OpenAI's Presence resolves 75% of queries with no human. See why enterprise AI agents may perform differently once deployed beyond OpenAI's own systems.
OpenAI launched Presence on 23 July 2026, a service that lets enterprises deploy voice and chat agents to handle customer and employee requests, including operating IT systems and taking approved actions without waiting on a person. The company disclosed that its own English-language phone support channel now resolves 75% of inbound issues through Presence with no human involved, a number Computerworld reported is already prompting analysts to question what enterprise AI agents can realistically deliver once they leave OpenAI's own environment.
Presence is not a plug-and-play product. Enterprises must join a limited availability programme, with OpenAI or selected global systems integrators handling the deployment. Each installation is scoped to one task type — billing queries, insurance claims, IT service requests — with the agent given only the data and system access that task requires. OpenAI is pairing Presence with Codex, a second tool that monitors agent performance and suggests process changes; in internal testing, Codex-driven suggestions cut handoffs to human staff by 15 percentage points over ten days.
Who is actually testing this and for what
Three named enterprises are piloting Presence, and their use cases show how differently the same platform can be applied. Spanish bank BBVA is evaluating it for everyday banking support in Mexico. Japanese technology group SoftBank is trialling Japanese-language customer interactions. Australian insurer IAG is assessing whether Presence can help it manage spikes in customer contact during severe weather events — a scenario where call volume swings sharply and staffing to peak demand is expensive.
That last case is the one most relevant to Australian mid-market operators watching this space. IAG's interest is not about replacing a steady-state support desk. It is about elastic capacity during a flood or storm event, when inbound volume can multiply overnight and a fixed headcount either sits idle most of the year or buckles during the surge. Whether Presence can hold up under that kind of variable load, rather than the steady, familiar traffic of OpenAI's own support line, is precisely the question analysts are raising.
Why 75% is a number to interrogate, not copy
Pareekh Jain, CEO of Pareekh Consulting, told Computerworld that CIOs should treat OpenAI's 75% resolution rate as proof the technology can work, not as a benchmark every enterprise can expect to hit. OpenAI's own deployment benefits from being built around its own products and its own data — a best-case environment. Jain said larger enterprises typically contend with fragmented legacy systems, uneven knowledge bases, and heavier compliance obligations, all of which tend to push initial automation rates lower before they improve with refinement.
That distinction matters for any business assessing enterprise AI agents against its own case load. A support desk running on a single, well-documented CRM with clean data is a different proposition to one spanning three systems accumulated through acquisitions, inconsistent naming conventions, and a knowledge base nobody has updated since 2021. The technology's ceiling is set less by the model and more by the quality of what it is connected to and told to follow — a pattern consistent with how document workflows determine whether AI can be trusted with a process at all, once codified procedures rather than tribal knowledge sit behind the agent.
The jobs question OpenAI didn't answer
OpenAI's announcement made no mention of the effect Presence could have on staffing, but the 75% figure invites the question anyway. Tulika Sheel, senior vice president at Kadence International, said the first workforce effect is more likely to show up as slower hiring than as immediate layoffs. The roles most exposed, she said, are repetitive, high-volume functions such as frontline customer support and routine back-office processing — work that follows a predictable script rather than requiring judgement.
Jain agreed that Tier-1 support agents handling predictable queries face the most exposure, but argued broader headcount reductions would only follow once companies reorganise operations around the technology rather than simply bolting it onto existing structures. Lian Jye Su, chief analyst at Omdia, offered a more measured view: Presence is unlikely to sharply increase displacement risk because enterprises have used comparable customer-support automation from vendors including Genesys, NiCE, Five9 and AWS for years already. Su expects Presence to sit alongside employees rather than replace them outright, with AI absorbing routine volume while people handle the cases that require empathy or judgement calls the agent isn't authorised to make.
The cost line CIOs are watching
Analysts flagged a second, less visible risk: as fewer cases get handed to humans, the cases that do reach a person may skew toward the hardest, most complex ones — a shift that could strain remaining staff even as headline resolution numbers look good. Sheel put the real test plainly: whether Presence can hold resolution quality as usage scales, not simply how many tasks it can technically handle.
Jain pointed to a cost structure that often gets underestimated in enterprise AI agents projects. "Often the biggest cost of enterprise AI is not tokens but integration and governance," he said, noting that connecting Presence to existing systems, and maintaining the oversight needed to audit what agents actually did, can offset early savings. Enterprises will need to decide what systems and data each agent can touch, monitor its behaviour continuously, and keep an audit trail of every action — work that sits closer to systems architecture than to prompt-writing. Su added that enterprise IT complexity will likely prevent OpenAI from automating entire workflows unassisted; CIOs will still need other technology providers and human staff in the loop, and will favour systems that can be integrated and audited rather than treated as a black box.
Diagnostics
Operations leaders assessing whether enterprise AI agents like Presence fit their own support function are asking themselves three questions:
Is our support caseload closer to OpenAI's own single-system environment, or to the fragmented, multi-system reality Jain describes — and which systems would an agent actually need to touch to resolve a case end to end?
If routine cases were resolved without a human, would the cases still reaching our staff be harder on average, and do we have a plan for that shift in workload composition?
Have we costed the integration and governance work — system access, monitoring, audit trails — separately from the AI itself, given Jain's warning that this is where the real expense sits?
