OpenAI Just Shipped Your Startup. As a Feature.

Read the feature list for OpenAI Presence and try to find a company it doesn't threaten. Agents wired into internal data and existing software. Policies and standard operating procedures. Guardrails covering what the agent can see and what it is allowed to do. Simulations against edge cases before launch. Automated graders that score whether the agent reached the right outcome, followed policy, used its tools properly and escalated when it should have. Human review on every change. Then, once it's live, Codex reads the production transcripts and proposes fixes.

Any one of those is a feature. Put together, they are the product page of a dozen funded startups, compressed into one announcement by the company that also sells those startups their models.

OpenAI announced Presence on July 22, describing it as a way to pair model reasoning with policies, guardrails and escalation rules. In OpenAI's own framing, the product brings together the pieces teams need to run agents in production: procedures, approved actions, simulations, evaluation tools and a Codex-powered improvement process. VentureBeat read it as an answer for companies that want agents but can't figure out how to stitch models, APIs, internal systems, security controls and evaluation tools into something reliable.

The pitch works because it names the real problem. Demos stopped being hard two years ago. What stays expensive is an agent that still follows policy in month nine, after the pricing page has changed twice and customers have learned new ways to argue with it.

The proof point is OpenAI's own phone line

The best evidence OpenAI has is that it runs Presence on its own English-language support number, 1-888-GPT-0090. According to AI News, the company says the agent met its internal benchmarks for frontline human support within weeks, now resolves 75 percent of inbound issues with no human involved, and shed 15 percentage points of human handoffs in ten days through the Codex improvement loop.

Those are good numbers, and AI News flagged the obvious caveat: they are OpenAI's figures, from OpenAI's channel, measured against grading criteria OpenAI wrote. Every vendor in this category computes containment in roughly that self-flattering way, which is why serious enterprise buyers stopped taking any of them at face value and started running their own bake-offs.

The 75 percent isn't the number to watch anyway. Fifteen points in ten days is a claim about rate of improvement rather than quality, and it's the one part of the announcement a competitor can't match by rewriting a marketing page.

The timing isn't a coincidence

The easy explanation is that OpenAI moved up the stack because the stack was sitting there. The better one is that the floor gave way underneath it.

Token prices have fallen by somewhere in the high-90s percent over three years. Pebblous, working from Stanford HAI and Epoch AI figures, puts the range at 94.5 to 99.7 percent depending on the estimate, with GPT-4-class capability going from $30 per million tokens in early 2023 to sub-dollar territory by this year. The same analysis finds open-weight models have closed to a low single-digit gap with the closed frontier, though it cautions that the gap has roughly stabilised rather than continuing to shrink.

Cheap and close enough turns out to be sufficient. Wavect reports that DeepSeek, Qwen, Kimi and GLM have closed most of the practical quality gap on coding and reasoning at prices commonly 15 to 30 times below the Western frontier, and recommends what most engineering teams have already built: route the easy majority to a cheap model and reserve the frontier for calls where a wrong answer costs real money.

Enterprises noticed, and it shows up in the P&L. Forbes columnist Peter Cohan calls the resulting squeeze "tokenomics," and his read is unsentimental: intense competition, buyers who now know how to shop, strong supplier power upstream, and a shrinking pool of profit for anyone whose revenue is denominated in tokens.

So the labs went where the contracts are. Not per-million-token pricing but per-deployment pricing, with a scoped job, a signed governance model and switching costs measured in quarters.

The tell is that you can't buy it online

Presence is not self-serve, and OpenAI says so plainly. AI News noted that deployments are run by OpenAI's own Forward Deployed Engineers alongside selected global systems integrators, with each engagement built around a single job: a billing dispute, an insurance claim, an employee IT ticket. The agent gets only the access that job requires, and the customer writes the rules for sign-off and handoff.

That is a services business wearing a product's clothes, which helps OpenAI and hurts it in equal measure. Staffing forward-deployed engineers against something is the clearest signal a company is serious. It's also a growth model that runs at the speed of hiring rather than the speed of a signup form, and the queue for those engineers is going to be long.

Which leaves the incumbents room. MLQ positioned the launch as a direct shot at Salesforce, ServiceNow, Zendesk and hundreds of CX startups building on OpenAI's own models, but those companies are not small and they sit closer to the system of record. CB Insights lists Sierra as having raised $1.585 billion across five rounds, with Wonderful, Parloa, Forethought, Fin and Decagon among its named rivals. A limited GA program does not clear that field.

There's a governance argument cutting against the platform owner, too. If inference has to run in a specific jurisdiction, or the compliance team wants model optionality on principle, a single-vendor stack is the wrong shape no matter how good the graders are. Most of these companies are multi-model by construction. Sacra notes that Decagon builds on OpenAI, Anthropic and Cohere at once, alongside its own fine-tuned models. A year ago that looked like hedging. Now it reads as positioning.

Worth noting on the way past: Contrary Research points out that Sierra co-founder Bret Taylor also chairs OpenAI's board, and argues that the relationship has been a distribution advantage rather than a liability.

Three things worth watching

Whether the Codex loop generalises. Continuous improvement driven by production escalations is the genuinely hard piece here, and OpenAI has shown it working on exactly one deployment: its own, in one language, on a workload it understands far better than any customer will understand theirs.

Whether limited GA ever opens. If Presence is still gated behind forward-deployed engineers a year from now, it's a top-50-logos business and everything below the enterprise tier stays contestable. If it becomes self-serve, the floor drops out of the category.

Whether the trust story survives contact. CX Today observed that Presence launched during the same week OpenAI and Hugging Face were investigating an incident in which advanced models exploited vulnerabilities to reach Hugging Face production infrastructure. Separately, explainX covered a Zenity disclosure of a ChatGPT flaw where a single crafted URL could create, authorise and publish an attacker-controlled workspace agent inheriting an employee's connectors; OpenAI removed the parameter after the report. Selling governance means making a promise about your own security posture, and procurement will price it accordingly.

The lesson generalises well past customer support. When the commodity underneath you deflates by 99 percent, your supplier becomes your competitor, because tokens no longer pay their bills and your contracts do. Any product that amounts to a thin layer of orchestration over someone else's model should assume it gets absorbed. The only open question is the quarter.