A Hong Kong insurer's claims hotline takes 4,000 calls a week. The operations director has been told AI can handle the first two minutes of every one of them. Six vendors have pitched. Every deck shows the same demo, and not one of them has explained what happens on the call where the AI gets it wrong.
On 22 July 2026, OpenAI launched a product built specifically around that question. It is called Presence, and it is not an API.
What is OpenAI Presence?
OpenAI Presence is an enterprise product for deploying voice and chat AI agents into production workflows. It combines model reasoning with company policies, guardrails, approved actions and escalation rules. Deployments are led by OpenAI Forward Deployed Engineers rather than self-serve signup, and each agent is scoped to one specific job.
The distinction matters more than it sounds. An API gives you a model and leaves the control layer to you. Presence ships the control layer as the product.
According to OpenAI's launch announcement, the reasoning was explicit: the challenge for enterprises is no longer proving that AI agents can work, it is making them reliable enough to do high-value work in production.
Why did OpenAI build a deployed product instead of another model release?
Because the failure rate of agent projects is a deployment problem, not a model problem. Gartner's June 2025 poll of more than 3,400 organisations investing in agentic AI concluded that over 40% of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls.
None of those three causes is fixed by a smarter model.
Gartner's 2026 Hype Cycle for Agentic AI places the category at the Peak of Inflated Expectations, with only 17% of organisations having deployed AI agents while more than 60% expect to do so within two years. That gap between intent and deployment is the market Presence is aimed at.
Gartner has also forecast that during 2026, one third of companies will damage customer experience by deploying AI prematurely, eroding brand trust and hurting both acquisition and retention. For a business leader, that is the real risk line. A failed internal pilot costs budget. A failed customer-facing agent costs customers.
How does a Presence deployment actually work?
Each deployment starts with one specific job, such as resolving billing issues, supporting insurance claims or handling employee IT service requests. The agent receives only the knowledge and system access that job requires. The company then sets policies covering what the agent may do, when it needs approval, and when a human takes over.
The sequence OpenAI describes has five stages, and every one of them is an organisational task rather than a technical one.
--- Scope the workflow. One job, defined narrowly enough that success is measurable.
--- Connect knowledge and systems. Only what the job needs, nothing more.
--- Set permissions and policies. Approved actions, approval thresholds, escalation triggers.
--- Test before launch. Simulations and graders check whether the agent reached the right outcome, followed policy, used tools correctly and escalated appropriately.
--- Improve after launch. Production sessions and escalations surface gaps, and Codex proposes updates that teams test against the live version before approving a controlled rollout.
Read that list again as a job description. Four of the five stages belong to your operations, compliance and service teams. Only one belongs to the model.
What does the 75% resolution figure actually tell you?
Presence powers OpenAI's own English-language phone support line, where it now resolves 75% of inbound issues without human assistance. Its Codex-powered improvement loop cut human handoffs by 15 percentage points in 10 days. Those are real published numbers, and they are also the best case: OpenAI running OpenAI's product on OpenAI's own support queue.
Treat it as a ceiling, not a forecast.
The transferable insight is the second number, not the first. A 15-point handoff reduction in 10 days tells you the improvement loop is where the value compounds. The launch-day accuracy is a starting position. The rate at which you can close gaps after launch is the actual capability.
That reframes the vendor question. Instead of asking how accurate the agent is on day one, ask how fast the platform lets you fix what you find in week three.
Who can actually buy OpenAI Presence today?
Presence is available to eligible enterprise customers through a limited general availability programme only. Deployments are led by OpenAI Forward Deployed Engineers and selected global systems integrators. OpenAI states plainly that Presence is not yet available as a self-serve product, and organisations must approach their OpenAI account team.
For most Hong Kong mid-market companies, that is a closed door for now.
The named design partners give a sense of the tier being served. BBVA is exploring AI voice support for everyday banking in Mexico. SoftBank is testing Japanese-language customer conversations. IAG is looking at support during severe weather events and natural disasters. These are multinationals with dedicated AI transformation leadership.
A 200-person professional services firm in Central is not the target customer of a Forward Deployed Engineer programme. Understanding that early saves a quarter of wasted vendor conversations.
What does this mean for a Hong Kong enterprise that cannot access it?
The architecture is more valuable than the access. Presence is essentially a published reference design for production agent deployment, and every component in it can be specified into a tender, a build project or a local vendor evaluation regardless of which platform you end up using.
Four requirements are worth lifting directly into your own criteria.
--- Job-scoped access. The agent sees only the data and systems its single job requires. This is also the cheapest way to keep an agent deployment defensible under the Personal Data (Privacy) Ordinance.
--- Explicit approval thresholds. Written rules for which actions the agent may take alone and which need a human sign-off, agreed before launch rather than after an incident.
--- Pre-launch simulation. Testing against edge cases and higher-risk scenarios, graded on outcome, policy compliance, tool use and escalation behaviour.
--- A named post-launch improvement owner. Someone accountable for reviewing escalations weekly and approving changes.
Local context sharpens the third and fourth points. The Privacy Commissioner for Personal Data completed compliance checks on 60 organisations in Hong Kong, with results published in May 2026 showing that 95% used AI in day-to-day operations and over half ran three or more AI systems. The PCPD found no PDPO contravention, but noted that several governance measures appeared to be weakening rather than strengthening as adoption spread.
Adoption is outrunning governance. An agent that takes approved actions on customer accounts is precisely where that gap becomes expensive. If you have not yet worked out how each agent is identified and authorised inside your systems, that question is worth resolving first, and we have covered it separately in our explainer on agent identity.
How is a deployed agent platform different from a chatbot or an in-house build?
Three delivery models now compete for the same budget line. A subscription chatbot answers questions from a knowledge base. A deployed agent platform takes approved actions inside your systems. An in-house build gives you full control and full responsibility for the control layer.
The choice turns on one question: does the workflow require the AI to do something, or only to say something?
Answering is a knowledge problem. If the job is to tell a customer what your refund policy is, you need accurate retrieval, multilingual coverage and honest handling of what the system does not know. A well-configured subscription chatbot is the proportionate answer, and it is priced accordingly.
Acting is a governance problem. If the job is to verify a caller, look up an account, apply policy and issue the refund, you have entered a different risk category. Now you need approved-action lists, approval thresholds, audit trails and escalation rules. This is the territory Presence occupies, and the cost reflects the governance scaffolding rather than the model itself.
Building in-house buys control and inherits the hard part. Teams that build often underestimate the simulation and evaluation layer, which is unglamorous, expensive and the single most common thing missing when a pilot stalls at 60% accuracy with no route to 80%.
Most organisations discover they need the first model for roughly 70% of enquiries and the second for the remaining 30%. Deciding that split before you shortlist vendors is worth more than any feature comparison, because it determines whether you are buying a subscription or commissioning a programme.
Where do agent deployments go wrong in practice?
Three failure patterns account for most of the cancelled projects Gartner describes, and all three are visible before launch if you know what to look for.
Scope drift. The pilot starts as "handle billing enquiries" and becomes "handle customer service". A narrow job with a measurable resolution rate turns into an unbounded remit with no pass mark. Presence resists this by design, one job per deployment. Most internal pilots do not.
No escalation contract. Teams define what the agent can do and forget to define when it must stop. The agent then either escalates constantly, which destroys the cost case, or escalates too rarely, which destroys the trust case.
No owner after go-live. The project team disbands at launch. Escalation logs accumulate unread. By month four, the agent is answering against policies that changed in month two. This is the quiet killer, and it is why an improvement loop matters more than launch-day accuracy.
A fourth pattern is worth naming for boards specifically. Agents that behave correctly in testing can behave differently when goals conflict under pressure, a risk we examined in our piece on agentic misalignment.
What should you ask before signing with any agent platform vendor?
Five questions separate a platform that will survive production from a demo that will not. Ask them in this order, and require the answers in writing.
--- What single job will this agent do, and what resolution rate counts as success? If the vendor cannot help you write a pass mark, you are buying a pilot.
--- Which actions can the agent take without human approval, and where is that list stored? The list should be a governed document, not a conversation.
--- How do I test a policy change before it reaches customers? Simulation capability is now a baseline requirement, not a premium feature.
--- What does the improvement loop look like, and who runs it? Ask for the actual cadence and the actual person.
--- Where does customer data sit, and who can access it? Under the PDPO, this answer belongs to you regardless of which vendor holds the infrastructure.
Notice that none of these five questions is about model quality. In 2026, model quality is table stakes. Deployment discipline is the differentiator.
The strategic takeaway
OpenAI Presence is worth understanding not because most Hong Kong enterprises can buy it this quarter, but because it settles an argument. The hard part of enterprise AI agents was never the reasoning. It was the policies, the escalation rules, the pre-launch simulation and the person who owns the improvement loop in month six.
That is an operating-model problem, and operating-model problems do not get solved by waiting for the next model release. They get solved by defining one job, writing down the rules, testing before launch and naming an owner. Any organisation can start that work this month, on any platform, at any budget.
The technology will keep moving. The discipline is what compounds. We understand AI. We understand you. With UD by your side, AI never feels cold.
Reviewed by the UD enterprise AI team.
Start with a readiness view, not a vendor list
Before you evaluate any agent platform, it helps to know which of your workflows are actually ready to be scoped, governed and measured. We'll walk you through every step, from AI readiness assessment to workflow scoping, policy design, deployment and performance tracking, backed by 28 years serving Hong Kong enterprises.
