OpenAI introduced Presence on July 22. It lets companies build voice and chat agents that can answer customer questions, work inside business systems, take approved actions, and hand a case to a person.
Small businesses aren’t the target. Presence is available only through a limited program for eligible enterprise customers, with OpenAI staff and selected partners helping with deployment. I still think the setup is worth studying because it deals with the part of automation that demos tend to skip: what happens when a real customer is waiting.
What OpenAI launched
Presence is built for jobs such as billing support, insurance claims, outbound sales, and employee IT requests. A company starts by giving the agent one job and access to the information it needs. Staff decide which actions are allowed and where approval or a human handoff belongs.
Before customers use the agent, teams can test ordinary requests and then throw harder cases at it. Did it follow policy? Did it use the right tool? Did it hand the case over at the right moment? The testing continues after launch. Codex can suggest changes based on production sessions, but staff approve those changes before they go live.
OpenAI says Presence now handles its English-language phone support line and resolves 75% of inbound issues without human help. It also reports that its improvement process cut human handoffs by 15 percentage points in 10 days. Those are results from OpenAI’s own operation. I’d read them as evidence that the approach can work in one closely managed setting, not as a benchmark every company should expect to match.
Permissions, policies, and handoffs
A bot that only recites a return policy can’t do much damage. It can’t do much useful work, either. Once an agent can verify an account, look up an order, issue a credit, or change a service request, both the value and the cost of a mistake go up.
Presence doesn’t get a vague order to ‘handle customer support.’ It gets specific permissions and rules. The company tests it before it reaches a customer, and someone remains responsible for the cases it can’t handle.
CIO reported that many companies may start with lower automation rates than OpenAI’s 75% figure. Real organizations often have fragmented systems, uneven internal information, and complicated compliance requirements. The part I’d keep in mind is the cost of connecting everything safely: according to the article, integration and governance can cost more than the model usage itself.
Borrow the pattern for one small workflow
You don’t need Presence to borrow the discipline behind it. Before connecting an AI tool to even one small workflow, I’d write down five things:
- Start with one job. If the task takes more than one sentence to describe, narrow it. Drafting answers to shipping questions from an approved policy page is a job. ‘Run customer service’ is not.
- Then look at access. Give the tool only the documents and systems the job requires. A shipping assistant doesn’t need payroll files, tax records, or the full contents of your Drive.
- Keep drafts separate from actions. Let the agent prepare a reply, refund recommendation, or appointment change before it is allowed to send, pay, cancel, publish, or edit a customer account.
- Decide when a person takes over. A disputed charge, a policy exception, sensitive information, a mention of legal action, or a direct request for a human should end the automated part of the conversation.
- Finally, test what makes the workflow uncomfortable. Try a missing order number, conflicting policies, an angry message, a vague request, and instructions hidden inside uploaded text. Save the results and run the tests again after the workflow changes.
What happens when the easy cases disappear
When an agent handles the routine requests, the support job doesn’t disappear. The people on the team are left with the exceptions, judgment calls, and customers who are already frustrated. CIO’s reporting suggests the first workforce effect may be slower hiring in repetitive, high-volume support roles rather than an instant wave of layoffs.
That makes the remaining queue harder, so the automation rate can’t be the only score. I would also want to know how often the agent took the wrong action, whether customers had to contact the company again, and how much time employees spent untangling the difficult cases.
Customers still need an obvious way to reach a person. Employees need to see what the agent did and have the authority to correct it. Otherwise, a high automation rate can hide a bad support experience.
What to watch next
I want to see what OpenAI publishes next about pricing, eligibility, audit records, and results outside its own support line. The early customers named in the launch are testing banking, insurance, and multilingual support. That is useful to watch, but testing isn’t the same as a full public rollout.
Policy changes may be the harder test. Presence can suggest updates based on real sessions, but people approve each change. That matters. An agent shouldn’t quietly rewrite its operating rules because one angry customer found an unusual path through the system.
For a small business, the sensible starting point is fairly boring: one repeatable task and the smallest amount of access that makes it useful. Keep sending, payments, refunds, publishing, and account changes behind approval until your own testing gives you a reason to do otherwise. Write down when a person takes over. That is enough to find out whether the workflow actually helps.
Sources
OpenAI - Introducing OpenAI Presence
CIO - OpenAI Presence raises new questions about enterprise automation and jobs
Help Net Security - OpenAI Presence connects AI agents to enterprise data with built-in guardrails
