What should a European company ask that a US buyer might not?

Which legal entity processes the data, in which jurisdiction, and under whose disclosure obligations. Ask where inference runs, which subprocessors sit underneath, and whether the deployer duties under the EU AI Act land on you or on the vendor. A delivery contract signed with a US parent can put your inference traffic under a jurisdiction your own regulator did not choose.

What survives the engagement, whoever you hire?

Ask for the answer in writing before signing: the code, the integration layer, the evaluation sets, the accumulated context about how your business actually runs, and the right to run all of it on a different model. If any of those five is described as the vendor's platform rather than your asset, you are renting the result, not owning it.

OpenAI's DeployCo, Anthropic's Ode, and the case for an independent forward-deployed partner. Nucleo

Q: Should we buy forward-deployed delivery from the model vendor itself?

It is a reasonable choice if you have already standardised on that vendor's model, you want proximity to its unreleased roadmap, and the workflow you are automating is not the one your business competes on. It is the wrong choice if the system will hold the knowledge that makes your company hard to copy, because a lab-owned team cannot be indifferent about which model that knowledge ends up shaped around.

Q: Is a lab-owned delivery team model-agnostic?

Not by construction. The delivery arm exists to increase consumption of the parent's model, and its engineers are deepest in that model's tooling. That is not a criticism of the engineers, it is a statement about incentives. An independent partner can run a head-to-head evaluation and pick the losing vendor's model without a conversation with its own board.

The short answer

The model vendors now sell the delivery model too, so the question a CIO faces is no longer whether to embed engineers, it is who they work for. OpenAI announced its Deployment Company in May 2026 and bought the London consultancy Tomoro to staff it. Anthropic, Blackstone and Hellman and Friedman announced Ode with Anthropic on 15 July 2026, and it has since absorbed two more consultancies, Fractional AI and, on 21 August 2026, Casper Studios.

Both build real systems with real engineers. What neither can do is be indifferent about which model your company ends up built around, and that indifference is the whole asset when the model market turns over every six months.

Key takeaways

The forward-deployed model stopped being a differentiator in 2026. When OpenAI, Anthropic and Google Cloud all fund embedded delivery in the same year, embedding engineers is table stakes, not a pitch.

A lab-owned delivery arm is single-model by construction. It exists to increase consumption of the parent’s model. That is an incentive, not an accusation, and it decides what gets recommended in week one.

The buying question moved. Not who deploys, but what you hold when they leave: the code, the integration layer, the evaluations, the context, and the right to run it on a different model.

Google took the third route. At Cloud Next in April 2026 it committed a partner fund to put its own engineers alongside Accenture, Capgemini, Cognizant, Deloitte, HCLTech, PwC and TCS. The integrators keep the client relationship, Google keeps the model underneath.

For a European mid-cap there is a fourth question. Which legal entity holds your data, in which jurisdiction, and whether the deployer duties land on you.

What actually changed this year

For most of the last decade the forward-deployed engineer was a Palantir peculiarity: send engineers into the customer’s environment because the customer cannot write a specification for a problem they have never seen solved. Vendors copied the idea slowly. Independent firms built practices on it. It was still niche enough in 2025 that explaining it was half of every sales conversation.

Then, in a single stretch of 2026, the three largest AI companies in the world all bought their way into it. OpenAI stood up a deployment arm and acquired Tomoro, which had been building enterprise systems in alliance with OpenAI since 2023. Anthropic did it as a joint venture rather than a subsidiary, standing up Ode with Blackstone and Hellman and Friedman and a long list of financial backers, then buying Fractional AI and Casper Studios to fill it with people who had already done the work. Google Cloud took the third path at Cloud Next in April 2026, funding the integrators instead of becoming one, and putting its own forward-deployed engineers alongside Accenture, Capgemini, Cognizant, Deloitte, HCLTech, PwC and TCS on customer deployments.

Read those three moves together and one conclusion is unavoidable: the labs have concluded that the bottleneck in enterprise AI is not model capability. It is everything between a capable model and a system that runs in your company on Monday morning. We have argued that for two years. It is now the stated position of the companies that build the models.

That settles the argument about the delivery model. It opens a different one.

What the lab-owned arms are genuinely good at

Being fair about this matters, because a buyer who reads only the criticism will discount the rest.

A delivery team owned by a model vendor has three real advantages. It sees the roadmap before you do, so it can build against a capability that ships next quarter rather than working around its absence. It has the shortest possible escalation path when the model behaves strangely on your data, which is worth more than it sounds at three in the morning during a rollout. And it has engineering density in one model’s tooling that an independent generalist takes longer to match.

If your company has already standardised on one lab, has no intention of moving, and is automating a workflow that is not the one you compete on, buying delivery from that lab is a defensible decision. Say so plainly at the board meeting and it will hold up.

The thing a lab-owned arm cannot do

It cannot run an honest head-to-head model evaluation and act on the result.

Not because its engineers are dishonest. Because the arm exists to increase consumption of the parent’s model, and every incentive above the engineer points one way. When the right answer for your document-extraction workflow is an open-weights model running on your own infrastructure at a fraction of the cost, an independent partner writes that in the evaluation and moves on. A lab-owned team has to have a conversation with its own board first.

This compounds over an engagement. The first system gets built around one model’s tool-calling conventions, its context window, its prompt-caching behaviour, its safety filters. The second system inherits the first one’s abstractions. Two years in, the cost of moving is not an API change, it is a rewrite of the layer where your company’s actual knowledge lives. Nobody made a decision to lock in. It accumulated.

The independent version of this argument is already being made in the market. QuantSpark, a UK firm working mostly with private equity portfolio companies, publishes a direct comparison of the lab-owned arms against model-neutral independents and lands on the same point about evaluating models per problem rather than per vendor. That is the right question for the buyer they serve. For a European company it is one of four.

Ownership is the question, not model choice

Model-agnostic is necessary and not sufficient. Swapping the model out solves one layer. Five accumulate during a serious engagement:

The model. Which provider generates the tokens. The only layer most vendor content discusses.
The orchestration. How agents are sequenced, retried, evaluated and rolled back. Usually the vendor’s framework.
The context. The structured, current record of how your business actually works: which systems disagree about a customer, what the exception path really is, how the best person on the team does the job. This is the layer that takes eighteen months to build and cannot be re-bought.
The governance evidence. Audit trail, traceability, access control, retention. The artefacts that prove to a regulator what happened.
The organisational knowledge. Who inside your company can change the system without breaking it.

A contract that gives you the code but leaves the context inside a vendor’s platform has given you the cheapest of the five layers. Ask about layer three specifically. Ask what format it is in, where it lives, and whether it means anything after you stop paying.

This is what we mean by owning your intelligence, and it is why we describe ourselves as a transformation company rather than an agent vendor. Agents are what runs on the intelligence layer. They are not the thing worth owning.

The four questions a European buyer should add

The comparisons written for a US or UK buyer stop at model neutrality. If you are a European company, the jurisdiction questions are not a compliance footnote, they are part of the architecture.

Which legal entity processes the data, and where. A delivery contract signed with a US parent can route your inference traffic under disclosure obligations your own regulator did not choose. Data residency and legal jurisdiction are separate questions, and only the second one decides who can compel access.

Where does inference actually run, and which subprocessors sit underneath. Ask for the list, in writing, including the ones the vendor considers its own supply chain rather than yours.

Who carries the deployer duties. Under the EU AI Act, obligations split between the provider of a system and its deployer, and most enterprises are deployers without having noticed. If the vendor’s contract is silent on this, you are the deployer.

What happens to the evidence if you switch vendors. Audit trails and traceability records are not portable by default. If they live in the departing vendor’s platform, your ability to answer a regulator about last year’s decisions leaves with them.

Where this leaves the decision

The lab-owned arms proved the delivery model works, at a scale no independent firm could have funded. That is good for every company buying AI, ours included. It also means the buyer’s question has moved on from who embeds engineers to what those engineers leave behind and whose roadmap it is tied to.

If the workflow is commodity and the model is already chosen, buy the delivery from the lab and get on with it. If the system you are building will hold the knowledge that makes your company difficult to copy, hire someone whose only loyalty is to the outcome, and get the answer about the five layers in writing before anyone starts.

FAQ

What is OpenAI’s Deployment Company? A services arm OpenAI announced in May 2026 to embed forward-deployed engineers inside enterprises and build production systems around its models. It was assembled in part by acquiring Tomoro, a London consultancy that had been working in alliance with OpenAI since 2023.

What is Ode with Anthropic? A standalone enterprise AI services firm announced on 15 July 2026 by Anthropic, Blackstone and Hellman and Friedman, backed by an investor group including Goldman Sachs, General Atlantic, Apollo, GIC and Sequoia. Its engineers embed inside client companies and build systems on Claude. It acquired the applied AI firm Fractional AI, and the consultancy Casper Studios on 21 August 2026.

Should we buy forward-deployed delivery from the model vendor itself? Yes if you have standardised on that vendor, want proximity to its roadmap, and the workflow is not the one you compete on. No if the system will hold the knowledge that makes your company hard to copy, because a lab-owned team cannot be indifferent about which model that knowledge gets shaped around.

Is a lab-owned delivery team model-agnostic? Not by construction. The arm exists to increase consumption of the parent’s model, and its engineers are deepest in that model’s tooling. An independent partner can run the evaluation and pick a competitor’s model without an internal conversation about it.

What survives the engagement, whoever we hire? Get it in writing before signing: the code, the integration layer, the evaluation sets, the accumulated context about how your business runs, and the right to run all of it on a different model. If any of those is described as the vendor’s platform rather than your asset, you are renting the result.

Nucleo builds the intelligence layer inside your company, model-agnostic, under your governance, and it stays yours when we leave. Talk to us.

Sources: Anthropic, Blackstone bet the next trillion-dollar AI business is implementation; Forward-deployed engineers are the AI industry’s latest talent obsession; Why OpenAI and Anthropic are hiring forward deployed engineer teams.