OpenAI has introduced Presence, a managed service designed to deploy voice and chat agents in production enterprise workflows, including billing, insurance claims and IT support. The launch matters less as a new conversational interface than as an attempt to package the operating systems around agents: monitoring, escalation, evaluation and feedback loops.
That operational layer is where many enterprise agent projects have stalled. A model may produce persuasive answers in a controlled demonstration yet remain unsuitable for customer-facing work if teams cannot reliably detect errors, route risky cases to people, measure outcomes or improve the system without rebuilding it. OpenAI is positioning Presence as infrastructure intended to close that gap. The company says its own phone-support deployment resolves 75% of inbound issues without human assistance, and that a Codex-powered improvement loop reduced human handoffs by 15 percentage points over 10 days. Those are company-reported figures, not independently audited performance results. [1]
By the numbers
- 75% — inbound phone-support issues OpenAI says its internal deployment resolves without human assistance.
- 15 percentage points — reported reduction in human handoffs from a Codex-powered improvement loop.
- 10 days — period over which OpenAI says that handoff reduction was achieved.
- 50% — share of organizations Gartner predicted would abandon plans for AI-driven customer-service workforce reductions by 2027.
From an agent model to a production service
Presence is aimed at a familiar but difficult enterprise pattern: an agent receives a customer request, gathers information from company systems, decides whether it can complete the task, executes or recommends an action, and transfers the case when it cannot proceed safely. Billing disputes, claims status questions and IT support all appear tractable because they often begin with repetitive requests. They also expose an agent to consequential edge cases, fragmented records, authentication requirements and policies that change over time.
A production platform therefore needs more than speech recognition, text generation and a connection to a knowledge base. It needs observability across the full interaction: what the customer asked, which tools or records the agent consulted, which policy or instruction informed the answer, whether a tool call succeeded, and what happened after the conversation. For voice deployments, that record must cover recognition errors, latency, interruptions, transfers and the relationship between a spoken answer and the system action behind it.
OpenAI’s framing of Presence emphasizes the operational controls required to manage those interactions. Evaluation is particularly important. A useful evaluation program does not merely score whether a reply sounds natural; it tests policy adherence, factual grounding, correct tool use, successful task completion, appropriate escalation and the quality of the handoff context supplied to a human agent. Organizations also need separate tests for routine, ambiguous and adversarial cases, because average answer quality can conceal failures in the small share of cases with the highest financial, legal or customer impact.
Escalation is the other essential control. Dependable automation is not defined by avoiding humans at all costs. It is defined by recognizing when a case falls outside the agent’s authority, evidence or confidence threshold, then transferring it with a concise record of the customer’s issue, verified facts, attempted steps and unresolved decision. Poorly designed handoffs simply shift work to staff while forcing customers to repeat themselves.
The feedback loop is the substantive product claim
The most significant element of the announcement is the reported Codex-powered improvement loop. OpenAI describes a system in which operational evidence can be used to identify sources of handoffs and improve the agent. In principle, this converts a support operation into a continuous engineering process: group failed or transferred interactions by cause, inspect transcripts and tool traces, modify instructions, retrieval content, workflow logic or integrations, evaluate the changes against a fixed test set, and deploy only when the new version improves the relevant measures.
Codex can potentially accelerate parts of that cycle, such as analyzing recurring failure patterns, drafting workflow changes, generating tests or helping maintain integrations. But its value depends on governance rather than code generation alone. Enterprises need version control, approval gates, rollback mechanisms, access controls and audit trails. An automated proposal to revise an escalation policy or a claims workflow should not become a production change without accountable review.
OpenAI’s reported 15-percentage-point reduction in handoffs in 10 days is encouraging as an indication that targeted iteration may improve coverage quickly. It does not, on its own, establish general performance. The announcement does not make that figure directly comparable across industries without details such as the initial handoff rate, the types of calls involved, the definition of a resolved issue, customer-satisfaction results, repeat-contact rates, error severity and the amount of human work needed to run the loop. A lower transfer rate can be beneficial only if it does not mean the agent is retaining cases that should have been escalated.
A credible route beyond pilots, with demanding proof points
Presence addresses a genuine reason that enterprise agent pilots struggle to become broad deployments. Pilots can be successful with curated prompts, a small set of intents and engineers closely watching results. At scale, organizations face changing source data, long-tail customer requests, outages in connected systems, security events and exceptions that do not fit an intent taxonomy. Managed monitoring, evaluations and feedback tooling can reduce the burden of assembling those capabilities internally.
Whether Presence becomes a credible path to dependable automation will depend on how well it performs under those conditions. Buyers should ask for measurable service-level evidence rather than a headline containment rate. Core measures include first-contact resolution, repeat contact, transfer accuracy, time to resolution, tool-call failure rate, customer satisfaction, cost per resolved case and the rate of material errors. Results should be segmented by workflow and risk level. An agent handling a password reset is not comparable with one explaining a disputed charge or processing a claim-related request.
Data governance will be equally important. Customer-support systems often contain payment information, account histories, health-related details, employment records or other sensitive data. Enterprises need clear rules governing authentication, permissions, retention, redaction, connector access, logging and the use of interaction data in improvement workflows. A managed service can simplify operations, but it also makes vendor due diligence, contractual controls and exit planning more consequential.
Market implications for contact centers and enterprise software
The launch puts OpenAI in closer competition with contact-center platforms, customer-service software vendors, systems integrators and specialist agent providers. These companies already sell pieces of the stack, including routing, agent assist, quality assurance, analytics, knowledge management and workflow automation. Presence’s opportunity is to make the model and its operational lifecycle feel like one integrated service rather than a collection of components an enterprise must assemble.
The more immediate effect is likely to be a shift in procurement criteria. Enterprises evaluating agents will increasingly look beyond model benchmarks and chatbot demonstrations toward the quality of evaluation tooling, escalation design, integration reliability and change management. Implementation partners may also become more important, because the highest-value work is often translating informal procedures and policy exceptions into controlled workflows.
The labor implications are more complex than a simple replacement narrative. Automation can reduce the volume of straightforward contacts, while increasing the share of cases that require judgment, empathy, exception handling or cross-system investigation. Gartner said in 2025 that 50% of organizations expecting major customer-service workforce reductions because of AI would abandon those plans by 2027, citing the difficulty of achieving those reductions. [2] Presence does not remove that constraint; it is an effort to make a larger portion of work automatable without abandoning controls.
What could limit adoption
The key risk is that a managed operational layer may create confidence faster than it creates reliability. Monitoring dashboards and evaluation scores are valuable, but they are only as strong as the test cases, labeling standards and business metrics behind them. If teams measure short calls and fewer handoffs while neglecting incorrect outcomes, repeat contacts or vulnerable customers, optimization can move in the wrong direction.
Voice support adds additional challenges. Background noise, accents, interruptions, poor mobile connections and authentication failures can turn an otherwise capable workflow into a frustrating experience. In regulated or high-stakes workflows, an agent must also distinguish between providing information, taking an authorized action and making a decision that requires a human reviewer. The escalation policy must be designed around those boundaries, not appended after deployment.
For OpenAI, the test will be whether Presence can demonstrate transparent, repeatable results across customers rather than just an internal support operation. For customers, the practical test is narrower: can the service improve resolution and service quality while preserving auditability, safe transfers and control over sensitive data? Those are the criteria that separate a convincing pilot from dependable automation.
Editor’s Take
I see the operational loop as the real story here. Enterprises do not need another agent that can answer a clean demo question; they need a system that shows why it took an action, knows when to stop, and gives an operations team a disciplined way to improve it next week. If Presence makes evaluation, trace review and escalation materially easier to run, it could eliminate a substantial amount of fragile, bespoke agent engineering.
I would treat OpenAI’s 75% resolution and 15-point handoff figures as promising internal signals, not a buying benchmark. The next evidence that matters is independently meaningful customer data: task-level error rates, repeat-contact outcomes, escalation quality and performance over months as policies and integrations change. The winning implementations will optimize for correct resolution, not merely lower handoff counts.
