What Is Forward-Deployed Engineering and Why Do AI Labs Use It?

What is forward-deployed engineering?

Forward-deployed engineering (FDE) means placing engineers inside, or right alongside, a customer's organization to turn purchased AI into working systems. They find where AI applies, then design, build, test and deploy it against the customer's own data and processes. The goal is not a demo. It is a payoff in everyday operations. AI labs now use this model because buying model access does not mean a company knows how to earn a return from it.

Why do companies struggle to get ROI from AI?

Procurement teams tend to get stuck on how to pay for AI. Per seat is the traditional SaaS measure. Per token meters compute. Neither says anything about whether the work done was valuable.

Tokens are useful units, and they help a finance team budget. But a token does not represent a fixed amount of useful thinking. Whether it was useful depends on the people using it. The usual contract leaves that assumption with the buyer: capture enough value to justify the cost.

General-purpose intelligence adds a second problem. Much of its value has to be discovered by the human using it. A company can buy the capability before it has learned which work to assign, how to change the surrounding processes, or how to tell when an answer is good. A model can look impressive in a demo while the organization stays poorly equipped to use it. That missing know-how explains a large share of weak AI returns.

The result is cautious spending. Many companies cap budgets because they have not worked out how to capture ROI at scale. Forward-deployed engineering is one answer to that specific gap.

How does forward-deployed engineering work in practice?

The engagement follows a rough sequence:

1. Choose where intelligence applies. Engineers work with business leaders and frontline teams to pick the opportunities worth pursuing.
2. Design and build. They create systems connected to the customer's data and processes.
3. Test and deploy. The system goes into production, not a sandbox.
4. Make it pay off. The work continues until results show up in daily operations.

Consider the kind of example a hypothetical equipment-maintenance company illustrates. An engineer spots a recurring failure customers keep reporting. An agent pulls maintenance records and compares prior cases. The engineer knows which operating conditions matter, which records are unreliable, and when an apparent pattern is really two different faults described in similar language. That judgment shapes which comparisons to request and which evidence to check.

Eventually the engineer must decide when to keep digging with the AI and when to talk to a person. Sending an agent off to generate more explanations does not fix a customer's problem. Repeat that across enough cases and the team may see a product: ongoing advice that helps customers prevent failures. Someone still has to price and package it, but the opportunity did not exist before.

What is OpenAI doing with forward-deployed engineering?

OpenAI has a deployment company that describes its engagements in exactly these terms. Its engineers work with business leaders and frontline teams to find opportunities, then design, build, test and deploy systems connected to the customer's data and processes.

OpenAI also took a deeper step with Thrive Holdings. It took an ownership stake and announced it would embed research, product and engineering teams inside Thrive's companies. The initial focus is accounting and IT services. How far it goes from there is still open.

The stake matters because it changes the economics. An AI supplier with real expertise in a domain can deliver value that goes well beyond what token pricing captures. When the supplier owns part of the upside, both sides are committed to making the intelligence productive.

What is Anthropic's joint venture?

Anthropic uses similar language around forward-deployed engineering. It announced a joint venture with Blackstone, Hellman & Friedman and Goldman Sachs. The venture is working with businesses like mid-size manufacturers and regional healthcare systems, the kind of businesses that may lack in-house AI teams but have plenty of processes to improve.

The source material does not detail the venture's structure, funding or timeline, so treat those as unknown. What is clear is the direction: financial backers and a lab are pairing capital with engineers who apply AI inside operating businesses.

Does outcome pricing replace per-token pricing?

Not yet, and probably not entirely. Outcome pricing is hard to measure, and no one is claiming the market will pivot overnight. Token pricing still helps buyers budget.

But the idea is gaining ground. Cedar, a healthcare payments company, runs an agent called Cora that answers patient calls about medical bills, confirms insurance and lets the patient pay on the call. The hospital does not care how many tokens that call used. It cares that the bill got paid. Sierra, a vendor working with Cedar, is exploring per-outcome pricing in deals where both seller and buyer are comfortable with it.

For a CFO, a deeper partnership changes the question. It stops being "what software discount can procurement win?" and becomes "how much more value could this create, what would we retain, what must we commit, and is the upside worth it?"

Who is the "manager of intelligence"?

Traditional economic inputs, people and capital, are managed from the top. An officer assigns a team and a budget, and the work follows. AI does not behave that way. Frontline engineers and customer-facing staff decide when to use AI and when not to, so they allocate intelligence every day, and those decisions carry real economic weight.

That makes everyone a manager of intelligence to some degree. The shift looks like this: instead of asking an agent to finish a regular report faster, you ask it to investigate the root cause of the customer problem behind the report and start fixing it. You move from user to manager.

It also makes motivation a practical concern. People need to feel equipped, supported and aligned with the mission, and to have some stake in the result. A motivated team applies AI better.

Should companies rely on forward-deployed engineering forever?

No. Keeping FDE as a service indefinitely means the company probably is not learning enough. Pushing a half-understood solution as a product too early has the same problem. The balance is sharp, empowered people on the front line who talk to customers, learn what works, and eventually shape that learning into something repeatable. Some opportunities grow large enough to justify a deeper relationship, such as the Thrive stake or a joint venture.