OpenAI Turned Its Own Call Center Into a Product. Presence Ships With Engineers Included and No Self-Serve Button.

TL;DR

OpenAI announced Presence on July 22: an enterprise platform for deploying AI agents that answer questions, touch company systems, take approved actions, and escalate to humans, across voice and chat. It bundles company context, policies, permissions, guardrails, pre-launch simulations with graders, and a Codex-powered loop that reads production sessions and proposes fixes. The proof point is OpenAI's own English-language phone support line, which it says now resolves about 75% of inbound calls with no human involved. The catch for builders: there is no self-serve tier. Deployments are led by OpenAI's forward-deployed engineers and select systems integrators, pricing is undisclosed, and the agent-ops layer a whole cohort of startups sells just became a first-party OpenAI product.

What Presence actually is

Presence is not a model. OpenAI describes it as a deployment platform: a shared foundation of company context, policies and standard operating procedures, permissions, guardrails, approved actions, and evaluations that sits under every agent a company runs, so the same rules follow the agent across voice, chat, and whatever channel comes next. Target workloads are the unglamorous, high-volume ones: customer support, sales development, procurement, IT, and HR.

The lifecycle is the interesting part. Before launch, teams run agents through simulations of common requests, edge cases, and higher-risk scenarios, with graders scoring whether the agent reached the right outcome, followed policy, used tools appropriately, and escalated when it should have. In production, guardrails can step in when an interaction drifts outside the company's boundaries, and agents only see the data and systems their specific workflow requires.

After launch, Codex reviews production sessions and escalations and proposes improvements, but staff have to test and approve every change before it goes live. Think of an airline whose maintenance crew reads every flight recorder overnight and writes up fixes, while nothing gets bolted onto the plane until an engineer signs the work order. That is the pitch in one sentence: the agent is cheap, the sign-off loop is the product.

The dogfood numbers

OpenAI's reference customer is OpenAI. The company says Presence has been running its own English-language phone support line, where it matched the benchmarks used to grade frontline human support quality within weeks and now resolves roughly 75% of inbound calls without human intervention. It also credits the Codex loop with cutting handoffs to humans by 15 percentage points in 10 days, which is the kind of curve you only get when the system fixing the agent reads every transcript.

Early external deployments exist but are cautious: BBVA is exploring AI voice support for everyday banking in Mexico, SoftBank Corp. is testing natural Japanese-language customer conversations, and Insurance Australia Group is looking at surge support during events like severe weather. Note the verbs: exploring, testing, looking at. Nobody outside OpenAI is publishing a 75% number yet.

The go-to-market is the real headline

Presence launched into a limited general availability program for eligible enterprises, and you cannot swipe a card for it. Deployments are scoped one at a time and led by OpenAI's Forward Deployed Engineers plus select global systems integrators, who wire up the company systems, permissions, and policies. Pricing is undisclosed, described only as coming as availability expands.

Sit with that for a second. The company that spent eight years teaching the industry that distribution means an API key and a credit card form now has a flagship product you can only acquire by talking to its account team and hosting its engineers. That is the Palantir playbook, not the Stripe playbook, and OpenAI is running it on purpose: enterprises keep paying for outcomes, not tokens, and outcomes need someone on-site who can be blamed.

What it means for builders

If you build agent platforms, the supplier just moved into your apartment. The evals, guardrails, simulation harnesses, and escalation logic that CX-agent startups like Sierra and Decagon, and platform plays like Salesforce's Agentforce, sell as their moat are now a first-party product from the company whose models most of them run on. OpenAI has been creeping up the stack all year, from the ChatGPT Work desktop push to Codex everywhere; Presence is the first time the whole operations layer is the SKU.

There is a carve-out, at least for now. Presence targets big-ticket enterprise workflows through a high-touch channel, and OpenAI physically cannot deploy FDEs to the mid-market. If you sell agent ops to companies too small for a forward-deployed engineer, the announcement validates your category more than it threatens you this quarter. The threat arrives whenever the playbooks those FDEs are writing get productized into something self-serve, which is historically what OpenAI does next.

The other takeaway is technical: OpenAI's internal bet is that agent quality comes from the loop, not the model. Simulation before launch, graders as the gate, transcripts feeding an automated reviewer, humans approving every change. None of that requires Presence; it is a described architecture you can build against your own stack today, and the 15-point handoff drop in 10 days is the argument for why you should.

The caveats

Every number here is OpenAI measuring OpenAI, on the friendliest deployment imaginable: the customer, the vendor, and the model provider are the same company, with unlimited engineering attention. The 75% figure covers one English-language phone line, and resolution rate says nothing about resolution quality from the caller's side. External customers are in exploratory pilots, not production at scale. There is no pricing, no self-serve path, no third-party evaluation, and limited GA means OpenAI decides who counts as eligible. File the claims under promising and unaudited.

Key Takeaways

• OpenAI launched Presence on July 22: an enterprise platform for voice and chat agents governed by company policies, permissions, guardrails, simulations, and graders.
• A Codex-powered loop reviews production sessions and proposes improvements, with mandatory human testing and approval before changes ship.
• OpenAI says its own English-language phone support line now resolves about 75% of inbound calls with no human, and the loop cut handoffs by 15 percentage points in 10 days.
• There is no self-serve: deployments are individually scoped, led by OpenAI Forward Deployed Engineers and select systems integrators, with pricing undisclosed.
• Early external users BBVA, SoftBank Corp., and Insurance Australia Group are in exploratory pilots, not published production numbers.
• The agent-ops layer that startups sell as a moat is now a first-party OpenAI product; the mid-market is safe only until the FDE playbooks get productized.

Sources: OpenAI (Introducing OpenAI Presence), No Jitter, Help Net Security, SiliconANGLE, VentureBeat