On May 4, 2026, Anthropic announced a $1.5 billion enterprise services firm with Blackstone, Hellman & Friedman, and Goldman Sachs. The firm is called Ode with Anthropic. It exists to embed Anthropic engineers inside customer companies and ship Claude into production work. Fortune called it a shot at the consulting industry. Bloomberg noted the private-equity backing was designed to push engineers directly into portfolio companies.
Seven days later, on May 11, 2026, OpenAI launched the OpenAI Deployment Company. It arrived with more than $4 billion in committed capital, TPG as lead partner, Advent and Bain Capital and Brookfield as co-leads, and a founding acquisition. The acquisition was Tomoro, a London and Singapore firm with roughly 150 Forward Deployed Engineers already at Tesco, Virgin Atlantic, and Supercell. Constellation Research and The Next Web reported the numbers within hours. OpenAI's own release framed the firm as a partnership with nineteen global investors, consultancies, and system integrators.
Two announcements. One week. Both structured almost identically. Both aimed at the same problem. That problem is not that Claude and GPT are hard to use. It is that most enterprise AI budgets have shipped nothing.
Executives are being told this is the fix. It is a fix. It is also a trade, and the terms of the trade are not what the deck implies.
The forward deployed engineer, from Palantir to now
The role has a lineage. Palantir invented it in the early 2010s because intelligence customers could not fully brief a vendor from the outside. Palantir sent its own engineers inside, gave them the same technical bar as the core product team, and let them write production code against real classified data. The pattern became the Forward Deployed Engineer, or FDE. Steve Banker wrote about the model in Forbes on July 10, 2026, and treated it as one of the most consequential organizational designs in enterprise software.
The FDE builds pipelines, models the ontology, wires the workflows, and ships. They pass the same interview as the platform architects. Anthropic calls the same person an Applied AI Engineer. Google Cloud, Databricks, Scale AI, and Salesforce each hire under a variant name. The MarkTechPost inventory from May 20, 2026 counted the role across every major frontier lab, plus most large enterprise-facing AI infrastructure vendors.
Compensation is public. FDE Academy pegs total comp at $300K to $550K for mid to senior levels at the labs. Anthropic's posting on the General Catalyst job board lists the base range at $200K to $300K plus equity for the applied AI role, with fifty percent travel expected. These are hard senior engineers being paid frontier-lab money and sent onto enterprise floors.
The obvious question is who pays for that. The stated answer is the customer. The real answer is more interesting.
The engineer is priced at cost. The runtime is priced forever.
Ode with Anthropic and the OpenAI Deployment Company both stand up services arms with billions of committed capital, hundreds of engineers, and a founder-level insistence that this is not consulting revenue. The Fortune coverage of Anthropic's launch was blunt: the venture exists to accelerate Claude adoption inside private-equity portfolio companies, and the labor bill is a means to an end.
Read that carefully. The point of the services firm is not services revenue. The point is Claude adoption. The revenue that matters is the token bill after the engineers ship the system.
This is a specific commercial pattern. It is the same one Snowflake pioneered with its professional services desks, the one AWS built with its Enterprise Support tier, the one Databricks perfected with its Delivery Solutions Architects. Sell the human at close to cost. Bill the runtime forever. The human is the delivery vehicle for the meter.
The reason the frontier labs are copying Palantir now is that inference is a recurring meter and models are a fungible commodity to any buyer who cannot ship. If the buyer cannot ship, the meter never turns on. The FDE is how the meter turns on. Everything that follows in this piece assumes this design is real, deliberate, and rational. None of it is a scandal. What it costs the buyer never shows on the invoice, and that is worth spelling out.
Forward deployed AI engineering is a knowledge trade
Every consulting model is a knowledge trade. You import a skill you do not have, you get an outcome, and if you are lucky you retain enough of the skill to run the next round yourself. Big Four consulting worked that way for decades. It was slow, and expensive, and the retention rate was often terrible, but the shape was legible. A partner scoped the engagement, a team of associates did the work, a deck was delivered, and the client kept the deck.
Forward deployed AI engineering does not produce a deck. It produces a running system inside the lab's runtime. That system embodies a translation of your business. Someone had to figure out which of your invoices count as revenue, which of your tickets are duplicates, which customer types get escalated, which decisions must go to a human, which can go to Claude, which can go to a smaller finetuned model, which trigger legal review. Every one of those decisions is now written in code. It sits on the lab's platform, calls the lab's models, and follows the lab's tool-use conventions.
The person who wrote all of that down was the FDE. They now know how your company works. Not in a deck. In a working artifact.
Their next posting will be another company in your sector. The pattern they extract from you will shape the scoping of the next engagement. The reference architecture they build with you will show up as a template inside the lab's engagement library. The system-prompt patterns they discover with your data will end up in the lab's model behavior guidance. The labs say so publicly. Applied AI teams publish learnings; that is how the model gets better.
There has been no theft. You have contributed. The direction of the trade is unusual, though. You paid the vendor, and the vendor learned your business. The next customer is your competitor, and the vendor arrives at their door already fluent.
The fluency drift
For twenty years, the argument for hiring outside engineering help ran through capacity. You did not have the bodies, or you did not have the specific skill in the moment, so you rented. The knowledge deficit was temporary and the direction of learning was inward. Your team paired with the consultants, absorbed the pattern, and after the project you were better than before.
Forward deployed AI engineering flips the direction of learning. The FDE arrives fluent in frontier model behavior, in tool-use scaffolding, in eval design, in agent orchestration, in the latest patterns from the lab's own research releases. Your team is not. Your team asks the FDE what to do, and the FDE tells them, and the FDE ships it, because the FDE ships faster than your team can absorb.
At the end of six months you have a system. You do not have people who can extend it in the same lab-native style. If the FDE rotates off, your internal team can operate what is there. Extending it in the same idiom means calling the lab back. And the lab knows.
This is the fluency drift. Your team learned the local domain. The vendor's engineer learned the local domain plus the frontier runtime plus how to translate between them. That last skill, the translation skill, is where all the leverage lives. The lab keeps it because the lab writes the runtime and rotates the engineers.
Palantir customers have been living inside this pattern for a decade. Some love it. Some resent it. The ones who thrive are the ones who understood, going in, that they were buying an ongoing relationship with a runtime, not a one-time delivery. The ones who struggle are the ones who thought they were buying software.
Ode, the Deployment Company, and the disintermediation of the integrator
Note who is not centrally featured in either announcement. The traditional systems integrators. Accenture, Deloitte, TCS, Infosys, Capgemini. Anthropic has a Deloitte partnership that is genuinely large, worth reading, and dates to a 470,000-seat expansion first covered by CNBC on October 6, 2025 and detailed by AI Magazine. Deloitte and Anthropic even co-launched a formal certification program to train 15,000 practitioners on Claude. It is a real deal.
Ode with Anthropic is a different animal. Ode is a new firm, funded by asset managers, staffed by Anthropic-native engineering talent, aimed at portfolio companies of those asset managers. The OpenAI Deployment Company is not Accenture either. It bought Tomoro, an Anthropic and OpenAI-native shop with 150 people. Both firms are engineered to skip the middle.
The reason is a specific bet. The bet is that a Deloitte consultant, however well trained, will always be one hop removed from the model release schedule, from the underlying tool-use behaviors, from the eval workflow inside the lab. A Deloitte-trained Claude practitioner is not the person the lab sends when a difficult finetune has to work by Friday. The lab sends its own engineer. And the lab wants that engineer's account permanent inside your company.
This is a real threat to the traditional integrator model, and the integrators know it. It is why Deloitte, PwC, and Accenture have all announced enterprise Claude and GPT partnerships in the past twelve months, and why they are hiring their own applied AI benches at frontier-lab compensation. The pricing tension will be extraordinary. In two years, an enterprise buyer will look at three possible bidders for the same AI engagement: the lab's own services arm, a traditional integrator's applied AI unit, and an independent firm. The lab's services arm will be cheapest on labor and most expensive on lock-in. The integrator will be neutral on lock-in and slower on delivery. The independent firm, if it exists, will be the only party with no meter to sell you.
What the invoice does not price
An eight-month FDE engagement produces a lot of things a CFO can count. Hours billed. Systems in production. Ticket volume reduced. Handle time down. Those numbers will be real. They will also be an incomplete picture of what changed.
The uncounted line items are the ones that matter. Whose head holds the reasoning behind the choice of a specific model for a specific step. Whose head holds the failure taxonomy that the eval suite is trying to catch. Whose head knows why a particular tool was scoped one way instead of another. Whose head holds the escalation logic that determines when Claude is allowed to act and when it must page a human. If those heads all belong to the vendor's engineers, you have not built an AI capability. You have rented one, and the invoice you paid was for the rental deposit, not the asset.
A CIO who reads only the visible metrics will report progress to the board. A CIO who reads the invisible ones will notice the fluency has moved out of the building.
What you actually control
An executive reading this could easily conclude that the honest move is to buy the lab's engineers and accept the lock-in. Sometimes that is the honest move. If your first six agents are being built to run on Claude, and Claude is the best model for those workflows, and Anthropic will send you people who know the runtime cold, refusing that offer to prove a point about vendor neutrality is a slow way to lose the AI cycle.
The trade is real, though, and the trade has a specific mitigation. The mitigation lives in what you keep in your building.
You keep the ontology. You keep the labeled data. You keep the eval sets. You keep the runbooks that describe when a human is required and why. You keep the failure taxonomies that capture what went wrong in production and how the fix was scoped. You keep the tooling that lets your own people read what a model did and why. These artifacts are portable. A well-written ontology can be rehosted. An eval suite can be rerun against a different model. A failure taxonomy is yours regardless of who ships next quarter's release.
What you cannot afford to hand over is the authorship of these artifacts. The FDE will offer to write them, and they will write them well, and they will write them in a shape that reads naturally on the lab's platform and awkwardly anywhere else. This is the natural bias of a good engineer who has spent five years inside a specific runtime. No one is sabotaging you. Your job is to insist that a member of your own team co-authors every one of those artifacts, understands them, and can defend them in a meeting where the lab's engineer is not present.
This is what architecting the change looks like. Buying the tool is picking the runtime. Buying the service is picking the engineer. Architecting the change is deciding, in advance, which artifacts you will own, in what form, and how you will keep authorship of them as the models turn over.
The next decade is about who owns the translation
Every wave of enterprise technology has produced a translation layer between the business and the runtime. Mainframes had COBOL programmers. Client-server had systems integrators. Cloud had DevOps. AI has forward deployed engineers. Each of these layers, in its time, quietly became the place where the real leverage lived. The businesses that thrived through each transition were the ones that decided, early and deliberately, that they would own that translation layer themselves and pay whatever premium was required to keep it in-house.
The businesses that struggled were the ones that outsourced the translation layer to save money and then discovered that the vendor spoke the business more fluently than the business did. Some of those businesses recovered. Many did not. The pattern is old enough to be predictable.
Forward deployed AI engineering is a genuinely useful role, and the labs offering it are not villains. The right move is to accept the engineer and refuse the packaging. The packaging that leaves you with a running system and no ability to reason about it is the default. It is what an unarchitected buyer signs.
The architected buyer signs for the engineer and keeps the authorship. That is a different contract, a different governance model, a different set of internal hires, and a different set of artifacts on the way out. It has to be designed. It cannot be bought as a package because no lab is going to sell you the shape that lowers their own lock-in.
That is the work. At Agor AI Advisory we do exactly this work. We help executives specify what to keep in the building before the vendor's engineer sits down at the desk, so that when the engagement ends the fluency stays. We work in your ontology, your evals, your failure taxonomies, and your authorship, not the vendor's. The runtime will change. The models will turn over. The lab's engineers will rotate. Your ability to read your own AI system, and change it on your own timeline, is what compounds.
Sources
- OpenAI launches the OpenAI Deployment Company, OpenAI, May 11, 2026
- OpenAI Acquires Tomoro to Boost Private Equity-Backed AI Venture, Bloomberg, May 11, 2026
- Building a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs, Anthropic, May 4, 2026
- Anthropic takes shot at consulting industry in joint venture with Wall Street giants, Fortune, May 4, 2026
- Anthropic lands its biggest enterprise deployment ever with Deloitte deal, CNBC, October 6, 2025
- Palantir And Forward Deployed Engineering: What Should We Believe?, Forbes, July 10, 2026
- What is a Forward Deployed Engineer: The AI Role OpenAI, Anthropic, and Google Are Hiring in 2026, MarkTechPost, May 20, 2026
- Forward Deployed Engineer, Applied AI at Anthropic, General Catalyst job board