The AI Adoption Paradox: Why Individuals Beat Enterprises

Two stories about AI are running at the same time, and they appear to contradict each other.

The first story: one person with a chat window can now do the work of twenty. Learn the tools, the story goes, and you can build anything, learn anything, ship anything. I mostly live inside this story. I run two companies where AI does most of the production work, and I would not trade my ten-person output for the fifty-person version of my life from a few years ago.

The second story: the largest technology companies on earth are spending billions of dollars on human engineers whose entire job is to sit inside enterprises and make AI work. Microsoft just committed $2.5 billion to embed 6,000 of them in customer organizations. OpenAI stood up an entire deployment business. Anthropic signed a $1.5 billion joint venture to do the same for financial firms. If AI were as easy as the first story claims, none of that money would need to exist.

This weekend the contradiction finally got named. Hamel Husain, an engineer whose work on AI evaluation I rate highly, pointed out that we are telling individuals they can do the work of twenty people by learning a chat tool, while telling enterprises they need billions in embedded engineering help because implementation is too hard to do alone. He called the two narratives incongruent, and the replies filled up with people picking sides. Aaron Levie compressed the whole argument into one line: “The battle in AI is shaping up to be a battle for context.”

Here is the position I want to defend: both stories are true, at the same time, about the same models. The same system that turns a solo founder into a small company produces a dead pilot inside a Fortune 500. And the variable that decides which story you get has nothing to do with model quality, prompt skill, or budget.

The variable is distance. Specifically, the distance between three things: the knowledge of what good work looks like, the authority to act on it, and the machine that now does the work. I call it context distance, and once you see it, you stop arguing about which narrative is right and start building around the mechanism. This post is the map.

Table of Contents

• The Paradox in Numbers: Both Sides Are Telling the Truth
• Context Distance: The Variable That Explains Both Stories
• The Individual Is the Integration
• What the Deployment Billions Actually Buy
• Adoption Is a Rep Count: The Two Loops
• The Window: How Long the Arbitrage Lasts
• The Founder Playbook: Selling Into the Gap
• The Contrarian Take: The 20x Story Is Also Selling You Something
• What to Do Monday Morning
• FAQ: The AI Adoption Paradox

The Paradox in Numbers: Both Sides Are Telling the Truth

Start with the enterprise side of the ledger, because it is brutal and well documented.

MIT’s NANDA initiative studied enterprise generative AI programs in 2025 and produced the number that has been quoted in every boardroom since: 95 percent of pilots showed no measurable impact on profit and loss. One industry tally put global enterprise AI investment at $684 billion, with more than $547 billion of it, roughly 80 percent, failing to deliver the value it was bought for. Gartner projects that over 40 percent of agentic AI projects will be cancelled by the end of 2027, and industry trackers put the share of AI agent pilots that never reach production near nine in ten. The average journey from working prototype to production system runs about eight months. In large enterprises, moving from pilot to implementation takes nine months or longer. Mid-market firms manage it in roughly 90 days, and the difference between those two numbers is going to matter a great deal in a moment.

Now the individual side, which is just as well documented and points the opposite way.

By 2024, Microsoft’s own workplace research found that 78 percent of AI users were bringing their own tools to work. A 2026 PagerDuty survey found two thirds of office professionals have used AI tools at work without approval. Other studies put the share of employees who install AI tools without consulting IT at 54 percent, and the share of American workers who use AI without telling their employer at 45 percent. Nearly half of generative AI use inside companies happens through personal accounts, invisible to the enterprise, and only about 30 percent of organizations claim full visibility into how their people use AI. The tools failed the security review, and usage went up anyway. People do not fight this hard for software that does not work.

And at the extreme end of the individual story sit the new company benchmarks. Cursor reached roughly $2 billion in annual recurring revenue with a team of about 50, which is in the neighborhood of $40 million per employee. Midjourney built about half a billion dollars of annual revenue with around 40 people and zero dollars raised. A well-run traditional SaaS company generates $150,000 to $250,000 per employee. The AI-native small team is not 20 percent better on this measure. It is two orders of magnitude better.

Put the two datasets side by side and the shape of the paradox gets sharp:

Here is the detail that should stop you: it is the same model on both sides of the table. Often the same person on both sides of the table, succeeding at home under a personal account and failing at work inside the sanctioned pilot. The weights did not change on the commute.

When the same input produces opposite outputs, the difference is the system around it. AI did not create that difference. AI priced it. For the first time, the cost of organizational structure shows up as a direct, visible multiple on the return of a specific technology, and the multiple is savage.

Context Distance: The Variable That Explains Both Stories

For AI to produce value on a task, three things have to meet: the knowledge of what good looks like on that task, the authority to act on the result, and the execution engine that does the work. The models collapsed the price of the third one. A competent draft, a working script, a first-pass analysis: execution is now cheap and instant. But the models did nothing to move the other two. Knowledge still lives where it always lived, in specific people’s heads and habits. Authority still lives where it always lived, in roles, permissions, and committees.

Context distance is the number of hops a task has to travel for those three things to meet. Count every translation, every approval, every queue between the person who knows what good looks like and the system that does the work. That count, more than any benchmark score, predicts what AI will return.

Every hop charges a toll, and the tolls compound. A translation hop loses fidelity: knowledge gets flattened into a ticket, the ticket into a requirements doc, the doc into a prompt written by someone who never did the job. A permission hop adds a gate with a calendar: security review, procurement, legal. A queue hop adds waiting, because the meeting to schedule the meeting is real. And an incentive hop adds noise, because somewhere along the chain sits a person whose job the automation touches, and nobody sprints toward their own automation.

Walk the four positions on the map:

Distance zero: the solo founder. Knowledge, authority, and execution share one chair. The prompt is a direct transfer from the person who owns the standard to the machine that does the work. The decision to ship is made by the person who sees the output. Change management is you, deciding, this morning. This is the position from which all the 20x stories are told, and from this position they are not exaggerated.

Distance one: the small team. The knowledge sits one desk away. There is a handoff, but the person who knows and the person who ships share standards, context, and often a room. Loss is small. Speed survives. This is where well-run startups live, and it is the position I try to defend at my own companies as they grow.

Distance three: the functional organization. Knowledge lives in operations, authority lives in management, and the AI initiative lives in IT. Connecting them takes tickets, briefs, and sponsors. Weeks pass between the question and the attempt. The pilot is now a project, with a project’s politics.

Distance five and beyond: the enterprise. Knowledge is fragmented across silos and locked behind permissions that exist for real legal reasons. Authority is distributed across committees precisely so that no single person can take a large risk. Execution is assigned to a center of excellence or an outside firm. The distance is now so long that a new job title exists specifically to walk it on foot.

That job title is the forward deployed engineer. An FDE is a person you rent to carry context across an organization by hand: they embed with the customer, dig the knowledge out of the silos, negotiate the permissions, encode what good looks like, and deliver it to the model. Palantir invented the role in the early 2010s because it learned that big-organization context never travels on its own. In 2025, postings for the role grew more than 800 percent, and it is now among the hottest titles in software. The market did not invent a new engineer because models got worse. It invented one because everyone finally noticed where the cost actually sits.

One boundary note, because I have written around this territory before. The production gap is about the vertical journey of a single AI system from demo to production, the five gates a build has to cross. The context engineering playbook is about the technical craft of feeding an agent the right information at the right time. Context distance is the third axis: the horizontal geometry of the organization around every system it tries to adopt. A company can pass every production gate, with perfectly engineered context, and still get a fraction of the return of a founder at distance zero, because the geometry eats the value before it lands. Wrapper depth is about your product. Model churn is about your vendor. Context distance is about you.

The Individual Is the Integration

Read the shadow AI numbers again, because they are the paradox in miniature. Two thirds of office professionals have used unapproved AI at work. More than half never asked IT. Nearly half of enterprise AI usage flows through personal accounts the company cannot see. While the official pilot spends its third month in security review, the building is already full of people quietly succeeding with the same models before lunch.

The enterprise calls this a governance problem, and it is one. It is also the largest natural experiment in software history, and it has a clean finding: remove the hops, and the value appears. Put the hops back, and it disappears.

At their own desk, on their own account, an employee operates at distance zero. They know what good looks like because they have done the task for years. They hold authority over their own drafts, because nobody reviews work that does not officially exist yet. And execution is one paste away. The sanctioned version of the same tool arrives with the geometry rebuilt: scoped to a workflow a committee chose, grounded in documents from a wiki nobody updates, wrapped in guardrails picked by someone who never did the task, measured by a dashboard that cannot see the actual work. The person did not get dumber between their kitchen table and their desk. The distance got longer.

I run this experiment on myself constantly. At ReBillion, my real estate transaction coordination company, AI runs most of the production work. That is not because I buy better models than a national brokerage can buy. It is because the standards those agents run on came out of my own hours doing the work with real estate teams, written down and wired directly into the systems. When a workflow breaks in the morning, the person who understands the failure, the person authorized to change the system, and the system itself are all within arm’s reach, and the fix is live by afternoon. A brokerage with 40 times my resources would route the same fix through a ticket, a vendor, and a quarterly release.

This is also why the standard enterprise response, training everyone on prompts, keeps disappointing. Prompt skill was never the binding constraint. Geometry was. Teaching a better prompt to a person who is five hops from authority is like teaching a faster sprint to someone standing in a queue.

What the Deployment Billions Actually Buy

Now look at where the smartest money in the industry went, because it confirms the map from the other direction.

In mid-2026, within a single stretch of weeks, the deployment economy went from a trend to an arms race. Microsoft launched Frontier Company, a $2.5 billion operating unit that embeds 6,000 of its own engineers directly inside customer organizations. OpenAI, which built its first forward deployed engineering team in 2024, was reported by May 2026 to have stood up a dedicated deployment business with more than $4 billion behind it. Anthropic announced a $1.5 billion joint venture with Blackstone and Goldman Sachs to embed engineers inside financial services firms. Accenture, the incumbent that has been selling this all along, reported $11.5 billion in cumulative advanced AI bookings across 11,000 projects, built an 80,000-person AI and data workforce, and then told investors it would stop reporting AI as a separate line because it is now simply the work.

Here is what none of that money buys: better models. Every company on that list resells access to the same handful of frontier systems. The billions buy something else. Line by line, they buy the tax bill of distance:

The last row is there on purpose. Some of the enterprise tax is waste, and some of it is load-bearing. Keep that distinction in mind for the contrarian section.

Satya Nadella gave the enterprise side of this thesis its vocabulary. In a long essay published in June 2026, read more than 28 million times, he argued that firms now compound two stocks of capital: human capital, the judgment and knowledge of their people, and what he called token capital, the AI capability a firm builds and owns, its workflows, its data, its evaluations, and the loop between the two. Without human direction, he wrote, you have compute running in circles.

Read cynically, the essay is marketing for the 6,000 engineers Microsoft would announce weeks later. Read structurally, it concedes the entire argument of this post: the value was never in access to the model, because everyone has access. The value is in the loop between what your people know and what your systems do. Enterprises are paying billions to construct that loop out of committees, connectors, and embedded engineers.

A solo founder wakes up inside it.

Adoption Is a Rep Count: The Two Loops

The second mechanism behind the paradox is speed, and it compounds harder than the first.

Adoption is not a purchase. It is a loop: notice a task, try it with AI, keep what works, kill what does not, fold the winner into how you operate. Every turn of that loop leaves behind either a working automation or a sharpened sense of what the tools can and cannot do. Both are assets. The return on AI is largely a function of how many turns of that loop you complete, which makes adoption a rep count.

An individual’s rep costs an afternoon and a few dollars of tokens. Mine happen between meetings: twenty in a good week, each one small, most of them failures, every failure cheap. An enterprise’s rep is a different organism. Vendor shortlist, security review, procurement, pilot cohort, steering committee, integration build, training, rollout. Gartner puts the average prototype-to-production journey at eight months; in large enterprises, pilot to implementation runs nine or more. That is not twenty reps a week. That is one rep, per workflow, per year. And when the rep fails, sunk cost keeps it breathing as a zombie pilot instead of letting it die as a lesson.

Run the compounding. A five-person company completing twenty small adoption loops a week banks roughly a thousand reps a year: a thousand lessons about where AI pays and where it lies. An enterprise completing one 9-month loop banks between one and two. The individual advantage was never a single 20x afternoon. It is a hundred-to-one difference in learning rate, applied to the fastest-moving capability shift in the history of the industry. Mid-market firms, at 90 days a loop, sit in between, which is exactly where their results sit too.

The vendor-side proof is Harvey, the legal AI company. At the end of 2025 it held $195 million in annual recurring revenue; by May 2026 it had passed $300 million, adding roughly $100 million in a handful of months, and it raised at an $11 billion valuation. More than 142,000 lawyers across 1,500 customers use it. Law firms could buy raw model access directly for a fraction of the price. They buy Harvey because it sells the loop pre-assembled for one vertical: the workflows, the evaluations, the interface where lawyers already work. That is Levie’s battle for context, fought as a product strategy: the applied layer is valuable precisely because it delivers a shortened context distance that the buyer cannot build quickly. The fastest-growing AI companies in the world are all, one way or another, selling distance.

There is a dependency cost to buying your loop from a platform instead of owning it, and I mapped it in the AI platform risk playbook. For now, hold the simpler point: whether you rent the loop, buy it, or build it, the thing being priced is distance, not intelligence.

The Window: How Long the Arbitrage Lasts

Every arbitrage invites its own closure, and it would be dishonest to sell you this one without an expiry analysis. Three forces are working to close the gap.

First, the deployment armies exist precisely to compress enterprise distance, and they will partly succeed. An embedded engineer who spends a year inside a bank does real work: context gets written down, permissions get renegotiated, dead pilots get buried. That is what the combined $2.5 billion, $4 billion, and $1.5 billion commitments are betting on.

Second, vertical products keep bundling context so enterprises can buy distance-zero as a subscription. Harvey for law. The same pattern is filling in for finance, health documentation, customer support, and code. Where a good bundle exists, the enterprise does not need to shorten its own geometry; it buys someone else’s.

Third, the token capital thesis is enterprises learning the right lesson: encode your knowledge into owned workflows, data loops, and evaluations, so the loop belongs to the firm instead of to the departing consultant. Some large companies will actually do this. Most will relabel a SharePoint folder and call it a knowledge strategy, but some will do it.

Against those three, two forces hold the gap open, and they are made of harder material.

Organizational physics does not repeal. Permissions exist because liability is real, privacy law is real, and blast radius grows with headcount. A 50,000-person company cannot give everyone authority; the gates are load-bearing. Committees can be shrunk but not deleted, and every surviving gate is measured in weeks. The enterprise can pay to walk its hops faster. It cannot pay to have no hops, because the hops are what being an enterprise means.

And incentives do not repeal either. The individual adopts AI for herself, on her own task, with the gains landing in her own evening. Enterprise adoption asks thousands of people to accelerate a change whose gains land on a slide they will never present. One of those two systems runs on its own fuel. The other needs 6,000 embedded engineers pushing it uphill.

So the honest forecast is a narrowing, not a closing. The gap compresses in commodity workflows: email, summaries, support macros, meeting notes, anything a vertical product can bundle. It persists, for years, wherever workflows are odd, judgment density is high, and headcount is low. Which is a precise description of where startups live. The durable version of the advantage is not that enterprises are slow this year. It is that a small organization can re-fuse knowledge, authority, and execution every single morning, at any budget, and a large one cannot, at any budget. Every hire you make is a hop you chose to add, which is why the hire versus automate decision is really a context distance decision, and why the judgment calls you refuse to delegate are the subject of the agent boss operating system.

The Founder Playbook: Selling Into the Gap

If context distance is the variable, the strategy writes itself in two directions: keep your own distance short, and sell to people whose distance is long. Six moves, in the order I would run them.

Move 1: read the deployment billions as a market signal. When Microsoft, OpenAI, Anthropic, and Accenture collectively point tens of billions of dollars at the same pain, they have done your market research for you. Deployment pain is the largest under-priced problem in software. You cannot out-hire their engineer armies, and you should not try. Build the thing that makes the army unnecessary for one narrow, deep use case.

Move 2: design for a distance-one install. If your product needs an embedded engineer to deliver value, you have inherited your customer’s geometry, and their nine-month loop is now your sales cycle. Design the wedge so a single team can adopt it: first value inside a day, no committee required, priced under the procurement threshold. The 66 percent of professionals already using unapproved tools are your beachhead, and they are self-identified: they have the knowledge, they lack sanctioned tools, and they have proven they will move without permission. Build the audit trail into the product from day one, so that when IT eventually finds you, the conversation is an upgrade, not a ban. That is exactly how the unauthorized tools of 2024 became the enterprise line items of 2026.

Move 3: compete against distance, not against products. Pick markets where the incumbent operates at distance five: their AI response to you is trapped in the same nine-month loop as every other initiative they run. If you ship weekly and they ship yearly, you bank fifty lessons for their one. Their headcount, their process maturity, the very things their org chart is proudest of, are the mechanism of their slowness. You are not outsmarting them. You are out-looping them.

Move 4: price the assembled loop, not the model call. Harvey’s growth is not a story about legal prompts. It is proof that buyers pay a premium for pre-shortened distance: workflows, evaluations, and interface, bundled. Charge for the outcome inside the workflow, and the model bill becomes your cost of goods, not your product. I wrote the pricing half of this argument in why per-seat pricing is dying.

Move 5: guard your own geometry as you grow. Growth is the process of adding hops, and most founders add them by default. Hire people who carry context rather than coordinate it. Write the standards down before you hire anyone, so the document, not a meeting, is the interface between your head and both your people and your agents. My internal AI stack post covers the tooling half; the principle is that anything worth doing twice gets a written standard, and every written standard is a hop you never have to add back as headcount.

Move 6: start compounding your own token capital now. Nadella’s term, but the small-company version is simple: workflows written down, evaluations that encode your taste, data loops that make your product better because you ran it yesterday. This is the asset that survives model releases, makes you portable across vendors, and reads as a moat in diligence. A founder who owns the loop can swap the engine; a founder who is the loop cannot take a holiday.

The Contrarian Take: The 20x Story Is Also Selling You Something

Everything above could be misread as a victory lap for the individual, so here is the part the 20x crowd skips.

The claim that one person now does the work of twenty is a claim about output, and output was never the scarce thing. Twenty people’s worth of drafts, code, and campaigns, absorbed by one person’s judgment, produces one person’s worth of decisions with twenty times the surface area for error. I wrote a whole post on how this plays out as the AI productivity paradox: volume rises, shipping stalls, because attention is the actual constraint. Distance zero makes you fast. It does not make you right.

Distance zero also has ceilings that no loop speed fixes. Distribution does not care how fast you iterate. Trust in regulated markets is bought with structure, insurance, and history, not with reps: an insurance carrier will not hand its claims process to a company of one, however brilliant, and I say that as someone who sells into regulated real estate workflows. The one-person company post maps these ceilings one by one: attention, trust, liability, and taste all stop scaling at different points.

And the sharpest edge: context in one skull is a single point of failure. The founder who never writes anything down has context distance zero and bus factor one. The same geometry that makes you unbeatable on Tuesday makes your company unownable, unsellable, and uninsurable as a going concern. If your knowledge, authority, and execution are fused in your head and only your head, you have not built a company. You have built a very productive dependency.

Some enterprise slowness, meanwhile, is the product. Audit trails, permission gates, four-eyes review: in money, medicine, and law, the customer is buying the gates. A startup whose pitch is that it skipped them does not have an arbitrage. It has a liability with good unit economics, and the market eventually prices those correctly.

So the honest synthesis: returns on AI are a function of context distance, and distance is a choice with tradeoffs in both directions. Enterprises are spending billions to learn how to shorten theirs. The founder’s mirror-image mistake is accidentally lengthening yours, hire by hire, tool by tool, until you wake up one morning running a small enterprise with a startup’s payroll. The window is real, and it rewards the builders who exploit it while encoding their context, because encoded context is the only version of this advantage that survives your own growth.

What to Do Monday Morning

Run a context audit on your five most valuable workflows. Thirty minutes. For each workflow, write three names: who knows what good looks like, who can authorize a change, who or what executes. Count the hops between them. That count predicts your AI return better than any model benchmark you will read this quarter.

Collapse one hop before Friday. Find the workflow where knowledge and authority already sit in the same person, hand that person an agent, and give them standing permission to ship. Measure time to first value in hours. Do not start with the workflow that needs a committee; start with the one that needs nobody’s permission, and let the result argue for you.

Write one context document. One page, for your single best workflow: the inputs, the standard including what gets rejected and why, the failure modes you have personally seen. Wire it into an agent’s instructions and run it. That page is your first unit of owned token capital, and writing it will teach you how much of your standard was never actually articulated. The context engineering playbook covers the full craft when you are ready to go deeper.

If you sell to companies, measure your install distance. Two numbers: hours from signature to first value, and count of humans who must say yes before a user touches the product. Kill one approval in the next thirty days, through product design, packaging, or pricing. Your growth rate is downstream of your buyer’s loop speed, not just your feature list.

Count your weekly reps. How many AI experiments did you run in the past seven days: tried, judged, kept or killed? If the answer is under five, you have built an enterprise-shaped loop inside a company with no enterprise to blame for it. Fix the loop before you buy another tool, and when a rep fails, check it against the production failure list before concluding the model is the problem.

FAQ: The AI Adoption Paradox

What is the AI adoption paradox?

The AI adoption paradox is the simultaneous truth of two opposing narratives: individuals report doing multiples of their previous output with off-the-shelf AI tools, while roughly 95 percent of enterprise AI pilots show no measurable return and companies spend billions on deployment help. Both are accurate observations about the same models. The paradox resolves once you measure context distance: the number of hops between the knowledge of what good looks like, the authority to act, and the system doing the work.

What is context distance?

Context distance is the count of translations, approvals, and queues a task crosses between the person who holds the standard for good work, the person with authority to act on the output, and the AI system that executes. A solo founder operates at distance zero because all three sit in one chair. A large enterprise often operates at distance five or more, with knowledge fragmented across silos, authority spread across committees, and execution assigned to a separate team. AI returns fall as distance grows because every hop loses fidelity, adds waiting, and introduces misaligned incentives.

Why do individuals get more value from AI than enterprises?

Because the individual is the integration. One person using AI on their own work carries the domain knowledge, holds full authority over the result, and executes immediately, so nothing is lost in translation and no gate adds delay. An enterprise has to reassemble those three elements across silos, permissions, and committees before any value appears, which is slow, lossy, and expensive. The shadow AI numbers make the point: two thirds of professionals use unapproved AI tools productively at work while their employer’s official pilot stalls in review.

What is a forward deployed engineer and why is the role suddenly everywhere?

A forward deployed engineer, or FDE, is an engineer who embeds inside a customer organization to make AI systems work in production: extracting knowledge from silos, negotiating data access and permissions, encoding quality standards, and integrating models into real workflows. Palantir invented the model in the early 2010s. Postings for the role grew more than 800 percent through 2025 as Microsoft, OpenAI, and Anthropic collectively committed billions to embedded deployment teams. The role exists because enterprise context does not travel on its own; an FDE is a person paid to walk it across the organization by hand.

Does the paradox mean enterprises should stop investing in AI?

No. It means sequencing the investment differently: shorten distance before buying scale. The enterprises seeing real returns tend to grant small teams direct authority to adopt AI on their own workflows, buy pre-assembled vertical products instead of raw model access where good bundles exist, and measure time to first value per team rather than pilot headcount. Buying more capability without shortening the geometry produces exactly what the data shows: nine-month loops and dead pilots.

How long will the individual and small-team advantage last?

Expect narrowing, not closing. Deployment teams and vertical products will compress enterprise distance in commodity workflows like support, summaries, and documentation over the next few years. The advantage persists where workflows are unusual, judgment density is high, and headcount is low, because a small organization can re-fuse knowledge, authority, and execution daily while a large one cannot at any budget. The permanent part of the edge is loop speed: a small company completes hundreds of adoption experiments in the time an enterprise completes one.

How do founders keep context distance low while scaling?

Treat growth as the discipline of adding as few hops as possible. Write standards down before hiring, so documents rather than meetings carry context. Hire people who carry context into new territory rather than coordinators who move it between others. Give the person closest to the work authority to change it. And encode your judgment into evaluations and workflows that agents and employees both run against, so the standard lives in the system instead of in your calendar.

Is shadow AI good or bad for a company?

Shadow AI is a signal before it is a risk. It tells you your people have already found distance zero and proven value your official program has not delivered; roughly half of enterprise AI usage now flows through personal accounts. The data exposure is real and needs governing. The mistake is responding with bans that rebuild the distance that caused the shadow usage in the first place. The better response is to formalize the winning use cases, provide sanctioned tools with audit trails, and keep the geometry that made them work.

I write one deep playbook like this every day at the intersection of AI, entrepreneurship, and personal growth. If this one reframed the adoption debate for you, the AI-native founder playbook is the parent guide, and the production gap is the sibling piece on why pilots die even when the geometry is right.