OpenAI Is Selling Its Most Ambitious AI Agents Like Management Consulting, Not Software
OpenAI Presence Review 2026: a governance-first platform for voice and chat agents, deployed entirely through OpenAI's own consultants — no pricing page, no signup form, and a 75% resolution claim that's currently self-reported.
Most AI agent products released in 2026 lead with what the model can do. OpenAI Presence, launched July 22, leads with something else entirely: what the agent is not allowed to do, who reviewed it before it went live, and what happens when it gets something wrong. The product is essentially a packaging of policies, guardrails, pre-launch simulations, evaluation graders, and a Codex-powered improvement loop around OpenAI's models — governance as the headline feature, not an afterthought bolted onto a chatbot.
It's also, by a wide margin, the least accessible thing OpenAI has shipped this year. There's no pricing page. There's no signup form. There's no public API. Getting access means contacting an OpenAI account team and going through a scoping process led by Forward Deployed Engineers — consultants, not a checkout flow. That's a deliberate choice, not an oversight, and it's the thing this review spends the most time on, because it changes what "reviewing" Presence even means.
What Presence Actually Is
Policies, SOPs, and guardrails
Every deployment starts from a company's own rules — what the agent can say, which systems it can touch, and where it must stop and escalate. Guardrails intervene when a conversation drifts outside those boundaries, and access to internal systems and databases is scoped rather than open-ended.
Approved actions with escalation
Agents can process refunds, update accounts, or resolve billing issues within a defined set of authorized actions, with approval gates on sensitive operations and a clear handoff to a human when policy, risk, or judgment calls for one.
Simulations and evaluation graders
Before anything goes live, the agent is tested against common requests, known edge cases, and higher-risk scenarios. Graders then assess whether it reached the right outcome, followed policy, used tools correctly, and escalated appropriately — a formal QA layer most consumer-facing chatbots skip entirely.
Codex-powered improvement loop
After launch, Codex reviews production sessions and escalations, proposes behavioral updates based on what it observes, and a human team tests and approves each change before it rolls out — a continuous-improvement cycle rather than a static deployment that goes stale.
Presence operates across both voice and chat. A secondary report states the voice side runs on GPT-Live-1, OpenAI's full-duplex voice model that reached the API on September 10, 2026 at $0.05/minute — though OpenAI's own Presence documentation doesn't commit to a fixed model, saying only that "model configuration is selected for the workflow." Named early evaluations include BBVA (Mexico, everyday banking support), SoftBank (Japan, natural Japanese-language conversations), and IAG (Australia, support during severe-weather demand spikes) — though as of this review, only one deployment is confirmed live in production: OpenAI's own phone support line, 1-888-GPT-0090.
Pricing — There Isn't Any
OpenAI has published no price, rate, or fee structure for Presence anywhere. The Help Center states plainly that "pricing and implementation scope are specific to each customer and deployment. No specific prices or pricing formulas are provided." A company spokesperson told The Register that deployments during this limited-availability phase are "scoped individually based on each customer's use case and implementation needs," with broader pricing details promised as availability expands — no timeline given.
It's worth being careful not to confuse this with OpenAI's published API rates. GPT-Live-1 costs $0.05/minute in the standalone API, and GPT-5.6 Sol runs $4/$20 per million input/output tokens — but those are rates for building your own agent on OpenAI's infrastructure, not what Presence itself costs. Presence's actual price includes implementation services, integration engineering, and ongoing managed deployment, none of which map cleanly onto a per-token or per-minute number. There is no public estimate of what any of that costs in practice.
The Consulting Model
This is the part of the story that explains everything else about the product. OpenAI acquired the consultancy Tomoro in May 2026, specifically to build out the Forward Deployed Engineer workforce that now leads Presence deployments — a Palantir-style delivery model where the company sells implementation capacity, not just model access. That's a meaningful strategic shift for a business that's mostly been understood as a model provider: OpenAI is positioning itself, at least for its highest-value enterprise customers, as a deployment partner as much as a technology vendor.
The commercial logic isn't hard to follow. As frontier models become more commoditized and price-competitive, the margin increasingly sits in the "plumbing" — the integration work, the governance layer, the ongoing tuning — rather than in the tokens themselves. Selling that plumbing as a managed service, at whatever rate a Forward Deployed Engineering team can negotiate per enterprise contract, is a different business than selling API access at a published rate.
Pros & Cons
✓ Strengths
- ✅ A genuinely differentiated governance-first architecture — simulations, graders, and approval gates as core product, not an add-on
- ✅ The Codex improvement loop is architecturally sound: production sessions inform proposed changes, which get tested and approved before rollout
- ✅ OpenAI is dogfooding the product on its own support line, giving at least one real, checkable reference point
- ✅ Each deployment goes through security, privacy, and legal review before production — a real process, not a marketing claim
✗ Weaknesses
- ❌ No public pricing anywhere, at any stage of evaluation
- ❌ No self-serve access — every deployment requires an OpenAI account team relationship and FDE-led scoping
- ❌ The 75% resolution figure and the 15-point handoff-reduction figure are both self-reported, with no independent verification published
- ❌ Named design partners are still in evaluation or trial, not confirmed production use — only OpenAI's own line is a live reference
Does It Actually Work? The 75% Claim
OpenAI states that Presence resolves 75% of inbound issues on its own phone support line without human assistance, and that the Codex improvement loop cut human handoffs by 15 percentage points in ten days. Both numbers come from OpenAI's own deployment, measured by OpenAI, with no third-party audit published alongside them.
Analyst Pareekh Jain's caution on this point is worth repeating rather than glossing over: the 75% figure "should be viewed as evidence that the technology can work, not as a benchmark that every enterprise should expect." Large enterprises carrying fragmented legacy systems, uneven knowledge bases, and heavier compliance requirements than OpenAI's own support operation may see meaningfully lower resolution rates — Jain specifically noted that integration and governance, not token costs, tend to be the real bottleneck in enterprise AI deployments. None of that makes the number meaningless. It makes it a best-case reference point from a company with unusually clean internal data and total control over its own workflow, not a promise transferable to any given buyer's environment.
Presence vs Genesys, PolyAI, ElevenLabs, Vapi
| Platform | Access model | Pricing | Best for |
|---|---|---|---|
| OpenAI Presence | FDE-led consulting deployment | Custom, unpublished | Large enterprises wanting governance-first, OpenAI-managed deployment |
| Genesys / NiCE / Five9 | Established contact-center platforms | Enterprise, generally quote-based | Organizations already invested in a mature contact-center stack |
| PolyAI / Parloa / Sierra | Self-serve or lighter-touch onboarding | Enterprise, estimable | Enterprise voice agents with more accessible onboarding than Presence |
| ElevenLabs / Vapi | Developer-first, self-serve API | Transparent per-minute rates | Teams that want to build their own voice agent rather than buy a managed one |
An independent comparison by o-mega.ai scored Presence 6.4 out of 10, ranking it 13th of 14 voice AI platforms evaluated — high on governance (9/10) but low on cost transparency (5/10) and control, since a customer doesn't own the code or the model relationship the way a developer-first platform allows. Omdia analyst Lian Jye Su made a related point worth sitting with: vendors like Genesys, NiCE, Five9, and AWS have sold similar customer-support automation for years, which means Presence's real competitive claim isn't that automated support is new — it's that governance-first, OpenAI-managed deployment is worth paying a consulting premium for over already-mature alternatives.
Who It's For
Consider Presence if: you're a large enterprise with a well-documented, high-volume workflow — billing support, claims processing, IT service requests — and you have both the budget for a consulting-style engagement and an existing OpenAI account relationship to start the conversation through.
Look elsewhere if: you want to evaluate or compare pricing before committing time to a sales process, you need something deployable in weeks rather than months, or you're a smaller organization without the internal resources to support an FDE-led implementation — a self-serve voice agent platform will get you further, faster.
Expert Editorial Opinion
The most interesting thing about Presence isn't the technology — it's the business model wrapped around it. OpenAI didn't build a self-serve product and add enterprise features later; it built a consulting practice with an AI product inside it, which is a meaningfully different bet than anything the company has shipped before. Whether that's the right call depends entirely on whether enterprises are willing to buy AI agents the way they buy management consulting — a slower, more expensive, more relationship-driven process than downloading an API key.
The skepticism visible in public forum discussion after launch is worth taking seriously precisely because it isn't really about bugs or reliability — there's too little external deployment yet for that kind of complaint to even surface. It's structural. Commenters questioned whether OpenAI's "battle-tested" framing squares with a product that still needs extensive guardrails to handle high-value work, raised real concerns about data handling on voice calls with no visible terms of service, and pushed back specifically on the idea of letting Codex make production code or account changes with a human just "mashing approve." One Reddit thread captured the tension well: is Presence a genuine threat to existing customer-support SaaS platforms, or mostly a bet that roughly 80% of the real work in any enterprise deployment — unifying fragmented data, policies, and systems behind one coherent layer — is exactly the unglamorous integration work a token-selling API was never going to solve on its own?
The 75% resolution figure is a real, if narrow, proof point — OpenAI is at least willing to run this on its own support line rather than only in a slide deck. But a single, highly controlled internal deployment doesn't settle whether the governance framework holds up across the messier data and compliance environments most enterprise buyers actually have. Until independent verification exists and a broader set of named customers move from evaluation to confirmed production, Presence is best read as a serious, well-architected bet rather than a proven platform — one that's asking enterprises to commit to a sales conversation before they can see a price, a demo, or an outcome that isn't OpenAI grading its own homework.
Final Verdict
Presence's governance architecture — simulations, graders, approval gates, and a Codex-powered improvement loop — is a genuinely differentiated approach to enterprise AI agents, and OpenAI running it on its own support line is a real, checkable proof point rather than pure marketing. The score reflects what's still missing rather than a flaw in the design: no public pricing at any stage, no self-serve access, unverified performance claims, and named design partners still in evaluation rather than confirmed production. For most organizations evaluating enterprise AI agents in 2026, that combination makes Presence a serious option to watch rather than a practical one to act on today.
❓ Frequently Asked Questions
🔗 Related ToolRadar Reviews
Official announcement and documentation: openai.com and the OpenAI Help Center. Given the complete absence of public pricing and the limited number of confirmed deployments, contact an OpenAI account team directly for current eligibility and cost rather than relying on any estimate.
Comments
Post a Comment