AI Customer Service for Small Business: Steal the Blueprint

OpenAI shipped its most complete customer service product of the year on July 22. There is no signup page, no listed price, and no free trial. You get it by having OpenAI send its own engineers to your company.

The short version: OpenAI Presence runs voice and chat support agents that resolve 75% of inbound issues without a human, and no small business can buy it. That matters less than it sounds, because the model was never the hard part. AI customer service for small business succeeds or fails on a wrapper of written policy, permission limits, and escalation rules, and every one of those is a configuration decision rather than a purchase. The expensive part of Presence is the part you can copy for free.

What is OpenAI Presence?

Presence is an enterprise platform for running customer-facing voice and chat agents in production, announced on July 22, 2026. The important thing to understand is that it is not a model. It is scaffolding built around one.

Per reporting from AI News, the platform bundles four things: a company’s own policies and standard operating procedures, permission controls that limit which data and systems an agent may touch, escalation rules that route a conversation to a human when it should, and pre-deployment simulations that test an agent before it ever speaks to a real customer.

Three early customers are named: BBVA Mexico for banking support, connected to account systems so agents can resolve issues rather than just describe them; SoftBank Corp., running Japanese-language voice agents; and Retail Insurance Australia, part of IAG. OpenAI reports that on its own English-language support line, Presence resolves 75% of inbound issues without human assistance.

Why can’t a small business buy OpenAI Presence?

Because it is not sold, it is delivered. Presence launched through a limited general availability program rather than a self-serve dashboard. Deployments are led by OpenAI’s Forward Deployed Engineers or a short list of approved global systems integrators, and as CMSWire noted in its coverage, pricing has not been disclosed. None of the named customers had reached full production at announcement.

It would be easy to read that as a snub. Read it as a signal instead.

OpenAI has the best distribution in the industry and an obvious incentive to make this self-serve. It chose not to. When the company holding the models decides that customer-facing support needs human engineers on site to configure, the conclusion is not that the technology is half-baked. It is that deployments fail in the wrapper rather than in the model. Those engineers are expensive because the work they do is the work that matters.

What actually made Presence work?

Strip away the enterprise packaging and Presence is four disciplines, none of which are proprietary:

Written policy. The agent is given the company’s actual rules and procedures as source material, not a vague instruction to be helpful.

Scoped permissions. The agent can reach specific systems and specific data, and nothing else.

Explicit escalation. The conditions for handing a conversation to a person are decided in advance, in writing, rather than left to the model’s judgment in the moment.

Testing before launch. The agent is simulated against realistic scenarios before a customer is exposed to it.

Notice what is missing from that list: a better model. Every item is something a two-person shop can do this month, and none of them require a procurement department. The reason most small business AI support experiments feel flimsy is not that the underlying model is weaker than the one BBVA is using. It is usually the same model. The difference is that nobody wrote the four things down.

How do you build AI customer service for a small business without an enterprise contract?

Mirror the four disciplines with tools you can already afford. Our roundup of the best AI tools for small business, organized by the job you need done, covers the website chat and missed-call categories in detail, but the sequence matters more than the vendor.

Write the policy before you pick the tool. Open a document and answer the questions your customers actually ask: your hours, your service area, your refund terms, your lead times, what you do and do not take on. Most small businesses have never written this down, which is precisely why the AI improvises. An hour here beats a month of tuning prompts.

Decide what it may reach. Read-only access to your FAQ and hours is a different risk profile than access to your booking calendar, which is different again from anything that can issue a refund. We covered the questions worth asking before giving any AI tool more access, and they apply directly here.

Define the handoff in advance. Pick the categories that always go to a person: complaints, anything involving money moving, anything about a job already in progress, anything where the customer sounds upset. Write them into the configuration rather than hoping the model reads the room.

Test it against your own history. You already have a simulation set. It is your last fifty emails, voicemails, or chat transcripts. Run them through before you point the thing at a live customer.

Is 75% a number you should expect?

No, and it is worth being clear about why. The figure is self-reported, and it comes from OpenAI’s own English-language support line: a technical customer base, a product documented in exhaustive public detail, and a high volume of repeat questions with stable answers. CMSWire also flagged that the metric is undefined, meaning we do not know what counts as resolved.

Your inbound mix is different. A landscaping company gets “can you come look at my yard Thursday,” which is not a knowledge question at all, it is a scheduling and judgment question. Treat 75% as evidence that the ceiling is higher than most owners assume, not as a forecast for your shop.

The part worth being optimistic about

The enterprise framing around this launch is deflection: fewer contacts reaching human agents, measured as cost avoided. That math does not transfer to a business with one person answering the phone, and it is the wrong thing to chase.

The small business version is not about reducing headcount, because there is no support department to reduce. It is about the inquiries that currently go nowhere. The second caller while you are already on the phone. The form filled out at 9pm on a Saturday that you answer Monday morning, by which time they have booked someone else. Those are not costs being cut, they are revenue that was already being lost.

And the escalation rule is the whole point of the design. Configured properly, the AI takes the repetitive intake and hands you the conversations that need a person, which is the opposite of what the word automation usually implies. Voice is getting good enough for this quickly, as we covered when ChatGPT’s voice model learned to handle interruptions. The constraint now is not capability. It is whether anyone bothered to write down the rules.

Frequently Asked Questions

Can small businesses use OpenAI Presence?

Not currently. Presence launched on July 22, 2026 through a limited general availability program with no self-serve signup, and deployments are handled by OpenAI’s Forward Deployed Engineers or approved global systems integrators. Pricing has not been published. Small businesses should instead apply the same configuration approach using standard website chat, helpdesk, or missed-call tools they can buy directly.

What does OpenAI Presence actually do?

It runs customer-facing voice and chat support agents in production, wrapping an AI model in a company’s written policies and standard operating procedures, permission controls limiting what data and systems the agent can access, escalation rules that route conversations to humans, and simulation testing before deployment. It is a governance and deployment layer rather than a new model.

Does AI customer service replace your support staff?

In a small business it generally cannot, because there is rarely a dedicated support role to replace. The practical gain is capturing inquiries that currently go unanswered, such as after-hours form submissions, second callers during a busy period, and repeat routine questions. Escalation rules are designed to send anything involving judgment, money, or an unhappy customer to a person.

Is the 75% resolution rate realistic for my business?

Treat it as a ceiling rather than a forecast. The figure is self-reported by OpenAI, measured on its own English-language support line, and the term resolved has not been defined publicly. That environment features a technical customer base and heavily documented products, so businesses with more scheduling, pricing, or judgment-based inquiries should expect a lower rate.

What share of your inbound questions are genuinely the same five questions over and over, and what share actually need you? We would like to know where the line falls in your business.