow actions directly inside ChatGPT. OpenAI's example: an IT staffer finds similar past incidents, assesses the risk of a proposed change, and updates a ServiceNow record without leaving the chat window. Customers connect ChatGPT to their own ServiceNow instance and select which capabilities to expose.

The governance layer is ServiceNow's Action Fabric, which the companies say applies existing business rules and permissions so that what a user can do through ChatGPT mirrors their configured access in ServiceNow. That is the right architectural answer — enforcement belongs in the system of record, not in the chat client — but the announcement does not describe how logging, approval gates or audit trails surface to compliance teams, nor what happens when a model proposes an action a user is authorized to take but shouldn't. Write actions against a CMDB or an incident queue are a different risk category than retrieval, and the release does not quantify accuracy or error handling.

ServiceNow is also previewing interactive experiences rendered inside ChatGPT. In the scenario OpenAI describes, an on-call engineer reviews linked monitoring alerts, recent changes and affected services, then assigns a team and works toward a fix in the same session. This is part of a broader pattern: vendors are increasingly treating the chat assistant as an application surface rather than a text box. OpenAI published a similar expansion with Oracle a day earlier, on October 27, per its own index of posts.

Three planned product areas

The companies outlined development work in three buckets.

Application development. A planned Codex integration in ServiceNow's AI Workflow Factory is meant to help developers build applications using ServiceNow's business context and platform capabilities. OpenAI models will also power ServiceNow's Build Agent and Autonomous Engineer for application development and modernization. Platform modernization — migrating aging custom apps off legacy stacks — is a credible fit for coding agents because the target behavior is already specified by existing code. No benchmarks, success rates or supported languages were disclosed.

Cybersecurity. A joint solution will pair what OpenAI calls its cybersecurity capabilities with ServiceNow's asset, identity and business context to help teams investigate vulnerabilities, prioritize risk and coordinate remediation. The asset-and-identity angle is the interesting part: vulnerability prioritization fails in practice less because scanners miss things than because teams cannot tell which exposed asset actually matters to the business. Whether a model improves that judgment, or simply produces confident rankings faster, is exactly what early access in January 2027 would need to demonstrate. The announcement does not describe evaluation methodology.

Voice. An OpenAI-powered experience in the Super Intelligence Control Tower will let someone describe an AI use case out loud, with the system populating assessment fields and asking follow-up questions to fill gaps. This is narrow — structured intake for AI governance reviews — and the narrowness is a point in its favor. Form-filling with clarifying questions is a task where current voice models are plausibly good enough.

Each company as the other's customer

ServiceNow plans to roll ChatGPT Work and Codex to all employees, not just engineers, and says it will use OpenAI capabilities in internal security work. OpenAI, in turn, says it is running ServiceNow workflows in its own operations, supported by the Super Intelligence Control Tower.

Mutual deployment is becoming standard choreography in these announcements — OpenAI published a similar item about Uber adopting ChatGPT Work for all employees on October 22. It is genuine evidence of commitment, and it also functions as marketing. Neither company disclosed seat counts, contract value, financial terms, exclusivity, or how the models are deployed and where data flows. For regulated buyers, those omissions are the first questions a procurement team will ask.

The framing to be skeptical about

ServiceNow CEO Bill McDermott's quote positions ServiceNow as the "Super Intelligence Control Tower" that orchestrates, secures and governs AI across the enterprise. OpenAI President Greg Brockman's framing is narrower and more defensible: models get more useful when connected to the systems a business already runs on.

The control-tower claim is a strategic bid, not a shipped fact. Every large platform vendor is making a version of it right now, and most enterprises will end up with several models, several agent frameworks and several systems of record. ServiceNow's advantage is that it already holds the workflow, asset and identity context that agents need to act responsibly. Its exposure is the same one Salesforce and others face: if the assistant becomes the primary interface, the vendor supplying the assistant accumulates leverage over the vendor supplying the workflow. The MCP plugin cuts both ways.

What to watch

• Whether the end-of-2026 "enhanced capabilities" for the MCP plugin include write actions with audit trails and approval gates, or remain primarily read-and-retrieve.
• Evidence from the January 2027 security early access: false-positive rates, how prioritization is evaluated, and whether remediation steps require human approval by default.
• Any disclosure of deployment model and data handling — where instance data goes, retention, and whether ServiceNow customers can restrict it.
• Whether ServiceNow's all-employee ChatGPT Work and Codex rollout produces measurable internal results, or stays an anecdote in press materials.
• How exclusive this is. Nothing in the announcement prevents ServiceNow from running comparable integrations with other frontier model providers, and nothing prevents OpenAI from doing the same with competing workflow platforms.