OpenAI and Anthropic Are Building Consulting Firms Because Models Alone Cannot Close Enterprise Deals

Two weeks. That is how long it took for both leading AI labs to announce they are no longer just model companies.

On May 12, OpenAI launched the OpenAI Deployment Company, nicknamed DeployCo, a standalone subsidiary backed by more than $4 billion in initial capital from TPG, Advent International, Bain Capital, Brookfield, and others, according to The Register. The new unit acquired UK-based consulting firm Tomoro, bringing approximately 150 Forward Deployed Engineers into the operation. Its mandate: embed engineers inside enterprise clients to diagnose AI opportunities, run proofs of concept, and build production systems.

One week earlier, Anthropic’s unnamed consulting joint venture, backed by Blackstone, Hellman & Friedman, Apollo Global Management, General Atlantic, GIC, and Sequoia Capital with $1.5 billion in committed capital, acquired San Francisco-based Fractional AI as its operational foundation. The deal ended Fractional AI’s 11-month partnership with OpenAI.

The timing is not coincidental. Both companies arrived at the same conclusion: selling API access is not enough.

The 95% Failure Rate That Explains Everything

The pivot makes sense when you look at one number. According to MIT NANDA’s State of AI in Business 2025 report, cited by MarkTechPost, 95% of enterprise generative AI pilots produce no measurable business impact. The models perform. Enterprise deployments consistently fail.

Enterprise AI fails at the integration layer. The client’s engineers understand their data schemas, compliance requirements, and legacy architecture. The AI lab’s engineers understand prompting patterns, retrieval strategies, evaluation frameworks, and failure modes at scale. Neither side has the other’s knowledge. A customer success manager cannot bridge that gap. Neither can documentation.

A Forward Deployed Engineer can. The FDE sits inside the client’s environment, writes production code against real systems, and stays until the thing works. The term comes from Palantir, which invented the role in the early 2010s to solve the same problem with intelligence agencies that could not clearly specify what they needed. Until 2016, Palantir had more FDEs than software engineers.

Palantir’s Q1 2026 results validate the model: 85% total year-over-year revenue growth, U.S. commercial revenue up 133%, according to Palantir’s investor release.

The Market Is Already Tilting

The enterprise AI market is moving fast enough that timing matters. According to Ramp’s May 2026 AI Index, Anthropic’s Claude models surpassed OpenAI’s GPT suite in business adoption last month for the first time. Claude penetrated more than one-third of the 50,000+ businesses on Ramp’s platform, while OpenAI fell nearly three percentage points to 32%. In the last year, Anthropic quadrupled its user base while OpenAI’s remained relatively flat, as reported by Channel Dive.

Ramp Lead Economist Ara Kharazian cautioned against crowning Anthropic the definitive leader, noting that “we have never seen a software industry as dynamic, where newcomers can disrupt market leaders in a matter of months.”

Meanwhile, the spending pool keeps expanding. Gartner’s latest forecast projects global AI spending will hit $2.59 trillion in 2026, up 47% year over year. AI model spending specifically will more than double to nearly $33 billion. But Gartner Distinguished VP Analyst John-David Lovelock told Channel Dive that the spending spike is “largely tied to model providers charging more for their products,” not organic volume growth alone.

That pricing pressure explains the urgency. OpenAI doubled API pricing with GPT-5.5, according to The Register. If enterprises are going to pay more per token, they need to see production value, not pilot dashboards.

The Margin Structure Tells the Real Story

Anthropic is on track to post its first operating profit in Q2 2026, projecting $10.9 billion in revenue for the June quarter, up 130% from $4.8 billion in Q1, according to PYMNTS. In Q1, Anthropic spent 71 cents on compute for every dollar of revenue. That ratio is expected to drop to 56 cents in Q2.

Those compute costs are the reason implementation services are so attractive. A consulting engagement generates revenue at the deployment layer without proportionally increasing compute costs. The model inference runs regardless. But the consulting margin sits on top of it, and the resulting enterprise lock-in drives sustained API consumption that pure self-serve cannot guarantee.

This is the Palantir insight, adapted for AI labs: the model is the loss leader. The implementation is the business.

Who Gets Squeezed

The obvious losers are the consulting firms that signed up as investors. Capgemini, McKinsey, and Bain all participated in DeployCo’s funding round, according to The Register. They are funding their own competition. OpenAI’s stated plan is to call in its FDEs when consultants need help proving AI value to clients. In practice, that means OpenAI’s engineers will be inside the client before the consultant’s team, building relationships and gathering proprietary workflow data.

The less obvious losers are agent infrastructure companies. OpenClaw, LangChain, and the broader agent tooling ecosystem depend on enterprises that self-serve: teams that pick their own models, build their own orchestration, and deploy through open-source stacks. When OpenAI or Anthropic sends an FDE into an enterprise, that engineer is not recommending OpenClaw. They are building on their employer’s stack.

Gartner’s Lovelock framed the endgame bluntly: “There simply isn’t sufficient revenue for several redundant LLM ecosystems.” If the model layer consolidates to two or three winners, and those winners also own the deployment layer, the surface area for independent agent infrastructure shrinks accordingly.

The Accenture Precedent

Anthropic is also building parallel channels. In the last two weeks alone, the company signed alliances with KPMG and PwC to deploy Claude Cowork inside their client delivery platforms, and launched Claude for Small Business to cover the lower end of the market.

This is the full-stack play: a consulting JV for mid-market enterprises, Big Four partnerships for large enterprises, and a self-serve product for small businesses. Every segment covered. Every customer touchpoint owned.

OpenAI is running the same playbook. DeployCo handles mid-to-large enterprise. The Thrive Holdings investment covers managed IT services. ChatGPT handles consumers and small teams.

Neither company is content being a model provider anymore. They are building the next Accenture, one where the consulting arm has a structural advantage because it controls the underlying technology. Accenture has to be vendor-neutral. DeployCo and Anthropic’s JV do not.

The question for the rest of the ecosystem is whether enterprises will care. The 95% pilot failure rate suggests many will welcome someone, anyone, who can make AI work in production. If that someone also happens to sell the model, so be it.