The last mile: what Anthropic's Forward Deployed Engineer reveals about the AI deployment gap

In May 2026, Anthropic announced a $1.5 billion joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman to launch an enterprise AI services firm. The stated reason, in Anthropic CFO Krishna Rao's words: "Enterprise demand for Claude is significantly outpacing any single delivery model."

OpenAI responded within weeks with its own deployment company, backed at $14 billion.

Read those two facts together. The two most capable AI labs on the planet have simultaneously decided that the bottleneck to enterprise AI adoption is not the model. It is the human presence required to get the model working inside a real organisation. Both companies are now, in effect, selling human deployment as a service.

That is a significant admission. These companies promised autonomous AI. What they are delivering, at scale, is embedded engineers. A few weeks before the Blackstone announcement, Anthropic posted a job listing for a role called Forward Deployed Engineer. The job posting is the small version of the same confession.

"Enterprise demand for Claude is significantly outpacing any single delivery model."

Krishna Rao, CFO, Anthropic

What the posting actually says

The job responsibilities, stripped of marketing language:

• Work inside customer systems to build production applications using Claude.
• Deliver technical artifacts: MCP servers, sub-agents, agent skills.
• Provide white-glove deployment support inside enterprise environments.
• Identify repeatable deployment patterns and feed them back to product and engineering.
• Travel to customer sites. Work alongside their teams. In person.

The salary range is $200,000 to $300,000. One of the preferred qualities listed is "high agency with an ability to navigate ambiguity present in complex organisations."

That last phrase is doing a lot of work. It is not describing a technical requirement. It is describing someone who can walk into an organisation they have never seen before, figure out where the bodies are buried, and ship something real anyway. That is a different skill from writing good code, and most people who write good code do not have it.

The deliverable is a running production deployment. Not a proof of concept. Not a pilot. Something that runs inside a real system, under real load, that real people depend on.

Where the term comes from

Palantir coined the role in the early 2010s, originally calling it "Delta." Until around 2016, the company employed more Forward Deployed Engineers than traditional software engineers. That ratio is not an accident. It reflects a specific thesis about what kind of product Palantir was actually selling.

The thesis: their software only worked when a capable engineer sat inside the customer's environment long enough to learn its constraints. Classification rules, legacy data schemas, security boundaries, internal politics. None of it appears in a sandbox. All of it determines whether the product ships.

Palantir's insight was that software plus embedded human was the unit of sale. Software alone did not work. The FDE was not a support function; the FDE was part of the product.

Anthropic has arrived at the same conclusion. A16z called the FDE "the hottest job in tech" in 2025. OpenAI now employs FDEs across more than eight cities. The role is spreading because every AI-native company selling into enterprise is hitting the same wall: the model works in the demo, and then the real work begins.

The four stages: and where things break

Most organisations think about AI adoption as a progression. They start at stage one and expect the path to stage four to be a matter of time and budget. It is not.

1. Chat window. A user opens Claude or ChatGPT in a browser and asks questions. Productivity is real but personal. Nothing is integrated with anything else.

2. AI-enabled workspace. The model has access to files, tools, and user identity. Context persists across sessions. Output starts to compound.

3. Agentic workflows. Models execute multi-step work using tools and memory. Humans approve, intervene, review. This is where most enterprise AI projects currently stall.

4. Production-grade agentic systems. Agents run inside real systems under real load, connected to real data and real downstream processes. Failures are observed. Recovery is automated. Other teams depend on the output.

The leap from three to four is where pilots go to die. A pilot can be run by a small team in a controlled environment with curated data and minimal dependencies. Production means integration with existing systems, security review, compliance checks, ongoing observability, and the ability to handle the edge cases that never appear in a demo.

Deloitte's 2026 State of AI survey of 3,235 enterprise leaders found that only 25% of organisations have moved 40% or more of their AI pilots into production. Talent readiness was the lowest-rated factor, at 20%. The bottleneck is not ambition. It is the capability to carry a project across the line.

That capability is what an FDE provides. It is also what the Blackstone joint venture is trying to manufacture at scale.

Why Anthropic's headcount cannot close the gap

Anthropic's total headcount sits at roughly 3,000 to 5,000 people. Their FDE function is new. The specific team size is not public. Whatever the number, Anthropic's own CFO has already admitted it is insufficient. That is why a separate company now exists.

80% of Fortune 500 companies are running active AI agents, according to Microsoft's February 2026 analysis of first-party telemetry. Almost every large organisation has an AI initiative. Most are parked at stage three. The supply-demand mismatch is not a temporary hiring lag. It is structural.

Both ventures were announced within 24 hours of each other. They are not the same thing.

The practical read: if your company is already in a Blackstone or Goldman Sachs portfolio, Anthropic's JV may come to you. If not, DeployCo is the broader-access option; neither is cheap or fast to engage.

The big consultancies are the obvious alternative and the obvious problem. Their staffing model assumes a large team of generalists, led by a partner who arrives for the steering committee and leaves before implementation. That model is mismatched to a job that requires two people with deep LLM deployment experience to sit inside a customer's environment for six weeks until something ships.

Internal teams face the inverse problem. They know the environment but are typically one to two years behind on hands-on agentic deployment experience. They have not yet built and broken enough production systems to know which patterns hold under load and which collapse quietly.

The supply is thin in three directions at once. The vendor cannot hire fast enough. The consultancies cannot move fast enough. The internal teams do not yet have the specific experience. Independent operators with production deployment scars are the fastest path through the gap, without vendor lock-in attached.

What it actually takes

The technical requirements of FDE work are more specific than they appear in job descriptions.

MCP servers are API surface boundaries. Designing one for an enterprise means deciding which internal systems the model is permitted to touch, what happens when an upstream system is degraded, how auth tokens are scoped and rotated, and how every call is logged for the compliance team that will eventually ask for it.

Sub-agents require a trust hierarchy. Which agent can invoke which tool, using whose credentials, under what conditions. What happens when a sub-agent hallucinates a tool call and the parent agent receives a malformed payload. You need observability at every boundary, or production debugging becomes archaeology.

Agent skills behave like small programs with unreliable inputs. They need deterministic behaviour under adversarial or malformed prompts, sensible defaults when context is missing, and tested fallbacks for the cases where the model does not understand what it is being asked. Skills that work on a clean evaluation set fail in week two of production because real users do not phrase requests the way the prompt designer expected.

None of this is in the demo. All of it is in the deployment.

The distinction that matters here is the same one Pragmatic Engineer draws between FDEs and Solutions Architects. Solutions Architects build offline proofs of concept using anonymised data. FDEs write production code on live customer infrastructure. The first proves the concept. The second ships the thing. The distance between them is exactly what Deloitte is measuring when it finds that only one in four organisations has moved the majority of its pilots into production.

The independent advantage

An Anthropic FDE is employed by Anthropic. Their mandate is to make Claude succeed inside the customer's environment. That alignment holds until the right answer for the customer is a different model for a specific step, or a deterministic script instead of an agent, or a simpler architecture than the one in Anthropic's reference stack.

An independent operator has no such constraint. They pick the right tool for each part of the job, design the architecture around the customer's systems rather than any vendor's roadmap, and stay as long as the engagement requires rather than rotating to the next logo.

The Blackstone joint venture will be good at deploying Claude into mid-market and portfolio companies at scale. It will not be good at bespoke architectural decisions that conflict with Anthropic's preferred patterns. Both things can be true at the same time.

A question to take to your AI vendor

Ask your AI vendor how many production deployments (not pilots, not proofs of concept) they have completed inside an enterprise environment like yours in the last six months. Ask who was on the ground. Ask what broke and how it was fixed.

If they cannot name the deployments, the engineers, or the failure modes, you are at stage three, and the work of getting to stage four has not started yet.

That work requires someone in the building. The job title is just the first sign that the industry is finally saying so out loud.

Sources: Anthropic FDE job posting; Blackstone/Anthropic enterprise AI services firm announcement; Pragmatic Engineer — Forward Deployed Engineers; Microsoft Security Blog — 80% of Fortune 500 use active AI Agents, Feb 2026; Deloitte State of AI in the Enterprise 2026 (n=3,235 leaders, Aug–Sep 2025).

Transparency Note

The ideas, arguments, and structure in this essay originated with the author. AI tools were used to assist with drafting, research, and revision. All claims, sources, framing, and final wording reflect the author's own thinking and were reviewed for accuracy before publication.

This essay is for informational and educational purposes only and does not constitute professional advice of any kind, including financial, legal, medical, or otherwise. The author makes no guarantees regarding accuracy or completeness. Readers should consult a qualified professional before acting on any information contained here. The author accepts no liability for decisions made based on this content.