AWS just set up an FDE division: a billion dollars to send thousands of engineers into client companies. Quite a production. OpenAI wasn't about to be outdone, putting $4 billion into its Deployment Company. Anthropic teamed up with Blackstone to form a consulting company too. GenerativeX had a more modest $4 million Series A. All of them are betting on FDE.
Chinese VCs caught the scent immediately. Now they're scouring the world for “China's Palantir.” One company positioning itself as a Palantir counterpart had its IPO oversubscribed 623 times—investors piling in as if the shares were free. MiraclePlus made it even easier: FDE and AI consulting went into the same investment category. They couldn't even be bothered to separate the labels. Apparently FDE is the miracle cure for AI commercialization, the next sexy thing after models, the new era's know-how moat.
Is it?
These so-called FDE businesses are taking people's money and going nowhere. The problem isn't difficulty or a need for more time. Physics won't allow it.
FDE = archaeology + fortune-telling
What do these FDEs do? It comes in three levels.
Level one: pure manual labor. Engineers work on-site, sit beside the client, and watch them use the software. They take notes and write a daily report at night. It's on-site outsourcing from twenty years ago, dressed up as “AI Deployment.”
Level two: manual labor with tools. Add some fancy stuff: session replay, AI interviews, user behavior analytics. Tools like trooly and cookiy conduct structured interviews with end users. Userpilot's Lia helps PMs look through session recordings. It sounds more sophisticated, like upgrading an archaeologist's probe to ground-penetrating radar. You're still digging up dirt.
Level three: manual labor, tools, and packaging. Turn what you've collected into an industry insights report. Show it to investors: “We're doing knowledge distillation.” Pull a few lessons from client A's custom implementation and paste them into client B's POC. The two clients happen to have somewhat similar needs? Announce that you've discovered a general pattern and are turning it into a platform. Double the valuation in the next round. One step closer to AGI!
All three are essentially the same thing: hoping to extract useful know-how from scattered snippets of partial, local context and user behavior. Basically archaeology.
An archaeologist digs up pottery shards, studies the patterns, and guesses whether the vessel held water or grain. A faster shovel and a larger excavation still leave you with fragments. You're using five percent of the pieces to reconstruct the other ninety-five percent.
VCs now call that knowledge distillation, a strategy, an investment category.
Ted Mabrey, Palantir's global head of commercial business, wrote an article in 2024 called “Sorry, that isn't an FDE.” It's free and under three thousand characters. He called every imitator half-baked: they copied the form without copying the function. His point, roughly, was that they'd taken only half a step and trapped themselves in low-value, low-complexity, extractive business models, with disappointing jobs for employees and disappointing cash flows for shareholders.
VCs, PE investors, LP partners: did any of you spend five minutes reading this free, 3,000-character death sentence for your portfolio companies? Or were you too busy—reposting “The AGI era is here” on WeChat Moments, talking about “structural opportunities in AI-native businesses” at conferences?
Why today's FDEs are all guesswork
Because archaeology can't capture the actual workflow. That's a physical constraint.
How much information can an FDE capture? Three months on-site, at most three hours of real observation a day. They see what the user does on the screen in front of them. Everything after the user leaves—the revision they make at home that night, the collaboration on Lark after closing the product, what they do next after receiving an email—is lost.
What does a real user workflow look like? Open Excel to review a report. Spot an anomaly. Switch to the browser to research it. Switch to Claude to ask a question. Try two versions and delete one. Copy the answer into Lark. A colleague says it's wrong. Go back and regenerate it. Paste it into an email. Send it to three people.
That spans seven applications and several parts of the day. An FDE watching during one three-hour window can catch two or three segments at best, and only the actions on the screen. They miss everything in the user's head: the versions tried and deleted, what the user thought about a colleague's feedback, why they changed the third paragraph but left the first alone, why they copied a particular person on the final email. That's where much of the workflow lives, and none of it gets captured.
They capture less than five percent of the information. That's a generous estimate.
Do the math: an FDE on an annual salary of around 500,000, stationed on-site for three months, costs 120,000 to 150,000 in salary alone for a single client. That doesn't include management, travel, recruiting, or replacing people who leave. But money isn't the real issue. The worst part is spending over a hundred thousand to get five percent of the fragments. How do you fill in the other ninety-five percent? Meetings? Interviews? PMs and FDEs relaying messages back and forth? The client calls to complain that the thing doesn't work, so you send someone back to patch it? You might as well ask an AI fortune-teller.
Five percent raw data, ninety-five percent subjective guesswork. And you call that knowledge distillation? Excuse me?
Michael Burry voted with his money. In November 2025, Palantir's stock hit an all-time high of $207. At the same time, Burry disclosed $912 million in PLTR put options through Scion, valuing Palantir at $46—less than a quarter of its share price at the time.
PLTR CEO Alex Karp called Burry's bet “batshit crazy” on CNBC. A year later, PLTR was down to roughly $112, forty-five percent below its peak. June 2026 was Palantir's worst month since going public: a twenty-five percent drop in a single month.
Burry sees FDE as a low-margin consulting business, not high-growth SaaS. Palantir took twenty years to reach $5 billion in revenue. Anthropic went from $9 billion to $30 billion in ARR in four months. When simpler AI platforms can get into enterprises at a lower cost, who still needs something that can only be delivered by sending engineers on-site?
Anyone still hyping FDE and throwing money at it is an idiot.
The money follows from the underlying problem: why can they capture only five percent?
They're watching from the wrong place: a chair in the client's office. The workflow happens between Excel, Lark, business systems, email, and the product, across browser windows, between Ctrl+C and Ctrl+V, or in a flash of insight at eleven at night. Watching inside the FDE's product misses all of that.
From that chair, the FDE sees only the few sessions when the product is open. Their vantage point limits them to fragments; everything else is outside their field of view.
Why doesn't Palantir have this problem? Its client is the CIA. All the data lives in one closed system. Users have no second application to switch to. Sitting beside them is enough. Can these Chinese FDE companies find clients like the CIA? Your clients jump between a dozen applications a day. Everything you see is a fragment. The only difference from on-site outsourcing twenty years ago: it costs ten times as much.
Enough about “doing FDE better.” The premise is wrong. Watch from one fixed location, and you can only see what happens during the few minutes you're present, however many applications and tools the full workflow spans over time. You get five percent in fragments and guess the rest. Fortune-telling.
Perception: useful context lives outside the model's window
What information do you need in the first place?
You need the complete workflow of how users accomplish something throughout the day, beyond the record of what they do inside your product: the problems they encounter, the tools they use, how they solve those problems, how many revisions they make, and where the final result goes. That spans applications, tools, and time.
How do you get it?
Send someone to sit beside them? A slice. Instrument session replay? A slice. Use AI for structured interviews? A slice with an AI wrapper.
These methods share exactly the same limitation: the capture layer is either a person or something inside a single application. It can only see fragments from one specific space, time, and product. Better tools won't fix it. The architecture has already set the ceiling.
We need the full picture.
There's only one way to get that full picture: build the capture layer on the operating system.
With capture at the OS level, you can see the complete stream of user behavior throughout the day, across applications, sessions, and situations. For every click, edit, undo, copy and paste, application switch, and AI interaction, you can follow what they entered, what was generated, which version they kept, what they changed, what they rejected, and where they ultimately used it. Capture inside your product gives you only snippets of that trail.
It's a dynamic map of how they work, a complete digital twin of their actual behavior. Session replay, interview notes, and on-site daily reports can't give you that.
And you don't have to ask the user a single question. They open that Airtable at 9:03 every morning, copy from A to B, and export to C. They try three answers and keep only the final paragraph. Users won't mention these things in interviews because they don't think they're important, or don't even realize they're doing them. The perception layer sees them and keeps watching.
Workflow distillation needs a sensor passively collecting data in the background around the clock. An archaeologist sitting beside someone with a probe can't do the same job.
Guessing versus seeing. Archaeology plus fortune-telling versus a real-time digital twin.
Models + perception will eat FDE
FDE exists on one premise: AI products aren't good enough yet. Models can't deploy themselves, adapt themselves, or understand user workflows on their own. So people have to fill the gap.
That premise is disappearing.
Model capabilities are growing exponentially. The perception layer supplies the complete stream of user behavior. When both exist, the FDE middle layer is no longer needed.
The model gets all the behavior captured by the perception layer directly, beyond the few questions you ask it. Every action, decision, and AI interaction over the past three months, including the full loop of what followed, becomes model context.
The model sees the user's workflow for itself, without on-site staffing, session replay, or AI interviews.
One of Burry's arguments for shorting Palantir is that simpler AI platforms like Anthropic are eating its business. He's telling only half the story. The other half: give models perception, and you don't even need Anthropic-scale infrastructure. Models run directly on the user's OS. The perception layer collects directly on the user's OS. The FDEs, deployment tools, consulting teams in between—delete them all.
Models plus a perception layer eat every FDE.
FDE isn't an investment category. It's an expensive temporary patch for a world where the models aren't good enough and the perception layer hasn't been built yet.
It was never a strategy either, just another so-called AI application in VC vocabulary: a hype narrative, a placebo. Nobody cares whether the product works or real know-how accumulates. They're not the ones using it. Its only purpose is to pump portfolio valuations, cash out, and walk away.
FDE—physics won't allow it. PowerPoint will.