E240 | OpenAI Teams Up with PE to Plow $4B into Deployment, and a Look at Silicon Valley’s Hottest New Role: FDE

Summary

• OpenAI has launched a Deployment Company and acquired Tomorrow, taking 150 FDEs with it; Anthropic has simultaneously formed a joint venture with Blackstone and other financial institutions. Both labs are pushing the same service: deploying models inside enterprises. Joe, who leads Cresta’s FDE function, says this confirms that “the model itself isn’t a product; getting a product into the real world still requires a huge amount of work”—exactly the work model companies “can easily overlook, or even look down on.” The line between model companies and application companies is becoming “an increasingly tangled relationship between friend and foe.”
• FDEs, or Forward Deployment Engineers, were invented by Palantir more than a decade ago, but have taken off because no one wants to hand AI deployment’s complexity back to the customer. Joe calls the role a “forward deployed CTO”: the FDE has to get the AI application running in production and lock in the customer, while feeding frontline lessons back into the product. A decade ago, when a Palantir FDE found a bug, they could only “write a letter or open a ticket and beg someone else to fix it.” Today, AI coding lets FDEs iterate quickly across 5 or 10 repos, while skills distill experience into reusable assets and create a “snowball effect.”
• FDEs are not billed separately by the hour. “We have never charged for FDE hours,” Joe says. Cresta sells outcomes, not tokens, and may even subsidize deployment with free labor—“like home delivery: it’s already at your door for you to try on, and you’re reluctant to send it back.” A single conversation may use 20-plus models, with the system choosing whichever is “good enough and priced right”—a degree of freedom Google’s FDE teams, tied to its own full stack of Gemini and proprietary voice technology, do not have.
• Model companies are turning to PE rather than consulting firms because PE can put a direct incentive in front of the top executive. “If you use this FDE model, some of my investments may move faster,” is a much more direct pitch than slowly negotiating scope with a consultancy. Oliver adds 3 PE motivations: the signaling value for LP fundraising—“I have to prove I’m at the leading edge of AI, or LPs won’t give me money”—portfolio value creation, and exposure to high-growth assets through the partnership structure itself. With “SaaS is dead” rhetoric running hot, GPs are highly anxious, while AI’s biggest value creation opportunity “is precisely not software companies”—it is traditional sectors such as business services, industrials, and healthcare. AI roll-ups are emerging as a new playbook.
• Oliver’s central view is that AI’s real value lies in creating revenue, not cutting costs. His question for clients is: “If I gave you 10,000 college-educated employees for free right now, what would you do?” A large asset manager used an AI sales assistant to reach customer segments it could not previously serve, while a dairy company gave employees back the time spent writing reports so they could focus on cow health. Deployment has 2 major failure points: the data platform is not ready—“no matter how smart AI is, without enough information and knowledge it can do nothing”—and companies let AI handle tasks that should be deterministic. In a 10-step workflow, 5 steps need to be deterministic.
• FDEs will not be replaced by AI in the near term. Over the next 1-2 years, tools such as transcription and translation software and Glean could lift the average FDE from 2 or 3 projects to 5 or 6. Further out, the market will bifurcate: high-end FDEs will handle the hard problems, while cheap FDEs will emerge for the long tail of small and midsize customers, including remote offshore teams. The logic is similar to small clinics that once could not afford software engineers but now want to hire one. By the time 99% of the role can be automated by AI, “the entire industry will already be different.” Humans become “AI’s tentacles”—collecting context AI cannot access and taking responsibility for decisions, with skin in the game. “You can’t expect a person to have coffee with AI.”

Deep dive

1. Two Frontier Labs Move into Deployment: Models Are Not Products

• In early May, OpenAI announced the launch of a Deployment Company, backed by private equity and consulting firms, and acquired a company called Tomorrow, “taking 150 FDEs with it.” Around the same time, Anthropic announced a joint venture with financial institutions including Blackstone. Both are offering the same service: bringing models into real enterprise workflows and making them work there.
• Joe’s first reaction was: “I knew this was coming, but I didn’t expect it to come this fast.” The move confirms 2 things. First, “the model itself isn’t a product.” Getting it into production requires a large amount of work that traditional model companies “can easily overlook, or even look down on”—customers who have the money to buy the model may still “have no idea how to use it.” Second, the boundary between model companies and application companies is blurring, and the relationship between “friend and foe” is becoming more tangled.
• OpenAI and Anthropic are making a straightforward staffing calculation: keep the best people training models, while deployment across all 360 lines of business requires more labor than they necessarily want to hire themselves. One is using acquisitions; the other is using capital to have someone else do the work.

2. What an FDE Is: A Forward Deployed CTO

• Joe’s one-line definition is that an FDE is “an engineer who works closely with customers and can get an AI application running in the real world,” while also being responsible for making the product better. “You have to close the deal, deploy the AI application, and lock in the customer,” then use those lessons learned to make the product stronger and stronger—even to the point of “self-disruption.”
• The model originated at Palantir. Military customers were not willing to explain clearly what they wanted, so “you had to go to the same tent in the same military camp and look at the data before they would tell you the details.” That led to 2 teams: Echo, made up of business operators who understood operations and rescue work, and Delta, a more forward-support software engineering team. One was technical and one was nontechnical. Joe says it is “hard to say what proportion of Palantir’s success came from FDEs,” but the company established the model.

3. The Cresta Playbook: Start with Data and Target the 80/20 Use Cases

• Cresta started in 2017, before the current AI wave, and compliantly accumulated customer-service text and voice data from major clients including Marriott. When building AI agents, it first identifies use cases with high volume, clear SOPs, and limited human judgment. “Eighty percent of the volume comes from 20% of the use cases.” The team extracts patterns from historical conversations and “can even use the data to train a small model for more simulations,” reducing the amount of guesswork.
• The FDE is an experienced AI deployment operator who decides which use cases to tackle first and whether the resources are in place. The team does not generally rewrite the customer’s APIs; it co-creates with the customer. Once the system is live, extensive testing, optimization, and monitoring remain. “Building it is only the first step.”

4. Forward Does Not Mean Permanently Onsite, and FDEs Are Not Billed by the Hour

• “The word forward catches people’s attention, but including myself, I’ve never spent more than a week at a customer site.” The typical rhythm is to fly to the customer’s office for 2 or 3 days of closed-door meetings, set high-level goals and KPIs, validate the APIs, and, if things go smoothly, build a small POC to get the customer excited. Then everyone returns home and continues remotely. The real value of being onsite is building personal rapport and trust: many things that are inconvenient to put in writing are conveyed through conversation, which builds rapport and reveals context that is difficult to reach remotely.
• FDEs are involved before, during, and after the sale. Once the contract is signed, 2 to 4 months to launch is a common expectation. After contractual metrics such as satisfaction, call duration, and case resolution rates are met, the FDE exits. “Small fixes and tweaks” are left to colleagues who are less AI-savvy.
• “We have never charged for FDE hours.” Cresta may even subsidize the work: a customer that has not paid anything may still receive 2 or 3 months of deployment support. “It’s like home delivery: it’s already at your door for you to try on, and you’re reluctant to send it back. The pace in AI is extremely fast right now. Everyone is racing to claim territory.”

5. FDE + FDPM: The CTO-CEO Split

• The ideal is “a one-person company, both CEO and CTO,” but people who can do both are hard to hire. Cresta therefore splits the role into 2 positions: the FDE is the forward deployed CTO, while the FDPM, or forward deployed product manager, is the forward deployed CEO. The FDPM uses people skills and negotiation skills to stay with the customer, and owns the agent’s overall behavior and quality. The FDE ensures the implementation is technically sound and the testing is robust, while bringing lessons learned back into the product.
• The rough ratio is 1 FDPM to 2 or 3 FDEs, regardless of project size, with each FDPM handling several projects at once. The team also deliberately builds domain specialists: some people go deep on healthcare insurance terminology and regulations, while others specialize in technical areas such as payments and search. “The world we have to operate in is extremely complex.”

6. Hiring Standards: No Juniors, and “AI Engineer” Means Nothing

• Joe’s goal is to “build the best FDE team in the world.” The bar has 3 parts. First, candidates must be qualified engineers who have developed and tested AI agents. “Writing ‘I’m an AI engineer’ on your résumé is meaningless. Which software engineer today doesn’t use Cursor or Cloud Code? If you don’t, you’re already badly behind. But not many people know how to develop and test an AI agent.” Second, they need strong customer-facing experience. Consulting, founding engineering, or freelance work all count, but “you have to be able to talk to the other side’s CTO or IT director, and sometimes say no.” Third, they must be dependable and resilient.
• Resilience matters because, as Joe puts it bluntly, “FDEs are really busy. You’re operating in a deeply imperfect world: the APIs are made of tissue paper, the SOPs are practically nonexistent, the documentation is all over the place, and the pressure is intense.” That is why he favors founders and co-founding engineers who have “been through storms and know that nothing is guaranteed to succeed.” “I don’t hire junior FDEs. A very junior person will struggle to build trust with the other side’s CTO. You can’t just open AI and ask it what to do.”

7. AI Coding and Distilled Skills: The Inflection Point in FDE Economics

• Why are FDEs taking off only now? “What could a Palantir FDE do 10 years ago even if they knew the product had a lot of bugs? They could write a letter or open a ticket and beg someone to fix it, and it might not be fixed for another 6 months.” Today, AI coding is powerful enough that “even if you have 5 or 10 different repos in different languages, you can easily get AI to make the changes, then find the right person to review them.” The feedback loop from the frontline back into the product finally works.
• The other lever is skills. Experience used to remain in people’s heads, requiring lengthy knowledge transfer. Now it can be written into a long Markdown file and a few scripts. “After doing 2 or 3 similar things, it becomes a hard skill. If 20 more people join our 30 FDEs, they can install the skill and access it without having to learn it from scratch.” The result can “very easily become a snowball effect.”

8. Will AI Replace FDEs? First Productivity, Then Bifurcation

• “The only constant is change, but compared with many other software engineering roles, FDEs are still a long way from being AI-ized.” Over the next 1 to 2 years, tools for transcription, translation, and searchable recordings, along with Glean for querying chat histories and code, will make multitasking more efficient. “The average person handles 2 or 3 projects today; that could become 5 or 6.”
• Further out, the market will bifurcate. High-end FDEs will handle problems tools cannot solve. At the same time, “people who previously did not need an FDE will ask, can I hire a cheap one?” The logic is similar to software engineering demand rising as small clinics and individual businesses that once felt they could not afford an engineer start wanting one. There will be remote FDEs serving the small and midsize long tail without ever going onsite, including teams in places such as Vietnam. “As long as customer complexity remains, there will always be a gap in what AI can fully automate. FDEs have to fill that gap.”
• The fallback judgment is simple: “Even SDRs have not been AI-ized very well yet. If one day an FDE can be 99% AI-ized, even with agent-to-agent alignment, the concern will not be the FDE itself. The entire industry will already be different.”

9. Why PE, Not Consulting: Put a Bait in Front of the Top Executive

• The point that surprised people was that OpenAI did not turn to an Accenture-style consultancy and instead went directly to private equity. Joe’s explanation: software outsourcing still leaves you with “a tool person,” along with the need to decide which team to use and fill in a project timeline. But if your investor says, “If you use this FDE model, some of my investments may move faster,” that puts a direct incentive in front of the top executive and accelerates the process. It is more “simple and blunt” than slowly negotiating scope with a consultancy, while also consuming more model tokens along the way.
• Consulting and software outsourcing firms face “not an existential crisis, but a major inflection point.” For the CRUD work they used to handle, “the customer may no longer feel compelled to hire an outsourcer; they can do it themselves.” Everyone wants to make the AI opportunity as large as possible, but every company will have to work extremely hard to capture its share.

10. Oliver’s Entry Point: The Huge Gap Between Personal Usage and Enterprise Adoption

• Oliver was previously a private-equity consultant at McKinsey on the Rewired team and is now VP of enterprise business at Invisible Technologies. The company’s name comes from the idea that “when technology is good enough, it becomes invisible.” His diagnosis is a huge gap between individual AI usage and enterprise adoption, driven in large part by supply. “Either the model provider sells it directly, or you get a wrapper product, such as Harvey for legal work. They are all good tools, but they haven’t changed the way you work.” The result is that many companies deploy AI without feeling any difference.
• His answer is not to deploy one tool at a time, but to enter “one workflow at a time.” Break the workflow apart: of 10 steps, 5 must be deterministic, such as mathematical calculations and compliance checks that cannot fail; 3 or 4 can use AI and allow flexibility; and 2 require human review. Every workflow must be customized for the individual company.
• On OpenAI’s deployment company, he says, “this is exactly the right move.” CFOs are talking about cost compression, while MIT and Stanford reports show that only a handful of companies have actually run AI at scale. The gap is unsustainable, and “just selling a chatbot” does not prove ROI. He adds one caveat: moving from a horizontal general-purpose model to building customized enterprise workflows is a completely different market move and sales motion. “I believe they can figure it out, but it will take some time.”

11. PE’s 3 Demands: Signaling Value Comes First

• The evolution over 3 years is stark. Three years ago, PE firms asked, “Can you come explain how AI works?” Two years ago, they asked, “How do we push AI through our portfolio?” This year, the message has completely changed: “I’m raising money from LPs, and I have to prove I’m at the leading edge of AI, or LPs won’t give me money.” Demonstrating an AI position can determine whether fundraising succeeds or fails, and partnering with the industry’s biggest names is powerful third-party validation. The second goal is real value creation in the portfolio. The third is returns: “The structure of these partnerships is highly attractive. At its core, it gives GPs exposure to high-growth assets.”
• The backdrop is that “the ‘SaaS is dead’ narrative has been extremely loud this year.” Over the past 5 to 10 years, the 2 largest PE asset categories were healthcare and software. “Almost every PE firm has exposure to software companies, so LPs and GPs are highly anxious.” Yet AI’s greatest value creation opportunity “is often precisely not software companies”—it is business services, industrials, healthcare, and essentially every sector where software previously could not do much.
• GPs themselves are an ideal target. Sourcing deals, valuation, investing capital, and managing assets are highly labor-intensive processes that require very expensive people. That workflow is precisely the kind of work AI can transform.

12. From Cost Cutting to Revenue Creation: “What Would You Do with 10,000 College-Educated Employees for Free?”

• Oliver gives the example of a large asset manager that wanted to distribute products through the small-asset-manager channel. The small managers required a sales representative to attend every customer meeting, eliminating the profit, so the firm built an AI sales assistant. The system combines data infrastructure for 1,000 products, an access-controlled input layer, a deterministic optimal-product-combination module—“it is essentially math”—pre-meeting talking points, in-meeting tools, and automatic post-meeting updates. The 7-step loop lets the large asset manager serve a much wider customer base. Other examples include a due-diligence platform with 10 workstreams that scans data rooms automatically and retrieves questions asked in previous deals—“I’ve seen far too many investors working weekends to do this”—as well as automating NAV calculations and reconciliations for fund operations.
• His central warning is that “many people understand AI only as a cost-cutting tool, but AI’s real value is often in creating revenue.” His signature question is: “If I gave you 10,000 college-educated employees for free right now, what would you do? What have you wanted to do but been unable to do?” A dairy company answered that it would write reports for every account. The result was a customized system generating health reports for every cow, “giving people back the time to actually maintain cow health—something that was not feasible before.”
• Investors in the AI roll-up era fall into 2 camps: instinctive avoiders who say, “We cannot invest in any sector with too much AI disruption risk,” and active adopters who acquire businesses that previously had limited technological sophistication and apply AI aggressively. Oliver cites one transaction as an example, but warns that “there is a huge gap between what people imagine can be achieved and what actually gets delivered in the real world.”

13. The 2 Deployment Traps, and a 3-to-5-Year Window for Consulting

• The most common mistake is trying to AI-enable everything at once. The first hurdle is the data platform. “Its value compounds, but no matter how smart AI is, without enough information and knowledge it can do nothing.” Data is scattered across Outlook, Gmail, ERP systems, and Salesforce. A PE-owned company may be running 4 different ERP systems; when it tries to understand why customers are leaving, it discovers the data is spread across 3 of them. Invisible addresses this with 4 modules: Neuron for the data layer and data mesh, Atomic for workflow automation, Synapse for evaluating AI performance, and Action for agent coordination.
• The second trap is “letting AI do things that should be deterministic.” “You would not want AI handling financial reconciliation. You want a deterministic result. AI can map out the logic of a workflow, but many execution steps should be hard-coded deterministic math.”
• Will consulting become obsolete? Oliver believes “consulting will see a wave of growth over the next 3 to 5 years.” Law firms are moving from hourly billing to outcome-based billing, and “the entire incentive structure has changed. For that kind of transformation, you need to talk to someone.” But “the people who truly unlock value are those who leave behind a transformed operating business.” The model of completing the work, leaving, and handing over a business that has actually been rebuilt is the real way to create value—not simply discussing how to transform.

14. The End State: Big Tech Has No Choice, and Humans Are AI’s Tentacles

• Google may also announce an FDE team, but Joe highlights the fundamental difference: big tech is bound to its own full stack. “It does not want to use Eleven Labs, Cartesia, or Deepgram for voice, and it wants to push Gemini for models.” Cresta, by contrast, may use more than 20 different models in a single conversation because “we sell outcomes, not tokens.” It has a strong incentive to choose the model that is “just good enough and priced right,” while also ensuring compliance and availability and preserving the option to switch to a self-hosted model when the network goes down. Big tech does not have that freedom of choice.
• Joe’s closing message is both a call to engineers and a clear statement of the human-machine division of labor. “More than 90% of our own code is written by AI,” but “you cannot fully trust AI to tune prompts. AI tuning AI can enter a state of overfitting.” Humans are “AI’s tentacles”: spending the day in customer meetings, meals, and coffee conversations to collect context AI cannot access. “You can’t expect a person to have coffee with AI.” AI can produce 5 or 6 options and a table full of pros and cons, but it cannot make the decision. A person still has to choose, then take responsibility for that choice—skin in the game.
• Not every engineer should become an FDE. Many will continue as cybersecurity or infrastructure specialists. But for people who are technically confident, do not mind travel or customer conversations, and “enjoy learning how a restaurant operates today and a dentist’s office tomorrow,” “FDE is like an entrepreneurial boot camp.” It is certainly not a comfortable way to work, but the skills it builds can be highly valuable for starting a company later.