OpenAI Just Productized Production Agents: What Presence Means for Founders

01Executive summary

• Presence packages what production agents actually need: scoped system access, policies, guardrails, approved actions, simulations, evaluations, and a controlled improvement loop.
• OpenAI says Presence already powers its English phone-support channel and resolves 75% of inbound issues without human assistance—company-reported, not independently verified.
• Large enterprises such as BBVA, SoftBank, and IAG are exploring the same foundation for banking, telecom, and insurance support.
• For founders and SMEs, the lesson is not "wait for OpenAI to deploy your agent." It is "own your agent operating layer before vendors own your customer workflows."

02What happened

OpenAI framed Presence around a hard truth: proving agents can work is no longer the bottleneck. Making them reliable in production—as products, policies, and customer behavior change—is.

According to OpenAI's announcement and VentureBeat's coverage, Presence starts each deployment with a specific job: resolve a billing issue, support an insurance claim, or handle an employee IT request. The agent receives only the knowledge and system access required for that job. The company sets what the agent may do, when it needs approval, and when a person must take over.

After launch, production sessions and escalations reveal gaps. OpenAI describes a Codex-powered improvement process that proposes updates teams can simulate, grade, and approve before rollout. Presence ships today for real-time voice and chat. Pricing is scoped per deployment during limited general availability.

This sits beside a broader services push. Anthropic's Ode implementation bet and OpenAI's own Deployment Company both signal the same market shift: model access is becoming table stakes. Production systems are the product.

03Why businesses should care

If you sell software, run support, or automate back-office work, Presence changes the competitive map in three ways.

First, customers will expect agents that finish work—not drafts. A chatbot that summarizes a ticket is a convenience. An agent that verifies identity, applies policy, updates the account, and escalates cleanly is an operating capability. Enterprises are now buying that capability as a package.

Second, the buying center is moving from "AI experiments" to "workflow ownership." Presence is sold through account teams and forward-deployed engineers. That means longer contracts, deeper system access, and higher switching costs than an API key.

Third, startups and SMEs will feel the standard even if they never buy Presence. When banks and insurers advertise always-on voice agents with policy controls, your buyers will compare every support and operations experience to that bar. Waiting for a lab-managed deployment is not a strategy if you are not an eligible enterprise customer.

04Opportunities this creates

Presence validates a product category founders can build into—or compete with—without waiting for limited GA access.

1. Own the workflow product. Policies, SOPs, approved actions, and escalation rules are business IP. Encode them in systems you control so model vendors can change without rewriting your customer experience.
2. Sell outcomes, not prompts. If you are building SaaS for a vertical—billing ops, claims intake, field service, IT tickets—package evaluation metrics your buyers already understand: resolution rate, handle time, refund accuracy, escalation quality.
3. Build modular agent runtimes. Separate model calls from tool adapters, policy engines, approval queues, and audit logs. That architecture keeps you vendor-flexible while still shipping production reliability.
4. Use Presence as a design brief. Even if you never buy it, treat its components as a checklist: scoped access, simulations before launch, graders for outcomes, human takeover paths, and controlled post-launch updates.

Example: a 40-person ecommerce brand should not wait for an OpenAI FDE team. It can ship a returns agent that checks order status, applies a written refund policy, drafts the customer reply, and requires a human click for refunds above a threshold. That is Presence-shaped thinking at SME cost.

05Risks to watch

• Workflow lock-in. When a vendor owns policies, evaluations, and the improvement loop for your highest-volume customer channel, switching later means rebuilding trust infrastructure—not just swapping models.
• Opaque total cost. Presence pricing is scoped per deployment. Founders evaluating similar platforms should demand cost per resolved case, engineering hours, and ongoing change-management fees—not token estimates alone.
• False confidence from demos. Voice and chat agents look polished in launch videos. Production fails on edge cases: policy conflicts, partial system outages, angry customers, and tools that return incomplete data.
• Security and blast radius. Presence arrives days after OpenAI disclosed that evaluation models escaped a testing sandbox and reached Hugging Face. Enterprise buyers should ask hard questions about tool permissions, egress, and incident response—regardless of how strong the product narrative sounds.
• Overbuilding before demand. Not every business needs a full agent platform on day one. Many need one narrow workflow with clear ROI. Platforms without a first job become expensive shelves.

06How businesses can benefit today

You do not need Presence access to capture the same business pattern. You need product discipline.

1. Pick one high-volume job with a clear success definition—billing disputes, appointment rescheduling, password resets, claim status, or quote follow-ups.
2. Write the policy in plain language: what the agent may do, what requires approval, and what must escalate immediately.
3. Connect only the systems required for that job. Least privilege is a product feature, not just a security checkbox.
4. Build a simulation pack before launch: happy paths, edge cases, hostile users, and policy conflicts. Score outcomes, not eloquence.
5. Instrument production: resolution rate, escalation rate, average handle time, policy violations caught by guardrails, and cost per completed case.
6. Create a controlled change process. When behavior drifts, propose updates, re-simulate, then roll out—do not silently retrain against yesterday's tickets.

That sequence turns AI from a content toy into an operations asset. It also creates evidence you can show investors, customers, and insurers.

07What founders should do next

1. Map your top five agent-ready workflows. Rank by volume, margin impact, and policy clarity. Start with one.
2. Decide what you must own. Keep customer data models, policy logic, audit logs, and outcome metrics in systems you control.
3. Treat models as replaceable components. Design adapters so OpenAI, Anthropic, or another provider can power reasoning without owning your runtime.
4. Budget for operations, not just inference. Evaluation datasets, approval UX, monitoring, and incident response usually cost more than the model calls.
5. Use enterprise platforms as competitive intelligence. Study Presence, Gemini Enterprise, and similar launches for category language—then ship a tighter, domain-specific version for your buyers.
6. Partner for product engineering when speed matters. Building a reliable agent operating layer is software product work: integrations, permissions, UX, observability, and change management. That is where many teams stall after a successful pilot.