The Self-Serve AI Dream Is Dead. Good.

OpenAI started selling human engineers last week.

The product is called Presence, and it went into limited release on July 22. It deploys OpenAI's agents inside your company to handle real work: billing disputes, insurance claims, IT requests. But the agents don't show up alone. Each deployment is led by an OpenAI Forward Deployed Engineer and a systems integrator, run through a six-stage process of scoping, security review, testing, staged rollout, and iteration.

Sit with that for a second. The company that builds the most capable models on earth just told its biggest customers that the model isn't enough. You also need their people, embedded in your business, for weeks.

The self-serve enterprise AI dream is dead. And that's the best news you've gotten all year.

The Forward Deployed Engineer Is Back

"Forward Deployed Engineer" isn't a new idea. Palantir built a company on it: send your best engineers to sit inside the customer's operation, learn how the work actually happens, and build the thing on the ground instead of shipping software over the wall and hoping.

For a decade, that model was treated as a workaround, the expensive thing you did because the software wasn't good enough to stand on its own. Now it's the strategy. OpenAI is doing it with Presence. Coforge just launched an enterprise AI platform built around the same Forward Deployed Engineer model. The pattern is everywhere once you look, and it's spreading because it's the only thing that works.

The tools got better. The need for humans to implement them got bigger, not smaller.

And it's not a pilot for scrappy startups. OpenAI's first Presence design partners are BBVA, SoftBank, and IAG, some of the largest and most sophisticated companies on the planet. If they need engineers embedded to make agents work, the idea that you'll wire it up yourself from a help doc is a fantasy.

Self-Serve Was Always a Fantasy for Real Work

Here's the thing nobody selling you a subscription wants to admit. Buying access to an agent and pointing it at your business was never going to work, and the numbers say so. Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027.

That's not a model quality problem. Those projects don't die because the AI can't reason. They die because nobody scoped the work, cleaned the data, defined what success looked like, or built the guardrails. They die in the gap between capability and implementation, the same gap I've been pointing at all year.

Look at what OpenAI's own service actually is. Six stages. Security reviews. Staged rollout. Human engineers on the ground the whole way. That's not a product you buy. That's a project you staff. OpenAI knows the difference, which is why they stopped pretending a login was the whole answer.

Buy the Outcome, Not the Login

So change how you buy. Stop shopping for seats and start buying deployed outcomes.

When a vendor sells you an AI tool, the real question isn't how smart the model is. It's who is going to sit with your team, understand your actual workflow, and make the thing work in your environment. If the answer is "you figure it out from the docs," you're buying a project with no one to run it, and you'll be in the 40% that gets cancelled.

Budget for the humans. The implementation is not the tax you pay to get the AI. The implementation is the product. The model is just the part that comes in the box.

Staff for It

This also changes who you hire. The most valuable person in your AI strategy isn't a prompt engineer. It's the one who can sit between the business and the technology, understand a messy real-world process, and translate it into something an agent can actually execute safely.

Call it a Forward Deployed Engineer, call it an implementation lead, call it whatever you want. That role is the difference between an agent that resolves 75% of your tickets and a pilot that quietly gets shut off in Q3. Human and AI, working the problem together. Not one replacing the other.

OpenAI just built a business on the idea that the model needs people standing next to it. If the company that makes the model believes that, the question isn't whether you need implementation muscle.

It's whether you have any.