explainx / blog
July 22, 2026: OpenAI Presence ships limited GA for voice/chat agents — policies, guardrails, Codex improvement loops, FDE-led deploy. Not self-serve.
Jul 24, 2026
A single crafted ChatGPT URL could silently create, authorize, publish, and schedule an attacker-controlled Workspace Agent that inherited the employee’s connectors. OpenAI removed the vulnerable parameter after Zenity’s report.
May 28, 2026
OpenAI's Secure MCP Tunnel enables enterprises to connect private MCP servers to ChatGPT, Codex, and the Responses API without opening inbound firewall ports or exposing internal systems to the public internet. Learn architecture, deployment patterns, security best practices, and enterprise integration strategies.
Jul 22, 2026
Free-tier ChatGPT now carries ads, clearly labeled and separate from answers, targeted using conversational context rather than just keywords. Paid subscribers are exempt for now. Here's how the ad platform works and why it reopened Sam Altman's old "ads are a last resort" quote.
On July 22, 2026, OpenAI shipped Presence — not another model card, but a deployed enterprise product for putting agents on real customer and internal workflows with policies, guardrails, and humans in the escalation path.
The pitch matches what every enterprise already learned the hard way: demos are cheap; reliable production agents that still obey policy when products and callers change are expensive. Presence packages that stack and sells it with Forward Deployed Engineers attached.
| Question | Answer |
|---|---|
| What shipped? | OpenAI Presence — enterprise voice + chat agents |
| When? | Jul 22, 2026 · limited GA |
| Self-serve? | No — FDE / SI-led only |
| Channels today? | Voice + chat (support, sales, internal IT-class flows) |
| Core stack? | Policies · SOPs · guardrails · approved actions · sims · evals · Codex improvement |
| OpenAI dogfood? | 1-888-GPT-0090 · ~75% resolve without human · −15 pp handoffs in 10 days |
| Design partners? | BBVA · SoftBank · IAG |
| Voice API still? | Yes — frontier models via API remain for voice customers |
OpenAI’s framing is blunt: the bottleneck is no longer “can agents work?” — it is keeping them reliable as policies, SKUs, and caller behavior drift.
Presence starts each deployment with one job (billing resolution, insurance claims, employee IT tickets). The agent gets only the knowledge and system access that job needs. The company sets:
That is agent harness + governance, not a chat widget. Same pressure curve as enterprise AI control-plane thinking and benchmark-before-you-buy guidance — outcomes and evals beat model brand names.
| Component | Role |
|---|---|
| Policies / SOPs | Shared behavior across deployments |
| Guardrails | Intervene when a session leaves company boundaries |
| Approved actions | Scoped tool use against company systems |
| Simulations + graders | Pre-launch checks: outcome, policy, tools, escalation |
| Eval tooling | Measure accuracy and performance over time |
| Codex improvement loop | Propose post-launch updates from production signals |
Companies decide what stays consistent across workflows (policies, evals, escalation rules) and what varies by channel or use case — so the second agent is not a greenfield rewrite.
At launch Presence supports real-time voice and chat: support, outbound sales, high-risk internal flows.
Example path OpenAI describes for billing: understand request → verify customer → look up account → apply policy → take an approved action.
If you already build duplex voice on OpenAI’s stack, keep the GPT Live / Realtime and Realtime-2 API guides in view — Presence sits above model APIs as a managed product, and OpenAI says API voice access continues for existing customers.
| Signal | Detail |
|---|---|
| OpenAI phone support | English channel 1-888-GPT-0090 on Presence |
| Resolve rate | ~75% inbound issues without human assistance |
| Codex loop | −15 percentage points human handoffs in 10 days |
| BBVA | Exploring AI voice support for everyday banking in Mexico |
| SoftBank | Testing natural Japanese-language customer conversations |
| IAG | Exploring support during high-demand events (e.g. severe weather) |
Treat vendor dogfood as directional. Your contact-center baseline, language mix, and compliance envelope will differ.
Before users: simulate common requests, edge cases, and higher-risk scenarios. Graders check outcome, policy adherence, tool use, and escalation.
After launch: production sessions, escalations, and quality signals show where the agent works and where it fails as products and behavior change. Codex (Presence plugin) investigates and suggests updates; teams test against the live version, then approve a controlled rollout.
When a use case exceeds the product surface, OpenAI FDEs and partners extend it — which is why Presence reads as much like a deployment company motion as a SaaS SKU. That matches the FDE demand spike and interview prep story we have been tracking all year.
| Topic | Official stance / buyer question |
|---|---|
| Availability | Limited GA · eligible enterprises · account team |
| Delivery | OpenAI FDEs + select global SIs |
| Self-serve | Not yet |
| Pricing | Not published; scoped per use case (broader pricing “to come”) |
| Model pin | Configuration can move with the workflow — ask contract language |
| vs Workspace Agents | Presence = managed production; Workspace = self-serve builder (AgentForger lesson still applies to any agent builder) |
Questions to put in the RFP:
| You | Action |
|---|---|
| Contact-center / CX lead | Ask account team for Presence scoping vs incumbent IVR/bots |
| CIO / AI platform | Map Presence vs build-on-Realtime/API vs Workspace Agents |
| Security / compliance | Demand sim coverage for high-risk actions + audit of Codex change proposals |
| FDE / SI partner | Presence is the product shape of the FDE economy — skill up on evals + policy packs |
| Builder staying self-serve | Keep shipping on API / Workspace; watch for future self-serve Presence |
Sources: Introducing OpenAI Presence · OpenAI product announcement (Jul 22, 2026)
Product scope, partner status, and resolve-rate figures reflect OpenAI’s July 22, 2026 announcement. Limited GA terms, pricing, and supported workflows change — confirm with your OpenAI account team before procurement.
