The scarcest enterprise AI skill isn't prompting. Anthropic is spending $100 million to train people who ship

Most enterprise AI programs do not stall because the model is weak. They stall in the weeks between a convincing demo and a system that a security team has approved, an operations team can run and a business team actually uses. Anthropic has now put a price on closing that gap: $100 million.

On October 2, Anthropic launched Claude Frontier Academy, with the goal of training 10,000 Frontier Deployed Engineers by the end of 2027. The first cohorts come from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. The interesting part is not the money. It is what the program chose to teach.

What Anthropic announced

The academy starts with a multi-day, in-person phase built around a simulated enterprise deployment. Participants who pass earn a Claude Resident Engineer badge and move into a 12-week residency in which they lead a real Claude project inside their own organization. Completing it earns the Frontier Deployed Engineer credential, which Anthropic expects to award for the first time in early 2027.

Participation is by nomination through Anthropic’s account teams, and the target profile is a hands-on software engineer who has already built with large language models. The first cohorts run in San Francisco, New York and London.

The program sits on top of a much larger certification base. Anthropic says professionals at 46,000 partner firms have earned more than 175,000 Claude certifications, and nearly 4,000 people have completed its Basecamp program. Frontier Academy is deliberately narrower: fewer people, judged on whether they can take a use case to production.

The curriculum is a map of where projects die

The curriculum covers use-case selection, implementation, security review, production handover and organizational adoption. Read that list as a diagnosis. Each item is a place where enterprise AI projects routinely get stuck:

1. Pick the use case
2. Build against real data
3. Pass security review
4. Hand over to operations
5. Drive adoption

• Where pilots usually stop: Pick the use case · Build against real data
• Where production is decided: Pass security review · Hand over to operations · Drive adoption

Engineering workOrganizational work

• Use-case selection decides whether the project is worth the governance effort at all. We argued in our five tests for agent-ready work that the best first agents slot into a decision an existing process already makes.
• Security review is where identity, data access and logging questions surface for the first time if nobody asked them earlier.
• Production handover is where a pilot becomes somebody’s operational responsibility, with runbooks, monitoring and an owner. Without it, the system belongs to the person who built it until that person moves on.
• Adoption is where a working system meets the people whose job it changes.

None of these is a prompting skill. They are the skills of an engineer who can also run a project through an enterprise.

Why this is a different role from “AI engineer”

The forward deployed engineer is not a new idea. Software companies have used engineers who sit with customers and build against their real systems for years. What is new is that a model provider is formalizing the role at scale, and that the credential depends on a production outcome rather than an exam.

For an enterprise, that changes the hiring and staffing question. An AI engineer who is excellent at retrieval, evaluation and model behavior is necessary but not sufficient. The person who gets a system live also has to write the threat model, negotiate data access, sit with the risk committee and design the handover. In a bank, that last part is not optional. The revised interagency guidance on model risk management that the Federal Reserve, OCC and FDIC issued in April 2026, replacing the long-standing SR 11-7, still expects sound development, validation and governance of models, scaled to their risk. A bank that treats an AI system as a model under that guidance needs someone who can produce all three. Morgan Stanley being in the first cohort is a hint about who feels this gap most.

This is the same conclusion our September synthesis reached from the other direction: the model is no longer the hard part.

What to keep in mind before relying on it

The credential is vendor-specific. A Frontier Deployed Engineer is trained on Claude. The organizational skills transfer to any platform; the product depth does not. If your estate is Microsoft-first, with Copilot Studio, Microsoft Foundry and Entra ID doing the heavy lifting, you need the same capability pointed at that stack.

The first wave goes to consultancies and large firms. Five of the eight launch organizations sell services. Most enterprises will meet this capability through a partner before they have it in house, which makes the handover step more important, not less.

Ownership after go-live is still yours. A residency ends after 12 weeks. The agent keeps running. We covered what happens when nobody owns an agent’s actions in our analysis of Copilot Autopilot.

What to do now

1. Name the role internally, whatever you call it: one engineer per production use case who owns the path from selection to handover, not just the build.
2. Write the handover package first: owner, runbook, evaluation set, access review, rollback plan. Make it the definition of done.
3. Pair builders with risk and security early. Put the security reviewer in the first design session, not the last.
4. When a partner delivers, buy the capability, not only the system. Contract for knowledge transfer and a named internal owner at go-live.
5. Grow a small cohort rather than a large audience. Broad AI literacy has value, but production depends on a few people who can carry a system through review.

The bottom line

Anthropic’s bet is that the bottleneck in enterprise AI is no longer access to models but people who can turn models into governed production systems. The curriculum, heavy on security review, handover and adoption, is a fair description of that bottleneck. Enterprises do not need 10,000 of these engineers. They need enough of them to ship, and a plan for who runs the system after they leave.