Enterprise AI agents: what gets your first one into production this quarter

I spent this week reading about billion-dollar commitments and kept landing on the same quiet realization. On June 30, AWS said it was putting $1 billion into a Forward Deployed Engineering unit that sends pods of its own engineers into customer offices to build AI systems on-site. Two days later, Microsoft announced its Frontier Company: $2.5 billion and roughly 6,000 embedded engineers doing the same thing. OpenAI and Anthropic had already stood up their own versions in May. Four of the most capable AI companies on earth just looked at the enterprise AI agents problem and decided the answer was people, not parameters.

That should change how a founder thinks about a first agent.

The gap between a working demo and a running agent

Here is the pain, stated plainly. The model works. It performed beautifully on the customer’s data in the pilot. Everyone in the room nodded. And then it sat there, because getting it to actually run inside the business meant wiring it into a CRM nobody fully understands, a data warehouse with three sources of truth, an identity provider, and an approval workflow that lives half in someone’s head.

That is not a capability problem. The model is fine. That is the last mile, and the last mile is where enterprise AI agents go to die quietly. The giants did the math and concluded the fix was worth billions in humans. The good news for a founder is that the fix they bought is not proprietary, and it does not require their budget.

What the billion-dollar embed teams are actually buying

Strip away the press releases and the forward-deployed model is a specific set of moves. AWS described its own approach, in VP Francessca Vasquez’s words, as “agentic-first,” compressing “timelines from months to days,” and structured “around shared goals and business results, not billable hours.” A one-person version of that is runnable this quarter.

1. Put one senior engineer inside the customer's actual environment

   Not a demo sandbox. Their Slack, their data, their approval chain. The whole reason the demo lied is that it ran somewhere clean. Give one strong engineer the mandate to sit where the real work happens and write the integration code by hand.

2. Pick one workflow narrow enough to finish and important enough to fund

   Not "transform support." One reversible, measurable task inside support. The 5% of pilots that reach production almost always started smaller than felt ambitious. Narrow is not timid. Narrow is how agents ship.

3. Set a 90-day pilot-to-production deadline and write it down

   The forward-deployed discipline is a clock. If integration eats six months, the agent never ships. A hard date forces the scope conversations that vague timelines let teams avoid.

4. Name one human owner who is accountable for the agent in production

   Not a committee. One name, on both sides: the engineer and the customer's operator who lives with the result. An agent nobody owns is an agent nobody ships, and an agent nobody can shut off.

5. Feed every reusable piece back into the core product

   The integration code, the eval harness, the change-management memo. This is what separates a forward-deployed engineer from a consultant: the first engagement is bespoke, the tenth is 70% reused. That is how deployment work becomes a product moat instead of a services trap.

Why the pilot stalled after the demo worked

Most teams read a stalled pilot as a signal to improve the model. So they swap in a better one, tune the prompts, and run the demo again. It works again. It stalls again. The mistake looks smart because the model genuinely did get better, and the demo genuinely does improve. But the loss was never about capability.

As IT Pro put it this week, forward-deployed engineers matter because “they sit inside the client, where they can drop barriers, cut through red tape, and get in front of the decision-makers quickly.” Read that again. Every noun in the sentence is organizational. Barriers, red tape, decision-makers. None of it is a benchmark. The reason the giants are hiring humans by the thousand is that the blocker was never in the model, it was in the building.

There is a second trap worth naming. This same week, Meta admitted its own agents are deploying slower than its $125 to $145 billion capital commitment implies they should. Even the company with near-infinite budget cannot buy its way past the last mile. Spending more does not move an agent from pilot to production. Owning the integration does.

The loss was never about capability. It happened in the meeting where nobody agreed on what "production-ready" means.

The numbers that should set the scope

The benchmarks are not subtle, and they all point the same way. MIT’s NANDA research found that 95% of enterprise generative AI pilots produce no measurable impact on profit and loss. Roughly 31% of enterprises have at least one agent in production, and Gartner expects 40% of agentic AI projects to be canceled by 2027 on cost, unclear ROI, and governance. Half of agentic projects, per this week’s coverage, are still stuck at the pilot stage.

What the forward-deployed model claims to buy against those odds is speed and proof. AWS points to embedded work that, in its telling, moved one real number:

Notice what “good” looks like there. It is one workflow, one number, one before-and-after. Not a platform, not a transformation. That is the bar to aim a first agent at. A win that cannot be expressed as a single measurable delta on a workflow a named person owns is not a production agent yet. It is another pilot wearing a nicer suit.

The Monday version of a forward-deployed engineer

Here is the move for Monday morning. Pick the one customer conversation that keeps almost closing and never quite does. Put the best available engineer inside that customer’s real environment for the quarter, not to sell, to build. Choose one workflow that can actually finish, attach one number to it, name the person on each side who owns the result, and set the ship date before a line of integration code gets written.

Keep one thing the giants will not hand over: a way out. Build the agent so it stays portable and so someone can turn it off in an afternoon. Meta’s own admission and the wider dependency risk this week are a reminder that a production system running on a model that can be repriced or retired by someone else is a bet, not an asset. Ship fast, own the integration, and keep the kill switch in reach.

The capability race is over for the purposes of a first agent. The model is good enough. The companies with the deepest pockets on earth just said, in billions of dollars, that shipping is the hard part. That is oddly reassuring, because shipping is the part a founder can actually control.