OpenAI Presence sells trusted agents. Here is the buyer’s checklist

Article Brief

Key Takeaways

4 Points24s Read

1. 01The product–Presence packages agent governance — policies, approved actions, intervening guardrails, graded simulations, human sign-off on changes — as a managed platform for voice and chat.
2. 02The delivery–Limited GA only: OpenAI Forward Deployed Engineers and select integrators lead deployments; no self-service, and pricing is undisclosed.
3. 03The claim–OpenAI runs its own support line on Presence and reports 75% autonomous resolution — real evidence, but from the friendliest possible customer.
4. 04The move–Treat the pilot as procurement: demand a guardrail intervention on your scenario, your own simulated resolution rate, and the undisclosed terms on paper.

OpenAI is now answering its own support phone with the product it wants to sell you. Presence, the enterprise agent platform the company introduced on July 22, runs OpenAI’s English-language phone line and — by the company’s own count — resolves 75% of inbound issues without a human touching them. That number is the pitch. The product underneath is something less glamorous and more important: a governance machine that decides what an AI agent is allowed to do, proves it in simulation, and puts a human signature between every change and production.

For teams that have spent 2026 assembling agent controls out of frameworks and toolkits, the interesting question is not whether Presence works. It is whether OpenAI’s version of the control checklist matches the one the industry has already converged on — and what it costs to find out.

What Presence actually is

Strip the launch language and Presence is a management layer for voice and chat agents in production. According to Help Net Security’s report on the launch, teams define policies and standard operating procedures, enumerate approved actions the agent may take, and connect company systems. Guardrails “can intervene when an interaction moves outside the company’s boundaries” — the platform’s phrase for stopping an agent mid-flight rather than auditing it after the damage.

The pre-production story is the stronger half. Before a deployment reaches users, teams run it against common requests, edge cases and higher-risk scenarios; simulations and graders score whether the agent “reaches the correct outcome, follows policy, uses tools appropriately, and escalates conversations when necessary.” When the system proposes improvements to itself, staff members test and approve the changes before they go live.

Anyone who has read TECHi’s coverage of CrewAI’s execution hooks or the AWS OpenSearch agent toolkit will recognize every item on that list. Policy boundaries, allow-listed actions, pre-production simulation, human sign-off, escalation paths: this is the control checklist the agent-tooling ecosystem spent the year standardizing. Presence’s real claim is not novelty. It is that the checklist now comes assembled.

Assembly is worth something, and it is fair to say so plainly. The joints between policy engine, simulator and telemetry are where self-built stacks leak engineering hours, and a vendor owning those seams end to end is the concrete thing being purchased here — provided the rest of the terms survive the scrutiny that follows.

The delivery model is the tell

How Presence ships says as much as what it does. Per VentureBeat’s reporting, the platform is available through a limited general availability program in which OpenAI’s Forward Deployed Engineers and select global systems integrators lead deployments. There is no self-service tier. Pricing, geographic limits and contractual terms are undisclosed — VentureBeat says it asked OpenAI about pricing twice without an answer.

Read that as a positioning statement. A product that requires vendor engineers to install is not a tool; it is a services engagement with software inside. That model has real advantages for the buyer — someone else owns the integration risk, and the vendor’s own people encode the policies — and one large disadvantage: the costs that matter, from engineering time to contract terms, are discovered inside the sales process rather than compared outside it.

It also changes who the product competes with. Self-assembled stacks built on open frameworks compete on flexibility and transparency. FDE-led deployments compete with consultancies and contact-center incumbents on accountability. Those are different purchases, evaluated by different committees, and pretending otherwise is how procurement teams end up comparing a platform quote against a framework’s free tier and learning nothing.

Claims versus evidence

The only rule that survives launch weeks: separate what is demonstrated from what is promoted. The demonstrated part is that OpenAI trusts Presence with its own support line — a real production workload with real callers, identity verification, account context and approved actions. Running your own traffic on your own product is meaningful evidence.

The promoted part is the numbers. The 75% autonomous-resolution figure and a claim that a Codex-powered improvement loop cut human handoffs by 15 percentage points over ten days are company-reported, measured on the friendliest possible customer: OpenAI itself, operating the platform with the people who built it. Nothing about that makes the numbers false. Everything about it makes them non-transferable. A support line staffed by the vendor’s own engineers, answering questions about the vendor’s own products, is the ceiling — not the baseline a retailer or a bank should budget against.

The honest translation for a buyer: treat 75% as an existence proof that the architecture can carry serious volume, and treat your own simulation phase — the one Presence itself insists on — as the only number that belongs in your business case.

The build-or-buy math, one year in

The assembled-stack alternative is not hypothetical; it is what most serious agent teams currently run. A framework supplies execution hooks, an infrastructure toolkit supplies connectors and controls, and an internal platform team owns the seams. The virtue of that path is that every control is inspectable and every component replaceable. Its cost is that the seams are yours: when the simulation harness disagrees with the guardrail layer about what “policy” means, the debugging belongs to your engineers at your expense.

Presence inverts the trade. The seams disappear into one vendor’s product, the accountability consolidates into one contract, and the integration burden shifts to people who deploy the same system weekly. What you give up is inspection depth and the option to swap parts — and, until pricing is public, the ability to know in advance what any of it costs.

One year into the agent-controls era, the honest comparison is not “platform versus framework” but “whose engineers debug the seams, and at what hourly rate.” Teams with a platform group that already ships production agents will find Presence’s pitch weakest, because they have paid down the integration cost it monetizes. Teams without one should notice that an FDE-led deployment is, functionally, renting that platform group — which is exactly why the undisclosed terms matter so much.

The question the governance layer cannot answer alone

One more context worth naming, because agent deployments do not live outside the law. Voice and chat agents that interact with the public are exactly the category the EU’s August 2026 transparency obligations reach, as TECHi detailed in its analysis of the AI Act’s enforcement gap. A platform can enforce escalation policy; it cannot decide for you whether a caller in Frankfurt must be told they are talking to a machine. Governance tooling and legal compliance overlap, but the second one stays on the buyer’s desk.

That is not a criticism of Presence — no platform can sign your compliance posture. It is a reminder that “trusted agents” is a stack, and the vendor sells only the middle of it.

What to ask before the pilot

The productive skepticism is specific. Five questions extract most of the signal from an FDE-led sales process:

• Show the intervention, not the demo. Ask to see a guardrail stop an in-flight action that policy forbids — the mechanism Help Net Security describes — on a scenario your team writes, not one the vendor brings.
• Price the exit before the entry. Policies, SOPs and simulation suites built inside Presence are configuration in someone else’s system. Ask what exports when the contract ends, in what format, and what a migration to a self-assembled stack would preserve.
• Demand your own 75%. Insist the simulation phase produce a resolution-rate estimate on your call mix, and get the grading rubric in writing. A vendor confident in the number will not resist the request; the platform was built to generate it.
• Trace the approval loop. Staff approve proposed changes before they go live — whose staff, with what review window, and what happens in the gap between a discovered failure and an approved fix?
• Get the undisclosed list disclosed. Pricing, geographic availability and contract terms are unpublished. None of those are unknowable; they are unnegotiated. The earlier they are on paper, the cheaper the pilot.

The bottom line

Presence is the strongest signal yet that the agent market’s center of gravity has moved from capability to accountability. The controls it productizes — approved actions, intervening guardrails, graded simulations, human sign-off — are the right controls, the same ones the open tooling has been racing toward all year. OpenAI operating its own support line on the platform is real evidence the assembly holds under load.

What remains unproven is everything a buyer actually budgets: transferability of the headline numbers, total cost of an FDE-led deployment, and the exit price of building your governance inside a vendor’s walls. Those answers exist — inside a sales process that currently publishes none of them. The checklist for extracting them is above, and unlike the resolution rate, it transfers to any vendor that walks in the door claiming their agents can be trusted.

The launch also sets a marker the rest of the market now has to answer. When the company with the most-used models decides the sellable unit is governed deployment rather than raw capability, every rival pitch deck gets rewritten around the same question. Buyers just inherited the leverage that comes with that — provided they walk in with the questions written down.