OpenAI just told every enterprise customer support team that 75% of their inbound calls could be handled without a human being. Not someday. Right now. OpenAI's own support line is already running this way — and the platform doing it is now available, in limited form, to enterprise customers.
That platform is called Presence. It launched on July 22, 2026, and it represents the clearest signal yet that enterprise AI agents have moved from proof-of-concept to production. If you're a CIO, COO, or VP of Customer Operations, this is the announcement you've been waiting for — and the one you need to evaluate carefully before your competitors do.
Here's what Presence is, how it works, why the 75% number matters more than it sounds, and what enterprise leaders should actually do with this information.
What OpenAI Presence Actually Is
Presence is not a chatbot. It's not a low-code workflow builder. It's an enterprise-grade deployment platform for AI agents that handle specific, bounded tasks across voice and chat channels.
The key word is specific. OpenAI designed Presence around the principle that enterprise AI agents should be narrow by design. Each deployment starts with a single defined job — handling billing disputes, managing insurance claims, processing employee IT service requests, onboarding suppliers in procurement — and the agent gets only the information and system access needed for that one job.
This narrowness is a feature, not a limitation. It's what makes guardrails actually work. An AI agent that can access everything is impossible to govern. An AI agent scoped to one function, with approved actions defined upfront, is manageable — and auditable.
The platform combines several capabilities that have traditionally required custom integration work: policies and standard operating procedures, safety guardrails, approved action lists, simulation tools, evaluation frameworks, and a continuous improvement loop powered by Codex (OpenAI's coding agent). That's a meaningful stack to get out-of-the-box.
The 75% Number: What It Means for Enterprise Leaders
OpenAI didn't pick 75% arbitrarily. That's the actual resolution rate on their own English-language phone support line — the line that handles support requests from ChatGPT's millions of users. They ran Presence on their own infrastructure before selling it to anyone else.
For enterprise leaders, there are two ways to interpret this number:
The optimistic read: Three-quarters of inbound support volume can be automated without degrading the customer experience. In a contact center handling 10,000 calls per month, that's 7,500 calls handled at effectively zero marginal cost once the system is deployed and optimized.
The careful read: 75% means 25% still escalates to humans. The question every operations leader needs to answer is: which 25%? Are those the easiest escalations or the hardest? Are they predictable or random? How does escalation affect customer satisfaction for the cases that do require human intervention?
According to reporting from SiliconAngle, Presence tracked human escalation rates as a key metric in its own deployment and uses those metrics to drive the Codex-powered improvement process. So the system is designed to move that 25% down over time — not through brute force, but through structured learning from production sessions that humans review and approve before changes go live.
That last point matters enormously for compliance-conscious industries. Changes to AI agent behavior require human sign-off before deployment. That's not just a nice-to-have — it's the kind of governance structure that legal, compliance, and risk teams need before they'll let any AI agent touch customer interactions at scale.
How Presence Works: The Technical Architecture for CTOs
For technical leaders evaluating Presence, the architecture is built around four phases: define, simulate, deploy, and improve.
Define: Each deployment starts with a job definition — the specific task the agent performs, the systems it can access, the actions it can take, and the policies it must follow. This is where standard operating procedures become machine-readable rules. Retailers, for example, can restrict which product databases an agent can query. Financial services firms can define which account types an agent can modify and under what conditions.
Simulate: Before going live, the platform runs the agent through simulated scenarios covering common cases, edge cases, and high-risk situations. The simulation tooling evaluates whether the agent reaches the correct outcome, follows stated policy, uses tools appropriately, and escalates at the right moment. This is the step that de-risks production deployment — you're catching failure modes before they hit real customers.
Deploy: Presence runs on GPT-5.6 Sol, the model that recently set records on OSWorld, a benchmark measuring performance on multistep knowledge work tasks. This is not the same as deploying a generic language model. Sol was specifically designed for agentic tasks — following multi-step workflows, using tools correctly, and knowing when to stop and ask for help.
Improve: Post-launch, OpenAI's Codex agent monitors production sessions continuously. It analyzes escalation patterns, identifies interactions where the agent underperformed, and generates specific recommendations for behavioral improvements. Those recommendations go to the company's internal team for review. Staff approve or reject changes. Approved changes go live. Nothing deploys automatically. This is the continuous improvement loop that will separate mature Presence deployments from early ones over a 12-18 month horizon.
For CTOs, the key integration question is system access: what data sources and APIs does the agent need, and how do you govern that access at the enterprise level? That's where Forward Deployed Engineers come in — Presence is not self-serve, and that's intentional.
Business Impact: What CFOs and COOs Need to Model
For business leaders, Presence creates a new calculation that most organizations haven't had to make yet: what is the total cost of a Presence deployment versus the total cost of the human operations it replaces or augments?
The industry data on AI agent ROI is promising but uneven. A Forrester Research study of 287 enterprise AI agent deployments across 14 industries found an average ROI of 540% within 18 months, with a median payback period of 7.3 months. Cost savings of 26-31% have been reported across supply chain and procurement, finance and accounting, and customer operations functions. (Source: onereach.ai ROI analysis)
The counterpoint: IBM research shows only about 5% of organizations achieve what they classify as substantial AI ROI — meaning investments that demonstrably improve the bottom line after accounting for total implementation costs including tooling, integration, training, and organizational change management. (Source: masterofcode.com AI ROI analysis)
The gap between those two data points — 540% ROI for the successful minority versus 95% of companies not hitting substantial returns — comes down to deployment discipline. Organizations that define measurable outcomes before deploying are four times more likely to achieve ROI than those that deploy and then try to measure impact.
Presence is architected for the disciplined approach. The scoped deployment model, simulation testing, and continuous improvement loop are all designed to produce measurable outcomes. But the platform does not do the business case work for you. CFOs evaluating Presence need to model:
- Current cost per resolved support interaction (fully loaded)
- Projected resolution rate based on call type complexity
- Integration costs (Presence requires FDE support — not a one-day setup)
- Training and change management costs for the human review function
- Expected improvement trajectory over 12-24 months
The 25% that still escalates to humans may actually become more expensive to handle over time if the agent handles the easy cases and leaves humans with the most complex, emotionally charged interactions. Contact center leadership needs to plan for that shift.
Where Presence Fits Across Enterprise Departments
OpenAI has signaled three primary use cases in public communications, but the model generalizes across several enterprise functions.
Customer Support (the flagship use case): Billing disputes, order status, returns, troubleshooting, account management. This is the highest-volume entry point for most enterprises and the use case OpenAI validated on their own support line.
IT Service Desk: Employee requests for system access, password resets, software installation, hardware issues. IT service desks typically handle thousands of repetitive requests monthly — this is ripe for automation with well-defined escalation rules.
Sales Operations: Lead data collection, prospect qualification, initial outreach coordination, pipeline data entry. Presence's task-specific architecture means you can scope a sales agent to data enrichment without giving it access to pricing systems or contract management.
Procurement: Supplier onboarding is specifically called out in OpenAI's materials. Verifying supplier information, collecting documentation, routing approvals — these workflows have clear policies and defined steps, making them well-suited for Presence's define-and-simulate approach.
HR Operations: Benefits enrollment support, policy questions, onboarding task coordination. HR interactions require careful guardrails (compliance with benefits regulations, privacy requirements), which is exactly the use case the platform's governance model was built for.
For enterprises with large operations teams, the highest-value entry point is whichever function has the highest call/request volume and the most clearly defined policies. Start there. Expand once the first deployment is optimized.
What Makes Presence Different from Other Enterprise AI Agent Platforms
The enterprise AI agent market has been crowded for two years. Salesforce Agentforce, Microsoft Copilot Studio, ServiceNow AI Agents, and a long list of venture-backed startups have all been pitching agentic automation to enterprise buyers. What makes Presence a different conversation?
Three things stand out:
The proof point is internal. OpenAI deployed Presence on their own support operations before selling it. That's a meaningful signal. Most platform vendors sell capabilities they've never operated at production scale. OpenAI ate their own cooking — and the number they're publishing (75% resolution rate) is specific enough to be audited and challenged. That specificity builds more credibility than vague claims about efficiency improvements.
The model underneath is purpose-built for agents. GPT-5.6 Sol was specifically optimized for multistep agentic tasks, not just language generation. The gap between a language model that can write about resolving support tickets and a model that can actually navigate enterprise systems, follow policy, and know when to stop has been the core failure point for most early enterprise AI deployments.
The improvement loop is human-gated. Codex-powered suggestions, human approval before deployment — this is the governance model that regulated industries have been asking for. Healthcare organizations, financial services firms, and insurers who've been watching AI agents from a distance because of governance concerns have a credible answer now.
The limitation: it's not self-serve, and that's a real constraint. Companies that want to move fast and experiment freely won't find that here. Presence requires OpenAI Forward Deployed Engineers and select systems integrator partners. For enterprise buyers, that means longer sales cycles, implementation timelines measured in weeks not hours, and costs that will vary based on scope.
What Enterprise Leaders Should Do This Week
This announcement warrants action, but not panic-buying. Here's a practical framework for different roles:
For CIOs and CTOs: Add Presence to your 2026 enterprise AI evaluation pipeline. Contact your OpenAI account team about the limited GA program. Use the time before deployment to inventory which internal workflows have the clearest policy documentation — those are your first candidates.
For COOs and VP Operations: Start the conversation with your contact center and operations leadership about baseline metrics: current cost per interaction, resolution rate, escalation rate, and customer satisfaction scores by channel. You need those numbers before any Presence business case conversation makes sense.
For CFOs: Don't approve a Presence deployment without a defined ROI model that accounts for all costs — not just licensing. The IBM data showing only 5% of companies achieve substantial AI ROI is a warning about deployment discipline, not a warning against deployment. The companies that succeed model outcomes upfront and measure them continuously.
For CLOs and Compliance Leaders: This is the moment to get ahead of AI agent governance policies in your organization. Presence's human-approval architecture is a starting point, but your organization needs documented policies on what AI agents can and cannot do, how interactions are logged and audited, and how errors are remediated. Get that framework defined before the technology arrives — not after.
The Bottom Line
OpenAI Presence is the most enterprise-ready AI agent platform announced to date, with a verifiable performance benchmark (75% call resolution) and a governance model built for compliance-conscious organizations. It's not cheap, not fast to deploy, and not right for every use case — but it's real.
The 75% number is the headline, but the architecture is the story. Narrow by design, governed by humans, improved by machine review with human approval — that's the template that makes enterprise AI agents trustworthy at scale.
The question for enterprise leaders isn't whether AI agents will handle a significant share of customer and internal operations. That question has been answered. The question is whether your organization will be the one optimizing a production deployment in six months, or the one still running pilots while your competitors are posting ROI results.
Presence is OpenAI's answer to that urgency. The limited GA window is the signal. Get in the evaluation queue now.
THE DAILY BRIEF covers Enterprise AI for Technical and Business Leaders. Published Tuesday and Thursday.
Follow on X: @rajeshberi | LinkedIn: Rajesh Beri
