Margin of Safety #55: Why OpenAI and Anthropic are building the new Accenture

OpenAI and Anthropic are aggressively launching massive, multi-billion dollar consulting arms

On May 4th, within hours of each other, OpenAI and Anthropic each announced a consulting firm. Anthropic partnered with Blackstone, Hellman & Friedman, and Goldman Sachs to form an unnamed enterprise AI services company with roughly $1.5B in committed capital and Anthropic’s own applied engineers embedded in client deployments. OpenAI went bigger: The Deployment Company(TDC) raised over $4B from 19 PE investors at a reported $10B valuation, majority-owned and controlled by OpenAI.¹

Coverage of these deals has framed them in one of two ways: either this proves the models aren’t enterprise-ready or it proves enterprise is hard to deploy into. We even had a version of this argument ourselves:

Kathryn: The more they invest in services, the clearer it is that the product isn’t easy to deploy at scale. If models were as capable as marketed for enterprise workflows, you wouldn’t need $4B and an army of forward-deployed engineers to make them work.

Jimmy: I don’t think that’s the right frame. Show up to a mid-sized company with a capable AI system and you’ll find legacy ERP, a security team that defaults to no, compliance requirements built in spreadsheets, and workflows that were designed around human-preferred interfaces and systems, with no eye towards automated systems.

This apparent disagreement came because we approached the problem from different angles. Jimmy’s chain of thought took him to the reasons for needing FDEs, while Kathryn framed it in terms of the business consequence. But we both arrived at the same conclusion; deploying AI into the enterprise requires wrestling a good-enough core model and a pile of elbow grease to fit that to the specific business in question. This elbow grease doesn’t necessarily generalize, because process correctness is measured per-customer in the context of all their technical debt and idiosyncratic preferences.

Most previous technology waves involved systems with a relatively easy-to-measure, binary correct or incorrect outcome. In the context of AI for business process, “correct” varies with each firm’s goals and capabilities and can be exceptionally hard to measure. For example, you can solve a customer’s immediate problem while leaving them dissatisfied in ways that don’t surface until they churn. On top of that, you can’t necessarily lift-and-shift a business process to Claude the way you lift-and-shift a workload to EC2.² Better models can indirectly improve things, since they make it faster and easier to deploy ad hoc code that bridges the gap between current process and automation, but they don’t solve for measurement difficulties, underspecified processes, and other human vagaries.

The real question

This changes the question of models’ enterprise readiness. We think the question of ready or not is a false dichotomy, and the key question is if or how the cost of successful deployments shifts over time.

Palantir is often cited as proof that efficiencies can be found. Palantir’s methodology has made generalization progressively less expensive, with commensurate improvements to operating margin – and the stock price to show for it. However, they haven’t eliminated the need for an army of forward-deployed engineers; we’d also note their adjusted figures are materially more flattering than GAAP. These ventures will presumably eventually be measured under GAAP.³

If the labs’ services layer can build an idealized version of the Palantir model, even their GAAP margins will steadily improve. But if engagements remain idiosyncratically expensive, especially despite repeat investments in a vertical, the Newcos will be more like Accenture than Palantir. That’s still a strong business, but with a much lower revenue multiple.

Why the labs want to own this layer

The natural question, especially given the automation and valuation risk, is why the labs are building this ground up versus partnering with existing SIs. Replicating Accenture and Deloitte’s GTM expertise while launching a $4B company from scratch is a heavy lift. We think two structural factors tip the scales toward ownership before we even get to the classic AI data argument.

First, quality in these engagements is hard to specify contractually. The same way that quality makes rollouts hard, it also complicates contracting – if you can’t agree on what good looks like, how can you agree on the deliverable for a SOW? Second, the pace of AI innovation requires a fast feedback loop between production failure and model/system improvement. An extra entity in the middle of the relationship is another player in the game of quality telephone; not only do they add latency, but they potentially blur or dilute the signal between the end user and the core engineering team. Both factors push towards ownership of a united team.

We often hear data ownership cited as a motivating factor in the space. Better training data is certainly valuable to the labs as they try to better fit enterprise workflows. However, enterprise contracts⁴ likely preclude outright data ownership for training purposes, so any valuable data will be organizational learning versus direct training datasets. And significantly, these sorts of organizational learnings are the ones that are often diluted when feedback passes through SI intermediaries. We expect the unfiltered feedback loop to have material value, but the extent will depend on contract terms.

Who’s actually under pressure

The obvious pressure point is traditional systems integrators. The new competitors created by Anthropic and OpenAI will presumably have preferential model access, faster escalation paths, and the ability to directly influence lab roadmaps. Competing from the outside with rate-limited model access and a disconnect from the model owners is a durable disadvantage⁵, one the labs are betting will outweigh existing SI strengths.

Staffing asymmetries may be equally durable. Effective AI integration requires in-demand AI engineers that Accenture and Deloitte may struggle to hire and retain, as their compensation frameworks and career tracks were built for a different generation of engagement. In contrast, the new firms can offer equity, different compensation bands, and a culture that attracts in-demand AI experts, without needing to blend those changes into an existing technical level and salary framework.

That’s not to say that the AI labs themselves do not have any pressure. They have a big need to prove enterprise adoption beyond coding. For Anthropic, coding is probably the one (and maybe the only) space where they can proudly declare widespread enterprise adoption. For OpenAI, seems like they need more even in the realm of coding. However, outside of coding, AI labs have not really found their way into true enterprise workflows besides simple chat and information retrieval features. That in turn will impact their valuations and ability to raise gargantuan amounts of capital to fuel model development & other products.

The pressure may extend further. If you believe that complex business automations pass the ownership test, that could ultimately have implications for even the newly formed consulting companies. After all, they still don’t own both the AI and the business process. In a world where ownership really matters, the best-positioned players may be operating companies like Google or Meta that own a full-stack business with an attached AI lab, running the feedback loop entirely internally without a contract in the middle. Against that structure, even the labs’ integration ventures are contractors.

1. These are reported figures from secondary coverage. Deal terms in both cases are not fully public and secondary reporting may contain errors.

2. The hyperscaler analogy is useful here. Cloud migration was also organizationally complex, and containers succeeded partly by allowing legacy workloads to be migrated without a complete re-architecture. The difference is that a containerized legacy app has a binary success condition: it runs or it doesn't. AI output correctness is semantic and domain-dependent; there's no equivalent of "the container started."

3. Palantir's adjusted operating margins are substantially higher than GAAP, primarily due to stock-based compensation treatment. We believe the army of forward engineers is a material portion of this expense, so it can't be separated from their deployment model. And the gap has historically been significant enough to matter when using Palantir as a margin ceiling reference.

4. One advantage to partnering with PE is that the joint ownership model may align incentives enough to allow for a greater degree of data sharing. However, to the extent that the goal is to eventually serve clients not owned by the same PE firms, there should be strategic reluctance to share data in ways that will ultimately serve those firms’ competitors. We expect this dynamic to ultimately cap the level of data the new firms can extract, even within PE-aligned structures.

5. A world in which it isn’t durable is a world in which the deployment challenges we opened the blog with are solved — at that point, the investors in these new companies have some explaining to do and the labs are facing a very different challenge.