I keep hearing a confident prediction. Once the models get good enough, the people who deploy them go away. The Forward Deployed Engineer becomes a rounding error. The model does it all.

I think that is exactly backward. And the data is starting to agree.

Start with what an FDE actually is. Palantir created the role and called it the Delta. A normal engineer builds one capability for many customers. A Delta takes many capabilities and makes them work for one customer, inside that customer's messy reality, on their infrastructure, against their real problem. Palantir's own job posting compares the work to a startup CTO. This is not support. It is deployment, judgment, and ownership of an outcome.

Now hold that next to the number everyone in enterprise AI is quietly staring at. An MIT NANDA report this year found that about 95 percent of enterprise generative AI pilots deliver no measurable impact on the bottom line. Read the reason carefully, because it is the whole argument. The failures were not about the models being too weak. They were about the learning gap. The tools did not fit the workflow. The deployment was flawed. The same report found that buying and partnering succeeded around 67 percent of the time, while building alone worked about a third as often.

So the bottleneck is not intelligence. It is the last mile. It always was.

Here is the part the "models will replace us" crowd misses. When intelligence gets cheaper and better, it does not remove the last mile. It moves all the differentiation into it. If every company can call the same frontier model, the model is no longer the edge. The edge is who can wire that intelligence into a specific business, with its specific data, its specific rules, and its specific risk. A better model does not shrink that work. It raises the ceiling on it. The smarter the engine, the more a skilled deployer can build on top of it, and the more damage a careless one can do.

The market is voting with real money. OpenAI stood up a Forward Deployed Engineering team, then went further and launched a deployment company with more than 4 billion dollars behind it, acquiring roughly 150 forward deployed engineers in a single move. The venture firm a16z called the FDE the hottest job in tech. These are the people sitting closest to the frontier models on earth. If better models made deployers obsolete, they would be cutting them. They are doing the opposite.

I have lived on both sides of this. I have been a CEO, a CIO, and a CISO five times. I have also shipped production AI systems myself, with Claude, GPT, and Gemini, wired to real payment and data infrastructure. So I can tell you where pilots die. They die in the gap between a demo that impresses a room and a system that survives a Tuesday. They die on data nobody cleaned, permissions nobody scoped, and edge cases nobody owned. No model update closes that gap. A person does.

There is a second job that gets bigger as the models get better, and it is the one most people skip. Assessment and audit. Before you deploy AI into a business, someone has to look hard at where that business actually stands, what data it can use, what it is exposed to, and what a failure would cost. The stronger the model, the higher the stakes of pointing it at the wrong thing. Governance is catching up to this. NIST published an AI Risk Management Framework. ISO published a management standard for AI. The EU AI Act now requires conformity assessments for high risk systems. This is the same assessment and audit discipline cybersecurity has run for years, arriving in AI. It does not get automated away by a smarter model. It gets more important.

Let me give the other side its due. There is a real version of the replacement argument. Better models and better tooling do absorb the low end of deployment. The simple integration, the boilerplate glue, the first draft of a connector. That work is already shrinking, and it should. But that is not the FDE job. That is the part of the FDE job a good FDE was always trying to automate anyway. What is left when you automate the easy part is the hard part. Judgment under ambiguity. Trust with an executive who has to sign the check. The call about what to ship, what to hold, and what to never let the model touch. Those do not commoditize.

So here is the honest conclusion. Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, on cost, unclear value, and weak controls. Those cancellations are not model failures. They are deployment and judgment failures. Every one of them is a job for someone who can stand in the room, read the business, wire the intelligence in correctly, and own the result.

Better models do not end that job. They are the reason it becomes one of the most valuable jobs in the company.

If you are deploying AI and your pilots keep dying in the last mile, that is not a model problem. Let's talk about the part that actually moves risk, cost, and advantage.

Sources