OpenAI Cannot Sell You an AI Agent Without Sending Engineers. That Is the Whole Story.

OpenAI launched an enterprise AI agent platform last month and said something quietly remarkable about it: you cannot buy it yourself.

Presence is not self-serve. Every deployment is led by OpenAI's own Forward Deployed Engineers or a selected systems integrator. It is sold as a project rather than a product, priced case by case, with no published rates. Each engagement starts from a single narrow job, resolving a billing dispute, handling an insurance claim, clearing an IT service request. BBVA Mexico, SoftBank and Retail Insurance Australia are the named launch customers.

Sit with that for a second. The company with arguably the best models on the planet, whose entire business is selling access to intelligence through an API, has concluded that for enterprise agents, shipping the model is not enough. They have to send people.

The admission inside the announcement

The implicit claim behind most AI marketing is that the model is the product. Better model, better outcome. Everything else is plumbing.

Presence is OpenAI conceding, in the most expensive way possible, that the plumbing is the product.

If the models were the hard part, OpenAI would win by simply having the best ones and letting customers connect them. Instead the difficult, expensive, human-intensive work turns out to be everything around the model: what your systems actually contain, what your process actually is, where the exceptions live, who is allowed to approve what, and what "resolved" means at your company specifically rather than in general.

That work cannot be shipped in a model release. It has to be discovered, at your company, by someone paying attention. Hence the engineers.

Forward deployed engineering is not new, and that matters

Palantir built a large business on exactly this shape, and the term comes from there. You do not sell software to an organisation that has messy data and undocumented processes. You send engineers in, they learn how the place actually works, and they build the last mile in situ.

It works. It also tells you something precise about the maturity of the technology: when a category still needs forward deployed engineers, the product is not finished. It is a capability plus a consulting practice that turns the capability into an outcome.

There is nothing dishonest about this. It is the rational thing to do when the integration work genuinely is that hard, and OpenAI is being unusually straight about it rather than pretending a self-serve product exists when it does not. Plenty of vendors sell you the demo and let you discover the gap.

But it does have a consequence, and the consequence is the whole point of this piece.

It does not scale down

Forward deployed engineering has a floor. There is a contract size below which sending humans is not worth it, and that floor is high. BBVA, SoftBank and a national insurer sit comfortably above it. A forty person logistics company does not.

So the integration work does not disappear for smaller businesses. Somebody still has to determine which systems the agent touches, what it is allowed to decide, where it escalates, what your exceptions are. That work is identical whether your company has forty people or forty thousand. What differs is who does it, and whether anyone will be paid enough to bother.

For most businesses the honest answer today is: you do it. You become your own forward deployed engineer, in the evenings, without the training.

That is the actual gap in the market. Not intelligence, which is abundant and cheap and improving on someone else's budget. The gap is the last mile, and right now it is priced as a consulting engagement or handed to the customer as homework.

The part of their playbook worth copying for free

Buried in the model is a piece of advice that costs nothing to take.

Every Presence engagement starts with one job. Not "transform customer service." One thing: this dispute type, this claim category, this ticket queue. A single workflow with a definable start, a definable finish, and a measurable outcome.

That is not a limitation of the technology. It is the thing that makes deployments work, and it lines up exactly with the ROI research: deployments that replaced a specific named process with a named owner return dramatically more than deployments that added a general capability nobody was accountable for. Same tools, wildly different results, and scope discipline is most of the difference.

If you are starting anywhere with agents, copy that. Pick one job that somebody currently does, that recurs, that has a clear finish line, and hand over that job completely. Do not buy a capability and go looking for uses.

What to do if nobody is sending you engineers

Three options, in ascending order of how much I would recommend them.

Hire the help. Systems integrators are building agent practices quickly. This works and it is expensive, and you should only do it for a workflow whose value you can already quantify, because you will be asked to justify it.

Do it yourself with general tooling. Feasible, and much slower than it looks. The trap is that the first workflow feels straightforward and the next five reveal that your processes were never written down. Budget for discovery, not just build.

Use something where the agent does its own integration. This is the direction we bet on. If an agent can be told about your business in plain language, given access to your actual tools, taught a new skill on request and made to remember all of it, then the last mile stops being a professional services engagement and becomes a conversation. That is not a claim that it is effortless. It is a claim about who does the work, and how much of it survives after they leave.

Because there is one more thing about the forward deployed model worth noticing: when the engagement ends, the engineers go. What they learned about your business lives in the system they built, if you are lucky, and in their heads, if you are not. An agent that accumulates that context itself, and keeps it, is solving a different version of the problem than one that had it installed by a visitor.

The short version

OpenAI just told the market that enterprise agents currently require humans in the loop of the deployment, not just the operation. That is worth believing, because it is against their commercial interest to say it and they said it anyway.

The follow-up question is the one nobody at that scale is asking: what happens to everybody below the threshold where sending engineers makes economic sense? Because that is most companies, and the work does not get any easier for them. It just gets less attended to.

Want to test the most advanced AI employees, without a deployment team attached? Try it here: https://geta.team