By Rajesh Beri | July 24, 2026
On July 22, OpenAI launched Presence — and with it, quietly declared that selling API access is no longer enough to win the enterprise.
Presence is not another model release. It is not another API endpoint. It is a fully deployed product for running AI agents inside enterprise customer service, sales, and internal operations — with OpenAI's own Forward Deployed Engineers sitting inside the customer's organization to make it work. Voice and chat. Policies, guardrails, escalation rules. A Codex-powered improvement loop that watches production sessions and proposes updates. And a resolution rate that OpenAI claims hits 75% on its own English-language phone support line — without human assistance.
The timing is deliberate. The AI agents market hit $10.9 billion in 2026, growing at 49.6% CAGR toward $182.9 billion by 2033. Gartner forecasts 40% of enterprise applications will embed task-specific AI agents by year-end, up from under 5% in 2025. But here is the number that matters more: only 17% of organizations have actually deployed AI agents to production. The gap between intent and execution is where OpenAI smells margin.
And it is betting billions to close it.
The $4 Billion Consulting Pivot
Presence did not materialize overnight. In May 2026, OpenAI acquired Tomoro, an Edinburgh-based AI consulting firm, as the anchor of a new subsidiary: the OpenAI Deployment Company. Backed by over $4 billion in initial capital from a consortium of 19 firms — including TPG, Advent, Bain Capital, Goldman Sachs, SoftBank, and McKinsey — the Deployment Company absorbed Tomoro's approximately 150 Forward Deployed Engineers.
These FDEs do not sell software licenses. They embed inside enterprises to identify high-value workflows, connect internal systems, configure policies, test agents, and bring them to production. The model is borrowed directly from Palantir, which pioneered the approach for government and intelligence work.
Presence is the Deployment Company's first production product. And it is available only through this high-touch channel — not self-serve, not through a dashboard signup.
"Deployments are led by OpenAI Forward Deployed Engineers and select global systems integrators," OpenAI stated. "Presence is not yet available as a self-serve product."
Translation: if you want OpenAI's agents in your customer service operation, you are buying consulting — not software.
What Presence Actually Does
Each Presence deployment starts with a specific job: resolving billing issues, supporting insurance claims, handling IT service requests. The agent receives only the knowledge and system access required for that task. The enterprise sets the policies — what the agent can do independently, which actions require approval, when a human takes over.
The architecture includes six components that work together:
- Policies and standard operating procedures — the rules that govern agent behavior, configured per workflow
- Guardrails — real-time intervention when interactions move outside defined boundaries
- Approved actions — explicit permissions for what the agent can execute (refunds, account changes, escalations)
- Simulations — pre-launch testing against common requests, edge cases, and high-risk scenarios
- Evaluation tools — graders that check whether the agent reached the right outcome, followed policy, and used tools correctly
- Codex-powered improvement loop — production sessions and escalation patterns feed into Codex, which proposes updates that teams can test against the live version before rolling out
That last component is what separates Presence from a standard chatbot deployment. The improvement loop reduced human handoffs by 15 percentage points in 10 days on OpenAI's own support line. It is designed to solve the drift problem — agents that work at launch but degrade as policies, products, and customer behavior change.
OpenAI's internal deployment serves as the reference case. Its English-language phone support channel at 1-888-GPT-0090 now handles open-ended requests, verifies callers, uses account context, and takes approved actions. The claimed 75% autonomous resolution rate — if independently verified — would place it above industry averages for AI-assisted customer service.
Three launch customers are in early deployments:
- BBVA is exploring AI-powered voice support for everyday banking in Mexico
- SoftBank is testing natural Japanese-language customer conversations, with frontline teams rating the quality highly
- IAG (Insurance Australia Group) is exploring support during high-demand events like severe weather
OpenAI has not disclosed pricing. "During this limited GA phase, deployments are scoped individually based on each customer's use case and implementation needs," a spokesperson told The Register.
The Elephant in the Room: Gartner Says Half Will Fail
OpenAI is launching Presence into a market where the dominant analyst firm is waving a red flag.
Gartner predicted last month that by 2027, 50% of organizations that planned to shift customer service to AI will abandon those plans. Separately, Gartner expects over 40% of agentic AI projects to be canceled by 2027 due to governance and ROI gaps. And only 21% of organizations have a mature governance model for autonomous AI agents.
"While AI offers significant potential to transform customer service, it is not a panacea," said Kathy Ross, senior director analyst at Gartner. "The human touch remains irreplaceable in many interactions, and organizations must balance technology with human empathy and understanding."
The contact center AI market itself is worth an estimated $2.98–4.15 billion in 2026, growing to $13.5 billion by 2034. But 56% of contact center AI projects miss their ROI targets due to integration failures. The problem is not the AI. The problem is the plumbing — connecting models to CRMs, billing systems, identity verification, and policy engines in ways that do not break when edge cases appear.
This is precisely the gap OpenAI is trying to fill with FDEs and governed deployment. Whether that approach can overcome the structural failure rates is the $4 billion question.
Framework #1: Enterprise AI Agent Platform Comparison Matrix
OpenAI Presence enters a market with four established competitors. Here is how they compare across the dimensions that determine production success:
| Dimension | OpenAI Presence | Salesforce Agentforce | Microsoft Copilot Studio | Google Vertex AI Agents | Amazon Bedrock Agents |
|---|---|---|---|---|---|
| Deployment model | High-touch FDE-led | Self-serve + SI partners | Self-serve + M365 admin | Self-serve + Google Cloud | Self-serve + AWS console |
| Pricing | Custom/undisclosed | ~$2/conversation | Included in M365 E3/E5 ($30/user/mo) + capacity packs | Pay-per-use (Vertex pricing) | Pay-per-use (Bedrock pricing) |
| Channel support | Voice + chat (email planned) | Chat, email, Slack, voice | Chat, Teams, web | Chat, voice, CCAI | Chat, voice |
| Governance & policies | Built-in policies, guardrails, escalation rules, simulations | Atlas reasoning engine, guardrails | Responsible AI controls, DLP | Vertex AI safety filters | Guardrails API |
| Improvement loop | Codex-powered auto-proposals | Einstein Analytics | Copilot Analytics | Vertex evaluation | CloudWatch + manual |
| Model flexibility | OpenAI models (3rd-party via API for tools) | Multi-model (OpenAI, Anthropic, Google) | Azure OpenAI + open models | Gemini + open models | Multi-model (Anthropic, Meta, Mistral, Amazon) |
| Best for | Enterprises wanting white-glove AI deployment | Salesforce-native orgs | Microsoft 365-heavy enterprises | Google Cloud customers | AWS-native enterprises |
| Key differentiator | FDE-led deployment, Codex improvement loop | 300+ prebuilt agent templates | Near-zero incremental cost for M365 customers | CCAI heritage + Gemini | Multi-model flexibility |
| Key risk | Vendor lock-in to OpenAI models, opaque pricing | Limited outside Salesforce ecosystem | Copilot fatigue, data sovereignty concerns | Smaller enterprise footprint | Complex multi-service integration |
The strategic insight: OpenAI is the only platform that requires human engineers in the deployment loop. Everyone else offers self-serve. That is either a competitive advantage (higher success rates through guided implementation) or a scaling bottleneck (custom deployments do not grow at SaaS speed). The answer depends on whether the 40% agentic AI cancellation rate is a technology problem or an implementation problem. OpenAI is betting it is the latter.
The Real Competition: Consulting Firms
Presence does not just compete with Salesforce and Microsoft. It competes with Accenture, Deloitte, PwC, and McKinsey — the firms that enterprises currently pay to deploy AI.
The AI deployment war escalated throughout 2026. TCS announced plans for 8,900 forward-deployed AI engineers. PwC signed a $2 billion deal with Anthropic. EPAM committed to 10,000 Claude architects. And now OpenAI has effectively built its own consulting practice, cutting out the middleman entirely.
This creates a fundamental tension for systems integrators. They can partner with OpenAI as "select global systems integrators" (the announcement mentions this channel), but they are also competing with OpenAI's own FDEs for the most valuable implementation work. Tomoro's pre-acquisition client list — Virgin Atlantic, Fidelity International, Tesco, Red Bull, Mattel, NBA — shows the caliber of enterprise OpenAI is targeting directly.
For CIOs, the question is whether to hire OpenAI to deploy OpenAI, or hire a vendor-neutral integrator who can deploy the best model for each workflow. Given that Presence allows third-party models only "through APIs for guardrails, tools, and other parts of their workflow" — not as the core agent — the model-lock-in concern is real.
Framework #2: Enterprise AI Agent Deployment Readiness Assessment
Before evaluating any platform — Presence, Agentforce, Copilot Studio, or otherwise — use this 10-question assessment to determine whether your organization is ready to deploy production AI agents for customer-facing or internal workflows.
Score each question 1–5 (1 = Not Started, 2 = Early Planning, 3 = In Progress, 4 = Mostly Ready, 5 = Production-Ready).
Infrastructure & Data Readiness
1. System Integration Maturity Can your target workflow systems (CRM, billing, identity, knowledge base) be accessed via APIs with sub-second latency?
- Why it matters: 56% of contact center AI projects fail due to integration issues. If your billing system requires batch exports, you are not ready.
2. Knowledge Base Quality Is your customer-facing knowledge base structured, current (updated within 30 days), and machine-readable?
- Why it matters: AI agents hallucinate when knowledge bases are stale. The agent will confidently cite a policy you retired six months ago.
3. Data Sovereignty Compliance Do you know where customer conversation data will be processed, stored, and retained — and does that comply with your regulatory requirements?
- Why it matters: Microsoft Copilot's subprocessor auto-enrollment caught enterprises off guard. Know where your data goes before it gets there.
Governance & Policy Readiness
4. Escalation Policy Definition Have you documented when AI should hand off to a human — not just for failures, but for emotional, legal, regulatory, and reputational scenarios?
- Why it matters: Only 21% of organizations have mature AI agent governance models. Escalation policies are the first test.
5. Action Authorization Framework Can you enumerate every action the agent should be allowed to take (refund limits, account modifications, data access) with explicit approval levels?
- Why it matters: The shadow AI SEC 8-K filing showed what happens when AI touches customer data without governance. Define the boundary before the agent crosses it.
6. Agent Identity & Disclosure Does your regulatory environment require disclosing that customers are speaking to an AI? Do you have a compliant disclosure script?
- Why it matters: China's agent recall regulation and evolving EU requirements are making agent identity disclosure a legal requirement in multiple jurisdictions.
Operational Readiness
7. Baseline Metrics Do you have documented baseline metrics for the workflow you want to automate — resolution rate, handle time, CSAT, cost per interaction?
- Why it matters: Without baselines, you cannot measure ROI. 56% of CEOs cannot prove their AI investments work. Do not join them.
8. Simulation Capacity Can you generate realistic test scenarios (common requests, edge cases, adversarial inputs) that represent your actual customer interaction distribution?
- Why it matters: Presence and Agentforce both emphasize pre-launch simulation. If you cannot simulate, you are deploying blind.
9. Human Oversight Staffing Do you have a team assigned to monitor agent performance, review escalations, and approve improvement proposals?
- Why it matters: AI agents require ongoing human oversight. The myth of "deploy and forget" is why 40% of agentic AI projects face cancellation.
10. Vendor Exit Strategy If the deployment fails or the vendor relationship ends, can you migrate to an alternative platform without losing your policies, training data, and workflow configurations?
- Why it matters: Presence runs on OpenAI models exclusively for the core agent. Agentforce runs inside Salesforce. Copilot Studio runs inside M365. Every choice has a lock-in cost.
Scoring Guide
| Total Score | Readiness Level | Recommendation |
|---|---|---|
| 40–50 | Production-ready | Proceed with vendor evaluation and pilot scope |
| 30–39 | Foundation in place | Address gaps in governance or integration before committing budget |
| 20–29 | Significant gaps | Invest 3–6 months in infrastructure and policy work before deployment |
| 10–19 | Not ready | Focus on basic API integration, knowledge management, and governance frameworks first |
What This Means for Enterprise Strategy
OpenAI's move to become a consulting company disguised as an AI lab reveals three strategic realities:
1. The API era is ending for high-value workflows. Self-serve APIs work for developers building features. They do not work for enterprises deploying customer-facing agents that handle billing, claims, and account management. The complexity of production governance — policies, escalation rules, compliance, continuous improvement — requires human expertise in the loop. This is why every major AI vendor is hiring thousands of deployment engineers.
2. The Gartner warning is a feature, not a bug — for OpenAI. If 50% of enterprises abandon AI customer service plans due to poor implementation, that validates Presence's high-touch model. OpenAI is not selling to the enterprises that would have succeeded with self-serve tools. It is selling to the ones that would have failed without FDEs. The question is whether the premium pricing (still undisclosed) makes the math work relative to a Salesforce Agentforce deployment at $2 per conversation.
3. Model lock-in is the hidden cost. Presence allows third-party models for guardrails and tooling, but the core agent runs on OpenAI models. In a market moving toward model-agnostic architecture, that is a deliberate choice. Enterprises buying Presence are not just buying deployment expertise — they are committing their customer interaction data to OpenAI's ecosystem.
The Bottom Line
OpenAI Presence is the most significant strategic pivot in the enterprise AI market this year. It is not a product launch — it is a business model declaration. The company that gave away ChatGPT to hundreds of millions of free users is now charging enterprise-scale consulting fees to deploy AI agents inside banks, insurers, and telecoms.
The 75% autonomous resolution claim needs independent verification. The pricing remains opaque. The model lock-in concern is real. And Gartner's prediction that half of AI customer service initiatives will fail should give every CIO pause.
But OpenAI is also the only platform offering to put its own engineers inside your organization, watch your agents in production, and use Codex to systematically improve them. In a market where 31% of enterprises have deployed agents but 71% of those agents are just chatbots — the gap between deployment and production is exactly where Presence operates.
The $4 billion question is whether white-glove AI consulting can scale.
Continue Reading
- $9 Billion in 60 Days: The AI Deployment War Just Went Nuclear — How TCS, Microsoft, OpenAI, and Anthropic are spending billions on forward-deployed engineers
- 71% of Your 'AI Agents' Are Just Chatbots in Disguise — The orchestration maturity gap separating real agents from wrappers
- 56% of CEOs Can't Prove AI Works. Here's the Fix. — OpenAI's CFO ROI scorecard and cost-per-task calculator
Rajesh Beri is Head of AI Engineering at Zscaler. Views are his own.
