He didn’t notice it at first.

The sprint velocity targets went up. The expected turnaround on feature requests got shorter. His backlog grew by a third in two months. When he raised it with his manager, the answer was casual — almost cheerful: “We just onboarded Copilot and Claude Code for the whole engineering team. We should be moving faster now.”

He’s a product manager at a mid-sized tech company. His job is to translate between business goals and engineering capacity. And overnight, the translation changed. Leadership had purchased AI coding tools and assumed the equation was simple: same team plus AI equals more output. His calendar didn’t get lighter. It got heavier — because now he was expected to scope more features at the same pace, for a team that was still figuring out how to use the tools they’d been handed three weeks ago.

When he pushed back — explaining that AI tools require workflow redesign, that the team needed training, that “faster” doesn’t mean “more” — management’s response was three words that captured a $684 billion problem: “Just add the chat bubble to the app.”

A button. A bubble. A little drawing on a screen.

He didn’t know it then, but he had just described what 80% of AI adoption looks like from the inside — and two of the biggest AI companies in the world had just spent $11.5 billion confirming he was right.

The Numbers Nobody Wanted to Publish

In 2025, global enterprises invested $684 billion in AI initiatives. According to RAND Corporation’s analysis, 80.3% of those projects failed to deliver their intended business value.

That number deserves to sit for a moment. Eight out of ten AI projects — funded, staffed, launched — did not do what they were supposed to do.

The breakdown is instructive. Of the failures, 33.8% were abandoned before reaching production — they died in pilot, in proof-of-concept, in the meeting where someone finally asked “what problem are we solving?” Another 28.4% reached completion but failed to deliver expected value — the system worked, technically, but nobody used it, or it solved the wrong problem, or the workflow it was supposed to improve had already adapted around it.

The percentage of organizations that abandoned at least one AI initiative jumped from 17% to 42% in two years. Financial services — the industry that spends the most on AI — leads the failure rate at 82.1%. Healthcare is at 78.9%. Manufacturing at 76.4%.

And here’s the finding that explains nearly everything else: projects with dedicated change management resources succeed at 2.9 times the rate of those without. The difference between the 20% that work and the 80% that don’t is not the model. It’s not the compute. It’s not the data. It’s whether someone prepared the humans.

The Button Problem

The decision-maker who thinks AI is a button represents a specific failure mode that the industry has been reluctant to name: leadership AI illiteracy.

It’s not that these executives are stupid. Many of them built successful companies, navigated complex markets, made decisions under uncertainty for decades. But AI is not like previous technology adoptions. When you gave a team Excel, they could start using it the next day — the spreadsheet metaphor was intuitive, the learning curve was manageable, the output was visible. When you gave a team a project management tool, the workflow change was significant but bounded — you could see where the old process ended and the new one began.

AI doesn’t work that way. AI changes the shape of work itself. A coding assistant doesn’t just speed up typing — it shifts what a developer spends time on, which changes what code review means, which changes how technical debt is managed, which changes how projects are scoped. A customer support agent doesn’t just answer faster — it changes what human agents handle, which changes hiring profiles, which changes training programs, which changes the entire support cost structure.

The executive who says “add the bubble chat” is applying the old software mental model to an AI problem. Install the tool, train the team, measure the output. But AI doesn’t have a bounded workflow change. It has a cascading one. And if leadership doesn’t understand that — if they think they’re buying a button when they’re buying a process transformation — the project will join the 80%.

There’s a failure mode the statistics don’t capture: shame. When a new tool doesn’t work the way someone expects, most people don’t raise their hand. They try again. And again. And again — silently creating duplicates, workarounds, and errors that compound before anyone notices. In one real deployment we observed, an operator created four duplicate records without telling anyone because she was embarrassed to ask her boss how the system worked. The problem wasn’t the tool. It wasn’t the training — there was no training. It was that nobody had created a safe space to say “I don’t understand.” Change management isn’t just a process framework. It’s permission to be confused out loud.

The Confession

On May 4, 2026 — the same day, within hours of each other — OpenAI and Anthropic each announced enterprise consulting ventures.

OpenAI launched “The Deployment Company,” a $10 billion vehicle backed by TPG and 19 investors, with a guaranteed 17.5% annual return. Led by COO Brad Lightcap, it is already in talks to acquire three AI services firms. OpenAI simultaneously deepened partnerships with McKinsey, BCG, Accenture, and Capgemini — the same consulting giants that charge $500 per hour to tell companies what their own employees already know.

Anthropic announced a $1.5 billion joint venture with Goldman Sachs, Blackstone, and Hellman & Friedman, explicitly modeled on Palantir’s forward-deployed engineering approach: embed engineers inside the client company to redesign workflows from within. In parallel, EPAM launched a program to train 10,000 “Claude-certified architects” and 250 specialized “Black Belt” engineers.

Two companies. Same day. Same admission: the model alone is not enough.

This is not a pivot. It’s a confession. For four years, frontier AI labs sold models — APIs, subscriptions, tokens. The implicit promise was that the model would do the work. Just plug it in. The smarter the model, the more value it creates. Scale the model, scale the value.

The 80% failure rate is the receipt for that promise.

What They’re Actually Selling

Look past the press releases and the partnership announcements. What OpenAI and Anthropic are selling now is not consulting. It’s not “AI transformation services.” It’s not even expertise.

They’re selling the manual that should have come with the product.

When you buy a car, the manufacturer assumes you can drive. If 80% of buyers crashed within a year, the manufacturer wouldn’t say “our car is fine, the drivers are bad.” They’d either redesign the car or start selling driving lessons.

The frontier labs chose driving lessons. But they didn’t call them driving lessons — they called them “enterprise AI services ventures” and priced them at $10 billion.

The genius of this move — and it is genuinely clever — is that it converts a product failure into a revenue stream. The model doesn’t work in enterprise? That’s not a problem. That’s a market. Sell the model. Then sell the consulting to make the model work. Then embed your engineers inside the company so deeply that leaving means ripping out not just the model but the entire workflow that was rebuilt around it.

This is lock-in at a depth that SaaS companies could only dream of. Salesforce embeds in your sales pipeline. Oracle embeds in your database. OpenAI and Anthropic are embedding in your operations, with their own engineers sitting next to your employees, redesigning how your company does its actual work.

The Palantir Pattern

The comparison to Palantir is not incidental — both Anthropic and OpenAI explicitly reference it. And the Palantir playbook is instructive.

Palantir’s forward-deployed engineers don’t just install software. They learn the client’s business, identify where Palantir’s tools can add value, build custom integrations, and become the institutional knowledge holders for how the AI system works within that specific organization. Over time, the Palantir engineer becomes indispensable — not because the software is irreplaceable, but because the engineer is the only person who understands how the software connects to the company’s unique processes.

When the contract comes up for renewal, the question isn’t “is this software worth it?” The question is “can we afford to lose the person who understands how everything works?”

That’s the model OpenAI and Anthropic are adopting. Not software licensing. Not API pricing. Organizational dependency. The kind of lock-in where the exit cost isn’t a migration — it’s a reorganization.

What This Means for the 80%

The uncomfortable truth is that the consulting ventures will probably improve enterprise AI adoption rates. Embedded engineers who understand both the model and the client’s business will produce better outcomes than a decision-maker who thinks AI is a button. The projects will succeed more often. The ROI will materialize. The 80% failure rate will come down.

And that’s exactly the problem.

The improvement will partly come from enterprises developing their own AI capability — people do learn through use, and the learning curve is real even if it’s slow. But the consulting model compresses that curve by renting expertise instead of building it. The company learns faster, but the deepest knowledge — how the model connects to the specific workflow, why certain configurations work and others don’t — lives in the consultant’s head, not in the client’s org chart.

When the next model generation arrives — and it always does — the company will need the same embedded engineers to adapt the new model to the same workflows. The consulting contract renews. The dependency deepens. The lock-in compounds.

The 80% failure rate was a problem. The consulting ventures solve it. But they solve it the way a drug solves a symptom: by creating a dependency that requires the same drug indefinitely.

The Alternative Nobody’s Selling

There is another path, and it’s the one nobody is packaging into a $10 billion vehicle: build the capability internally.

The companies that fall in the 20% that succeed have one thing in common — not better models, not bigger budgets, but dedicated change management. Someone inside the organization whose job is to understand the tool, redesign the processes, train the people, and iterate on the implementation. Not a vendor’s engineer. Not a consultant who leaves after the engagement. Someone who stays, who accumulates context, who understands why the company does things the way it does.

The irony is sharp. The AI labs are building forward-deployed engineering teams because they know that context — deep, specific, institutional context — is what makes AI work in practice. And they’re right. But they’re selling that context as a service instead of helping companies build it for themselves.

A developer in Santiago runs seven AI instances on a mini-PC that costs less than one month of enterprise consulting. He pinned his CLI version, disabled auto-updates, set his own reasoning effort levels, and built a communication protocol between instances using plain text over HTTP. No vendor. No consultant. No embedded engineer. Just someone who decided to understand the tool instead of waiting for someone else to understand it for him.

That’s the 20%. Not a budget. Not a partnership. A decision.

The Needle in the Haystack

The product manager who described his company’s understanding of AI as “a button in the app” didn’t know he was describing a $684 billion problem. He was just venting about his job.

But that frustration — the gap between what leadership thinks AI does and what it actually requires — is the market that two trillion-dollar companies just spent $11.5 billion to address. Not because they care about the product manager’s frustration. Because they realized the frustration is a revenue stream.

The 80% failure rate was the best thing that ever happened to frontier AI labs. It proved the model wasn’t enough. And in proving the model wasn’t enough, it created the market for everything else they wanted to sell.

Why They Really Fail

Underneath the statistics, underneath the consulting ventures and the lock-in strategies, there is a simpler explanation for why 80% of AI adoption fails. It’s not technical. It’s not financial. It’s philosophical.

The companies that fail ask: “How many people can we replace?”

The companies that succeed ask: “What can these people do now that they couldn’t before?”

The first question leads to the button — install the tool, reduce headcount, increase margin. AI as cost reduction. The decision-maker who says “add the chat bubble” is answering the first question. He doesn’t need to understand the tool because the tool’s job is to eliminate the need for understanding. Plug it in, cut the team, report the savings.

The second question leads to infrastructure. It leads to a developer who pins his CLI version because he wants the model to do exactly what he needs, not what the default provides. It leads to a fleet of nine AI instances distributed across three machines, each one specialized, each one accumulating context that makes it more useful over time. It leads to an operator in a notary office who — after failing four times in silence — gets a dedicated support channel built specifically so she can ask questions without shame. Nobody got replaced. The system gained capabilities it didn’t have before.

The 80% treats AI as a replacement for human judgment. The 20% treats it as an extension of human capability. The difference isn’t the model. It isn’t the budget. It’s what the person at the top believes they bought.

The confession is on the table. The question is whether enterprises will hear it as a warning — prepare your people, or someone else will do it for you at a price you’ll never stop paying — or as an invitation to sign the next contract.