OpenAI's Presence launch is the tell: the AI labs are becoming deployment companies, not just model makers

What launched — and why it’s a signal, not just a product

On July 22, 2026, OpenAI introduced Presence, a managed platform for deploying and governing production AI agents across voice and chat. Per OpenAI and VentureBeat, it targets customer support, outbound sales, and internal IT, and the emphasis is entirely on control:

• Task-scoped deployments — each agent is limited to a specific job.
• Company-defined policies, guardrails, and escalation rules — what the agent can do, when it needs approval, when a human takes over.
• Simulation-based testing — run agents against common requests, edge cases, and higher-risk scenarios before production.
• Codex-powered improvement — OpenAI’s coding agent analyses signals and proposes updates that human teams approve.

It’s in limited general availability for eligible enterprise customers. Taken alone, Presence is a customer-experience product for large orgs — narrow, and not something most readers will buy. But it’s the clearest single data point in a shift that affects everyone who buys AI, so it’s worth reading for the signal rather than the spec sheet.

The shift: from selling models to selling deployment

Step back and the pattern is unmistakable. Two moves are happening at once, across every major lab.

Move one — platformizing agent deployment. The labs are building managed layers that sit above the model and handle the hard part: getting an agent into production and keeping it safe.

• OpenAI Presence — deploy and govern production voice/chat agents.
• Google’s Gemini Enterprise Agent Platform — a unified platform for building, orchestrating, and governing agents, explicitly positioned as the evolution of Vertex AI (future roadmap ships through the agent platform).
• AWS Bedrock AgentCore — multi-model agent orchestration governed by AWS’s own IAM controls.

Move two — standing up implementation-services arms. The labs are also building (or backing) consulting-style organisations that do the work of adopting AI for you.

• Ode with Anthropic — a $1.5B AI-implementation firm backed by Anthropic, Blackstone, and Hellman & Friedman, on the explicit thesis that “the next trillion-dollar AI business is implementation, not just models.”
• OpenAI’s Deployment Company — a $4B services arm (with the Tomoro acquisition) built to close the enterprise deployment gap.

Both moves point the same direction: up the stack, away from the raw model.

Why this matters

1. It’s what commoditization looks like from the vendor side. When GPT-5.6 ships a $1/$6 Luna tier, Grok 4.5 undercuts on cost-per-task, and open-weight models close the gap on specific tasks, raw model access stops being a defensible business. The labs know this. Presence, Gemini Enterprise, and the services arms are the answer to “if the model is a commodity, what do we sell?” The answer is: the layer that turns a model into a working, governed, improvable system — because that’s the part nobody has solved, and the part customers will pay a premium for.

2. The value — and the lock-in — is moving up a layer. This is the buyer-relevant core. An API key is portable: you can swap models in an afternoon. A deployment platform is not. Once your policies, guardrails, evaluation suites, escalation rules, and workflows live inside Presence or Gemini Enterprise, switching vendors means rebuilding all of that. The labs are, quite deliberately, moving the lock-in from the model (easy to leave) to the deployment layer (hard to leave). That’s smart strategy and a real risk for buyers to price in.

3. “Which model is best” is becoming the wrong question. For most of the last three years, buying AI meant picking a model. Increasingly it means picking a platform — and the model becomes an interchangeable component inside it. That reframes tool selection: the question shifts from “is Claude or GPT-5.6 better” to “whose deployment, governance, and testing tooling fits how I actually ship.” Model quality still matters for hard tasks, but it’s becoming table stakes rather than the differentiator.

4. The governance framing is doing quiet strategic work. Notice how much of Presence is about control — policies, approvals, escalation, testing. That’s partly genuine (deploying agents that take real actions is scary), and partly a moat: governance is sticky, auditable, and exactly what regulated enterprises need. It also dovetails with the government-gated release regime — a world where AI is expected to be reviewed, gated, and controlled is a world where a governance platform is a requirement, not a nicety. The labs are building for the regulatory environment they helped create.

5. It changes what a small buyer should watch, even if they’ll never buy Presence. You may run a two-person business and never touch an enterprise agent platform. But the trend still reaches you: the consumer-grade versions of this — ChatGPT Work, Claude Cowork — are the same “deployment layer” logic aimed downmarket. The features, the lock-in, and the pricing structure that get built for enterprises tend to arrive in the prosumer tiers a few quarters later. Watching the enterprise layer tells you what your subscription is about to become.

What this means for you

• Reframe your evaluation. Alongside “which model,” start asking “whose deployment and governance layer am I standing on, and how hard is it to leave.” For agents especially, that’s now the bigger question.
• Keep the portable things portable. Store your prompts, evaluation sets, and data in forms you own, not locked inside one vendor’s platform. That’s the single best hedge against the lock-in this trend is building.
• If you’re deploying agents at any scale: the managed platforms genuinely reduce production risk (testing, guardrails, escalation are hard to build yourself). Buy the convenience deliberately, and negotiate exit terms as if you’ll use them.
• If you’re a small buyer: you don’t need any of this yet — but expect the same deployment-layer features and pricing to show up in ChatGPT Work / Cowork-style products. Compare those in the best AI agents guide.

The honest caveats

• Presence is a real but narrow product. This piece uses it as a signal; on its own it’s a limited-GA enterprise CX tool, not a mass-market launch. Don’t over-read the single product — read the pattern.
• “The labs are becoming deployment companies” is a thesis, not a fact. It’s a well-supported reading of several confirmed moves (Presence, Gemini Enterprise, AgentCore, Ode, the Deployment Company), but the labs remain, first and foremost, model makers. This is a shift in emphasis, not an abandonment of models.
• Lock-in is a risk, not a verdict. Deployment platforms create switching costs, but they also deliver real value; “lock-in” isn’t automatically bad if the platform is genuinely better. The point is to choose it knowingly and keep exits open, not to avoid it reflexively.
• The competitive map is still forming. AWS, Google, OpenAI, Anthropic (via Ode), and others are all positioning; who wins the deployment layer is undecided, and vendor claims about governance and portability should be tested, not trusted.
• Model quality hasn’t stopped mattering. For frontier-hard work, the underlying model is still decisive. The “model is a commodity” framing is true at the good-enough middle, less true at the top.

The grounded summary: Presence is a small launch with a large implication. The frontier labs have read the same tea leaves everyone else has — models are commoditizing — and they’re moving the value, and the lock-in, up to the deployment and services layer. For buyers, the smart response is to see the shift early: keep evaluating models, but start choosing deployment layers as deliberately as you once chose models, and keep the keys to your own house.

Questions or corrections? Email Pick Right. Want the full list? See all news.