OpenAI Presence Puts Voice and Chat Agents Into Production: Who Takes Over When the AI Gets Stuck?

OpenAI Presence brings managed voice and chat agents into live workflows, but the operational test is not whether the AI can answer routine questions. It is whether a Sydney business can transfer an uncertain, emotional or high-risk interaction to the right person with the context, authority and time to act.

Human takeover must be designed into the service, covering escalation triggers, queue ownership, after-hours coverage, approvals, records and recovery.

OpenAI has moved its enterprise-agent proposition beyond model access and into managed production operations. Announced on 22 July 2026, OpenAI Presence combines voice and chat agents with company policies, standard operating procedures, permissions, guardrails, approved actions, simulations, evaluations and controlled improvement processes.

The product is aimed at workflows such as customer support, outbound sales, insurance claims, employee services, procurement and internal IT assistance. OpenAI says each deployment begins with a defined job and limited access to the knowledge and systems required for that job. Enterprises decide what the agent can do independently, when approval is required and when a person must take over.

That final element may prove to be the most consequential. An agent that can resolve straightforward matters creates value. An agent that recognises its limits, transfers control without losing information and reaches an authorised human quickly creates an operational service.

The Launch Changes the Question From Conversation Quality to Service Continuity Recent discussion about voice AI has focused on latency, natural speech, interruption handling, accents, background noise and whether customers will tolerate an automated conversation. Elyment examined those issues in its analysis of whether GPT-Live voice agents are ready for real customer calls .

Presence introduces a different operational question: what happens after the agent determines that it cannot safely or confidently continue?

OpenAI says Presence now resolves 75 per cent of inbound issues on its English-language phone-support channel without human assistance. It also reports that its improvement process reduced human handoffs by 15 percentage points in 10 days. These figures indicate substantial automation potential, but they do not make human operations irrelevant.

Lower handoff volumes can leave human teams with a more concentrated queue of difficult, emotional, unusual or high-risk matters. The organisation may handle fewer interactions manually while requiring greater skill, authority and judgement for each one.

Handoff reduction is therefore not the same as handoff elimination. It changes the work performed by the human team.

A Handoff Is a Transfer of Operational State Many automation projects treat escalation as a routing instruction: send the conversation to a person when confidence falls below a threshold. That is only the beginning.

A reliable takeover requires five connected elements:

A trigger. The system must recognise uncertainty, risk, distress, conflicting information, missing authority or a policy exception. A destination. The case must reach a named role or accountable queue, not a generic inbox that may remain unattended. A context package. The human must receive the verified identity, conversation summary, actions taken, unresolved issue and any commitments already made. Decision authority. The recipient must be authorised to solve the matter, approve the exception or direct the next operational step. A response target. The business must define how quickly the human accepts control and what happens when the first recipient is unavailable. Without all five, the customer may be transferred but the problem remains stationary.

Example Escalation Settings for a Sydney Property and Service Operator These are operational examples, not universal legal standards.

Customer disputes a quoted inclusion or price

AI response: Stop making commitments and compile the quote history.

Human owner: Estimator or commercial manager.

Example service target: Acceptance within 15 business minutes.

Caller reports an injury or potentially dangerous site event

AI response: Prioritise safety information and trigger the emergency pathway.

Human owner: WHS or incident duty manager.

Example service target: Immediate takeover.

Resident reports active water entry after hours

AI response: Confirm location, occupancy and contact details without giving unsafe instructions.

Human owner: After-hours property coordinator.

Example service target: Immediate alert and rapid callback.

Agent detects distress, vulnerability or repeated misunderstanding

AI response: Offer a human conversation and preserve the interaction context.

Human owner: Senior customer-service lead.

Example service target: Warm transfer where available.

Requested action exceeds the agent’s approval limit

AI response: Explain that authorisation is required and create an approval record.

Human owner: Delegated approver.

Example service target: Based on the commercial or operational deadline.

Voice Failure Is Operationally Different From Chat Failure Voice and chat can share policies, knowledge sources and escalation rules, but they should not be managed as identical channels.

A chat user may tolerate a short delay while a specialist joins the conversation. A caller is already occupying a synchronous channel. Silence, repeated verification, an unexpected disconnection or a vague promise of a callback can feel like abandonment.

Voice interactions also contain information that may not appear in the words alone. Pace, repetition, confusion, interruptions and emotional intensity can change the risk profile of the call. The agent may need to escalate because of how the interaction is unfolding, not simply because a keyword appeared.

A production voice service should therefore distinguish between:

Warm transfer: The human joins while the caller remains connected. Priority callback: The caller leaves the channel but receives a defined response time. Specialist referral: The issue moves to a qualified team with a case reference and full context. Emergency diversion: A safety or critical-property event bypasses normal queues. Controlled refusal: The agent declines an unsupported action and explains the approved next step. The correct pathway depends on urgency, risk, customer expectations and the authority required to resolve the matter.

Sydney Property Operations Expose Weak Escalation Design Quickly Property, renovation and infrastructure workflows rarely remain inside one clean digital process. A customer conversation can move rapidly from a routine enquiry to a live access issue, a site variation, a safety concern, a strata restriction or a payment dispute.

Consider an AI agent receiving enquiries for a Sydney apartment renovation:

A resident asks whether workers can enter while the building manager is unavailable. A contractor reports that the actual floor build-up differs from the quoted scope. A neighbour complains about noise outside the approved strata work window. A customer asks the agent to approve additional floor levelling without a written price. A worker reports an injury, hazardous material or dangerous electrical condition. A purchaser asks whether keys can be released before the authorised settlement process is complete. The AI may be able to collect facts, verify identities, locate records and explain the standard process. It should not improvise an approval simply because the caller demands an immediate answer.

The business needs a precise takeover map. Access questions may belong to a project coordinator. Scope changes may require an estimator or project manager. Strata complaints may require a building manager. Payment and contractual decisions may require an authorised commercial or legal professional. Safety reports require a separate incident pathway.

For notifiable workplace incidents, SafeWork NSW states that a death, serious injury or illness, or dangerous incident must be reported immediately. Its incident-notification guidance allows notifications 24 hours a day. An AI intake system must not allow a potentially serious report to wait in an ordinary customer-service queue.

The Human Team Needs Named Ownership, Not General Availability A common implementation mistake is to confirm that human support exists without specifying which human is accountable at each moment.

The operating model should name at least the following responsibilities:

Service owner: Accountable for the performance of the full voice or chat service. Queue owner: Responsible for ensuring escalated cases are accepted and distributed. Subject-matter owner: Resolves technical, legal, commercial or project-specific exceptions. Incident owner: Takes control of safety, security, privacy and critical-service events. System owner: Manages integrations, permissions, failures and recovery. Quality owner: Reviews failed interactions and approves changes to policies, prompts or workflows. In a small Sydney operator, one person may hold several roles. The requirement is not a large organisational chart. It is explicit accountability.

The model must also cover absences, lunch periods, weekends, public holidays and after-hours incidents. An escalation pathway that only works when one senior employee is watching a screen is not a production pathway.

The Customer Should Not Have to Reconstruct the Conversation Poor handoffs often force the customer to begin again: state their name, repeat the problem, explain what the agent said and identify which actions may already have occurred.

That repetition increases frustration and introduces new errors. The customer may describe the situation differently the second time, while the human may not know which part of the first interaction was verified.

A useful handoff package should include:

The customer’s verified identity and contact details. The service address, job number, property or account reference. A concise summary of the request and the reason for escalation. The policies, records and knowledge sources consulted. Every system action attempted or completed. Any prices, dates, timeframes or outcomes communicated. The unresolved decision requiring human judgement. The urgency, risk classification and required response time. The customer’s preferred channel, language or accessibility requirements. The appropriate transcript or recording reference, subject to privacy controls. The summary must also distinguish between facts supplied by the customer, information verified against a system and conclusions generated by the agent. Blending those categories can cause a human operator to treat an assumption as an established fact.

Automation Can Reduce Volume While Increasing Human Complexity Workforce planning for AI-supported service operations cannot be based only on the expected number of handoffs.

A basic planning equation is:

Expected human workload = total interactions × escalation rate × average human handling time

That calculation still requires adjustments for peak periods, simultaneous calls, specialist availability, language coverage, callback delays and the proportion of matters requiring senior approval.

If the agent resolves simple password resets, booking confirmations and status enquiries, the human queue may become dominated by disputes, unusual site conditions, distressed callers and cross-system failures. Average handling time can rise even as total case volume falls.

This matters in property operations, where demand is rarely even. Severe weather, building access failures, settlement deadlines, contractor delays and urgent maintenance can produce short periods of intense contact. OpenAI has identified insurance support during severe weather as one of the use cases being explored with Australian insurer IAG.

The correct staffing model is therefore not simply fewer people. It may involve a smaller, more capable frontline team supported by on-call specialists and clearly delegated decision authority.

Money, Safety and Commitments Need Explicit Approval Boundaries Presence is designed around approved actions and permission limits. For operators, this means translating broad policies into precise boundaries that software and staff can follow.

A voice or chat agent should not be permitted to improvise decisions such as:

Changing a contracted price or approving an unquoted variation. Promising a refund, discount or compensation above a delegated threshold. Confirming that a site, product or property condition is safe without qualified assessment. Giving legal, engineering or technical advice beyond an approved information scope. Releasing keys, access codes or security information without verified authority. Changing a settlement, payment or project milestone without approval. Overriding strata, building-access or work-hour restrictions. Concealing a service failure by offering an unsupported commitment. The Australian Consumer Law remains relevant to the service delivered to customers. The ACCC’s guidance on consumer rights and guarantees explains that services may need to be provided with due care and skill, be fit for their stated purpose and be supplied within a reasonable time where no timeframe has been agreed.

From an operational-risk perspective, businesses should treat statements made by their agents as statements made through their service channel. An internal disclaimer that the response came from AI does not repair an incorrect price, unsafe instruction or unauthorised promise after it has affected the customer.

Privacy and Call Recording Must Follow the Interaction Into the Handoff Voice and chat agents can process names, addresses, phone numbers, payment information, insurance details, employment information, property records and descriptions of personal circumstances. The handoff process can increase exposure if full transcripts are copied into multiple systems or distributed to employees who do not need them.

The Office of the Australian Information Commissioner states in its guidance on commercially available AI products that the Privacy Act applies to uses of AI involving personal information.

Sydney operators using recorded voice interactions should also obtain advice on the application of the NSW Surveillance Devices Act 2007 , along with any federal privacy, employment, telecommunications and sector-specific requirements relevant to the deployment.

Practical controls may include:

Clear notification that the customer is interacting with an AI system. Appropriate notice where calls or conversations are recorded or transcribed. A defined lawful purpose for collecting and retaining information. Role-based access to recordings and transcripts. Redaction or masking of unnecessary sensitive information. Retention periods linked to genuine operational requirements. A human pathway for privacy questions or correction requests. Controls preventing production conversations from entering unapproved tools. The privacy principle should be simple: the human receives enough information to resolve the matter, not every available piece of information about the person.

NSW’s AI Framework Offers a Useful Accountability Test Private businesses are not automatically subject to every NSW Government AI policy. However, the governance questions used by the public sector provide a useful benchmark for commercial deployments.

Digital NSW’s AI Assessment Framework responsibilities call for documented accountability for risk management, service continuity, appeals and evidence supporting decisions and actions.

Those principles translate directly into a production-agent review:

Who is accountable when the agent produces an incorrect outcome? Can a customer challenge a decision and reach a capable human? Does the business retain enough evidence to reconstruct what occurred? Can the service continue when the agent, integration or telecommunications channel fails? Who may approve a policy change after reviewing production conversations? How is a failed update reversed without losing active cases? These are operating-model questions rather than model-performance questions.

A Practical Production-Readiness Sequence Businesses considering Presence or another production-agent platform should complete the following sequence before opening the service to customers.

Map the real workflow. Review actual calls, chats, exceptions, complaints and operational failures. Elyment’s analysis of why bad automations continue to cost businesses explains why teams should map decisions and handoffs before scaling technology. Classify the interaction types. Separate routine information, approved transactions, judgement calls, regulated matters, vulnerable-customer interactions and emergencies. Define the agent’s authority. Document what it can read, change, approve, communicate and refuse. Assign every escalation destination. Give each trigger a named role, backup owner, operating schedule and response target. Design the context package. Decide what verified information, summaries, records and system actions accompany the transfer. Test failure paths. Simulate unavailable staff, disconnected calls, conflicting databases, failed identity checks, distressed customers, urgent hazards and unauthorised requests. This extends the production preparation discussed in Elyment’s review of what businesses should prepare before agents move into production . Confirm containment and recovery. Establish credential revocation, tool isolation, log preservation and shutdown authority. Elyment’s examination of autonomous-agent containment and incident readiness explains why production access requires more than ordinary software permissions. Run a controlled launch. Begin with narrow intents, limited actions and staffed monitoring before expanding volumes or authority. Review every material escalation. Determine whether the agent escalated at the correct time, reached the right person and transferred sufficient context. The Metrics That Reveal Whether Human Takeover Actually Works Resolution rate alone can conceal a weak service. A business may report that the agent completed most interactions while customers who needed assistance waited too long, reached the wrong team or abandoned the channel.

Operational Measures for Voice and Chat Escalation Performance Escalation acceptance time

What it reveals: How long it takes an accountable person to take control.

Successful warm-transfer rate

What it reveals: Whether voice callers remain connected to the correct human.

Repeat-explanation rate

What it reveals: Whether the handoff context is sufficient and trusted.

Wrong-queue rate

What it reveals: Whether escalation rules identify the correct operational owner.

Abandonment after escalation

What it reveals: Whether customers leave while waiting for assistance.

Unauthorised commitment rate

What it reveals: Whether the agent promises actions outside its delegated limits.

Human override reason

What it reveals: Which policies, tools or knowledge sources require improvement.

Reopened-case rate

What it reveals: Whether the initial resolution was complete and accurate.

Critical-event escalation latency

What it reveals: Whether safety, security and urgent property events bypass ordinary queues.

Privacy overexposure events

What it reveals: Whether handoffs disclose more personal information than the recipient requires.

These measures should be reviewed by operational leaders, not confined to the technology team. They reveal staffing weaknesses, unclear authority, broken procedures and gaps between the service promised to customers and the service the organisation can actually deliver.

The Human Is Not a Backup System OpenAI Presence signals that voice and chat agents are becoming managed production infrastructure. They can connect to business systems, follow procedures, take approved actions and improve through controlled evaluation.

Yet the value of that infrastructure will depend on what surrounds it.

A Sydney business deploying an agent across customer service, property coordination, renovation enquiries, claims or internal operations must know who accepts control when the interaction becomes uncertain. The recipient must receive usable context, possess the necessary authority and respond within a timeframe suited to the risk.

The human should not be treated as an emergency attachment added after the automated experience has been designed. Human takeover is part of the product, part of the workflow and part of the customer’s service entitlement.

The defining question is therefore not whether the AI can handle the conversation. It is whether the organisation can still handle the customer when the AI cannot.

Design the Takeover Pathway Before the Agent Reaches Live Customers INTAKE • ESCALATE • ACCEPT • RESOLVE

Review escalation triggers, service ownership, approval limits, privacy, staffing, after-hours coverage and system records before voice or chat agents enter production.

Review Your Operational Handoff Model

Sources and References OpenAI: Introducing OpenAI Presence Elyment: Are GPT-Live Voice Agents Ready for Real Customer Calls? SafeWork NSW: Incident Notification ACCC: Consumer Rights and Guarantees OAIC: Privacy and Commercially Available AI Products NSW Legislation: Surveillance Devices Act 2007 Digital NSW: AI Assessment Framework Responsibilities Elyment: Why Bad Automations Continue to Cost Businesses Elyment: What Businesses Should Prepare Before Agents Move Into Production Elyment: Autonomous-Agent Containment and Incident Readiness