Almost every company I know announced its AI strategy last year. They bought ChatGPT for the team, ran a couple of workshops, dropped the word AI into the earnings call. Ask them today what changed and almost none can point to a single dollar of return.
What caught my attention wasn't that AI failed. It's that it didn't. The companies did.
I watched a video by Freddy Vega, the founder of Platzi, that puts a finger right on that wound. And the interesting part is that Freddy is not an AI skeptic. He's one of the most bullish people you'll find. His point is more uncomfortable: the tool works, what doesn't work is how they're installing it. And right there, in that gap, is the asymmetry almost nobody names: the same tool that produces nothing inside a two-thousand-person company is the cleanest lever a solo builder has ever had.
##Everyone bought it. Almost nobody got anything.
Adoption is record-breaking. In 2017, fewer than 20% of companies used AI. By 2025 it's more than 80%, and by the end of this year it will be pinned near 100%. Everyone bought it.
The return tells a different story. The MIT report that went viral last year, "The GenAI Divide," found that of the companies that invested in custom generative AI, roughly 95% saw no measurable return on their P&L. Only about 5% achieved real revenue acceleration. The exact number is contested (it isn't peer-reviewed and several academics have pushed back), but the direction isn't: almost nobody is capturing the value.
And many are already pulling out. According to S&P Global, the share of companies that abandoned most of their AI initiatives jumped to 42% in 2025, up from 17% the year before. The euphoria is turning into fatigue.
Freddy sums it up with a line that stings: three quarters of corporate AI strategies are pure theater. We implement AI so the boss can say we implemented AI, so we can tell the press we implemented AI.
The gap between the 80% that bought it and the 5% that benefit is the whole story. This isn't an adoption problem. It's a problem of what happens next.
##The tool works. Watch where.
Before blaming the model, look at where AI is actually delivering, because it is.
The cleanest case is coding. You hand an AI a set of instructions, it takes control of the computer and builds complete solutions. This genuinely works. At Uber, AI coding tools went from a fraction of the company to nearly every engineer. And yet Uber's COO came out to say it's getting harder to link that spend to real improvements for users: "that link is not there yet." They had already burned through their 2026 Claude Code budget ahead of schedule. It works, and they still can't justify it.
The other case is customer support, the "hello world" of AI in companies. In the a16z and Pylon data Freddy breaks down in the video, AI shines far more as a copilot than as a replacement: it filters, triages, and hands the hard ones to a human. Companies born with AI at their core get much more out of it than traditional ones running the same systems. And the bigger and more valuable the customer, the more a human stays in the loop. On small accounts AI resolves plenty. On accounts over a million dollars, you're almost always the one answering.
Same model. Different container. When the result swings that much based on who runs it, the problem stopped being the tool.
##Why it dies inside companies
The good part of the video is that it doesn't stop at the easy diagnosis of "they need more training." It goes deeper. I'll group the reasons here, because they all point at the same thing.
They treated it as a purchase, not a process change. Buying licenses isn't transformation. Redesigning a process around AI is. Most did the first and waited for the result of the second. It's like handing someone a car who still thinks like a pedestrian. The problem isn't the car. MIT calls it the "learning gap": neither the tools nor the organizations retain context, learn, or improve over time. The bottleneck is organizational, not the model.
They got burned before, so they stall. Freddy makes a point few people say out loud: a lot of companies pattern-match AI to the fads that let them down, badly implemented remote work, blockchain, the metaverse, and decide to wait. Waiting is exactly the mistake, because this time the tool actually delivers.
The gains are invisible to the income statement. Martínez used to spend 12 hours building a dashboard and now spends 5 minutes. That saving lives in Martínez's afternoon, not in the financial report, and Martínez has no incentive to show it off. This isn't an anecdote: MIT documented a shadow AI economy where employees at more than 90% of companies use personal AI for work, while only about 40% bought an official license. The value already arrived. Just not through the org chart.
The wrong owner measures it. Usually HR hands out the accounts. But the gain lands in other teams' metrics, HR only sees an average cost, and the CEO sees a line item and cuts it. The improvement existed. The person measuring it didn't own it.
Fear leads to pinching pennies. Corporate AI spend is tiny as a share of revenue, well under 2%. It's cheaper than a good office chair. And still, scared executives get advised to use the cheapest model "to control spend," optimizing to save ten cents on a decision worth thousands. Fear is a bad accountant.
Leaders won't get technical. AI is a technical topic, and educating yourself means facing what you don't know. A leader rarely gets trained for real because nobody likes feeling dumb in front of their team. Exactly like when computers arrived and bosses didn't want to learn what computing was. Today the biggest companies in the world are run by engineers. We already know how this story ends.
##Companies are humans times AI
The best frame in the video is a multiplication. A company, from here on, is humans × AI.
That multiplication has two factors. AI keeps growing on its own, so the factor on the right goes up no matter what. The human factor depends entirely on leadership. And anything times zero is zero. If your people aren't at the level, you can buy all the AI in the world and the result is still zero.
That's why the transformation is human work, not a purchase. Changing people and adapting processes is 80% of the job. The technology is 10 or 20%. And that, by the way, is why the "AI layoffs" story is mostly fiction. The real work isn't deleting people. It's growing them.
##What to do with this
Three frames, depending on where you stand.
If you work inside one of these companies. Don't wait to be trained. The gain is yours to take and it's invisible anyway. Become the person who uses AI, because that person quietly replaces the one who doesn't, inside the same company. The advantage is also exclusive right now: ChatGPT has around 800 million weekly users and only about 50 million pay. Learning to actually work with the tool puts you on a very small frontier. Those advantages don't last. That's why you take them early.
If you run a company or a team. Stop buying and start redesigning. And one data point that runs against the corporate instinct: that same MIT report found that buying AI from a specialized vendor works about 67% of the time, while building it in-house works only about 33%. Buying beats building. Pick ONE process, a painful and expensive one, and rebuild it end to end around AI. Measure a real business metric (time to close, conversion, capacity), not how many licenses you handed out or how many tokens you burned. Give it to the people with systemic thinking. Give it months, not weeks. And stop optimizing to save ten cents. AI also exposed something uncomfortable: that you can't measure your own operation. Fix that first, because without it you'll never be able to measure the impact of anything.
If you're an indie builder. Here's the asymmetry, and it's the cleanest edge in the game. You are the org. There's no income statement to blur the gain, no HR to mismeasure it, no leader's ego, no change-management quarter. And look at that buy-versus-build number: you are, structurally, a buyer. You grab the best tool off the shelf and use it today, which is exactly the posture that wins. The one thing that kills AI inside companies, the change of people and process, is the one thing you don't have to coordinate with anyone. The same tool that produces nothing spread across two thousand employees makes one person ship like a team. While the incumbents run theater, you run the real transformation, alone, today.
##The close
What people feared was that AI would hand even more power to the giants who could afford it. At the layer that matters, the opposite is happening. A multiplying lever is slowest to install where there are the most humans and the most process between the idea and the action. It's fastest to install in one decided person.
And the next wave of theater is already coming. Gartner predicts that more than 40% of agentic AI projects will be canceled by 2027. Same pattern, new model. The tool gets better every year. The companies' ability to absorb it barely moves.
The 95% isn't the story of a tool that doesn't work. It's a map of who can actually move.
AI isn't expensive or hard. What's expensive and hard is changing how people work. That's why it loses inside companies and wins inside you.
PS. Thanks to Freddy Vega for the video that unlocked this post. If the topic interests you, it goes much deeper than what fits here.