Follow a double-charge call to see how Presence works

The product is general: deploy Agents by role, connect systems, enforce policy, finish work, escalate when needed. To make that concrete, we walk one sample case from the product UI. Role: support. Channels: phone and text. Fictional retailer Swiftcart; the customer says a subscription was charged twice. From understanding the ask through refund, step by step—what Presence relies on in a real session.

Sample UI: left is a phone call (waveform + turn transcript)—the customer says they were charged twice; after the Agent checks the account it enters refund flow. Right is text chat—the customer asks about order status; the Agent calls LOOKUP ORDER STATUS and surfaces a located invoice card. Same capability stack for voice and text.

What Presence actually ships

Pull that refund call apart and the parts are all there. The same kit can fit other roles—billing support, insurance claims, employee IT—with shared policy templates and evaluation methods, each role with its own permissions and allowed actions. The case is support; the product is not limited to support.

When refund policy changes, simulate first, then let Codex propose

Stay with the support case. Suppose the company updates its annual refund rules—how double subscriptions are handled, whether verbal abuse counts as a crisis ticket. Presence does not pour those changes straight into the Agent already taking calls.

Simulation comes first: push a batch of fake but realistic requests. A grader scores each group—right outcome, policy compliance, correct tools, correct escalation.

Sample: after policy becomes "new annual refund rules," simulations run across guardrails, refunds, cancellations, email and OTP verification, and related groups. Groups in the figure score 80% each; the batch header shows Pass. How the pass line is set, and whether it is contractual, is not public—read it as marketing UI.

After the Agent is taking real calls, the system keeps watching: session quality, handoff rate, what customers ask. Where it starts to strain, the coding assistant Codex—with a Presence plugin—investigates and writes edit proposals. The team tests the candidate against the running production version; a human hits approve before it replaces live.

Sample production dashboard: delivery quality score, inbound volume, customer intent mix, response latency, per-task performance. Used to spot signals like "refund tickets are escalating more." How metrics are calculated, and whether they sit in a service contract, is not public.

OpenAI's own phone line, and three enterprises trying it

Among production instances of Presence, OpenAI used itself first. The English phone support line 1-888-GPT-0090 already runs on Presence: open-ended requests, identity checks, account context, approved actions. OpenAI reports roughly 75% of inbound calls resolved without a human, and that the Codex improvement loop cut handoffs another 15 percentage points in 10 days. That is the same product installed on the "own support" role—the product itself is not limited to support.

Only BBVA is explicitly a design partner; SoftBank and IAG are testing and exploring. None of this is "we already replaced the full contact center."

What Presence is, and what it's for

If you only take one sentence, take this:

What it is not: not another web chatbot that only talks, and not a ChatGPT membership feature for individuals. Individuals cannot open it; companies cannot self-serve from a web signup either. It is also not a single-purpose "voice support only" tool. Voice and text are the dialogue channels open today; the support phone case is the most complete public walkthrough—billing, claims, employee IT, and more can each get their own deployment.

Why it matters: many companies already proved "AI can talk to people." Presence targets the next step: once the assistant is in real work, who sets permissions, who writes rules, was it drilled before launch, who fixes production failures, and will a fix break live traffic. OpenAI packages the policy, simulation, dashboards, and change-approval work companies used to stitch themselves—and sends engineers on-site to install it into real workflows.

What you can use it for: for a business or ops lead, think of it as the "badge + permission sheet + job trial + QA edit flow" for an AI assistant on a given role. The role can face customers or staff:

The full "subscription charged twice" walkthrough was support demonstrating those four blocks. By now you should have the whole picture: Presence covers enterprise AI Agents from install through ongoing rule changes, with the emphasis on getting work done under control. Voice and chat are today's dialogue front doors; the support case is one teaching path, not the full product boundary.

Who can use it, and what is still unknown

Presence is limited access for now: enterprise customers talk to OpenAI; forward deployed engineers (FDEs) and a small set of systems integrators install it—pick workflows, wire systems, set permissions, test, then go to production. No self-serve web signup. Eligibility gates and commercial terms are not on a public checklist.

Customers already building voice on the OpenAI API keep that path. Presence is a separate "product + on-site" track: policy, simulation, evaluation, approval, and post-launch change live in one surface.

Around the same time OpenAI also disclosed a security incident in which a model in an evaluation environment escaped to touch external systems. Presence stresses policy, simulation, and human approval for controlled go-live; do not read that as having fixed that incident. This site has a separate piece on it.

Each deployment does one job type: billing disputes, insurance claims, employee IT tickets. The assistant only gets that role's knowledge and interfaces; the company hard-codes what it may do, when a human must approve, and when it must escalate. Before go-live, batch-run simulated tickets; after go-live, Codex (the coding assistant) only proposes edits—new versions must be compared to production and approved by a human.

We install a "subscription charged twice" support role for fictional retailer Swiftcart and chain pre- and post-launch into one loop: lock permissions and policy first, pass simulation before real calls; when production signals fire, Codex with a Presence plugin writes proposals; the team compares to production; a human approves before swap. The Agent cannot overwrite itself.

OpenAI first ran its English phone support line 1-888-GPT-0090 on Presence. Externally, design partner BBVA is exploring banking voice support in Mexico; SoftBank is testing Japanese customer dialogue; IAG is exploring peak support such as extreme weather—all exploring or testing, not full contact-center replacement.