Most Companies Are Adopting AI Wrong, Including Yours
In 2000, Netflix was a DVD-by-mail startup hemorrhaging cash. No streaming, no original content, and no path to profitability anyone could see. Just a website and a warehouse full of red envelopes burning through venture money while the dot-com bubble collapsed around it. Reed Hastings flew to Dallas and, by cofounder Marc Randolph’s telling, offered to sell the whole company to Blockbuster for $50 million, a trivial amount to Blockbuster at the time. They passed.
When they realized the mistake they made, they scrambled to build their own answer. Launched in 2006, Total Access was a Netflix competitor with a convenience Netflix couldn't match — rent online and return the disc at any store. It grew faster than Netflix, pulling in two million subscribers in its first year. For a moment, it looked like Blockbuster might actually pull ahead.
But the company couldn't escape its own gravity. Blockbuster was drowning in a billion dollars of Viacom-spinoff debt and leaned heavily on the $145 million a year its late fees generated. (Netflix, on the other hand, charged no late fees at all.) The future was clear enough for Blockbuster to act on it, but nobody on the board wanted to trade a guaranteed revenue stream for a digital bet.
The board made its choice as it pushed out the executive who built Total Access. His replacement turned the company back toward the stores, announcing on the record that "neither Redbox nor Netflix are even on the radar screen." He wrote off the DVD business itself as a "melting glacier," melting slowly enough to keep betting on stores.
Blockbuster filed for bankruptcy two years later.
Our Artificial Intelligence Moment
Every corporate leadership team on earth right now is running some version of Blockbuster's math. I have an unusually good seat for watching it happen; I am an AI implementation specialist for small businesses and now for a Fortune 100 enterprise. Every AI implementation I've worked on has the same structure:
- Train the staff on how to use AI chatbots and projects, and
- Make a list of everything the staff does, identify responsibilities that are easy to automate, and automate them.
The results of such a rollout can be impressive by pre-AI standards, with double-digit productivity gains on the automated tasks. But they can also be nothing. MIT's 2025 State of AI in Business report found that about 95% of corporate gen-AI pilots deliver no measurable impact on the bottom line. Not a disappointing return, none.
Businesses capturing orders-of-magnitude value out of AI approach things differently. They ask one question: how can we alter the business model so AI agents can operate it end-to-end?
You can measure how rare that question is. McKinsey's 2025 State of AI survey looked at 25 factors that might separate the companies seeing real bottom-line impact from the ones treading water. The single biggest was redesigning workflows, and only about a fifth of organizations had done it. If barely a fifth will redesign a workflow, imagine how few are willing to redesign the business around it. Everyone is busy building a better Blockbuster.
An AI-First Approach is Different
AI is not an operational upgrade. It is a new technology in the electricity-and-internet sense, the kind that rearranges how humans do nearly everything it touches. Something that large changes what the business is, not just how efficiently it runs. The structure of a true AI first initiative looks like:
- Redesign business processes so they are easy for AI agents to perform.
- Redesign technical systems so it’s easy for AI agents to connect to and operate them.
As you can see, this is a totally different approach from the first one. The first one uses AI to assist humans in doing the job; the second has AI doing the job, with humans only stepping in to review.
What that looks like is different for every enterprise, but a few pieces of the work are always on the table:
- Migrating data into machine-readable formats like Markdown and HTML, leaving proprietary and specialized formats behind.
- Building APIs and/or MCPs and/or CLIs for every business-critical system.
- Doing a complete security permissions audit, because the first two steps almost always surface security and access issues that were buried before.
This is definitely more work than the AI-assisted approach, but it’s not that much more work (I’d say only 40% more) and the results are just daggone spectacular. So much drudgery saved you wouldn’t believe it, and the new opportunities to take service to another level abound.
It’s actually more unbelievable than that, because in addition to getting more spectacular results initially, the investment compounds; it is easy to train agents for such a business, and as agents continue to grow more sophisticated these businesses will grow naturally into the new capabilities. Enterprises that compromised with an AI-assisted approach initially, however, will keep paying the price.
Why Leadership Says No
On paper, all of this wins easily. But I've been in the rooms where this decision actually gets made, and leadership teams are skeptical for many reasons, some of which are good and some of which are not.
Major renos suck
Any middle manager who wants to remain employed rightly shies away from a pitch involving a major technology alteration. They’re long and hard and expensive and have a greater chance of failing. Leaders who operate on long horizons are more likely to be interested.
Wait, why HTML?
Decision-makers who are not technologists (which is most of them) associate machine-readable formats like plain text and HTML with “code” and “code” with moving backward. Wasn’t the whole point of WordPress/Salesforce/Oracle so staff didn’t have to look at code? Now, anyone technical realizes that using AI to build a human-GUI compatible with a machine-readable format is trivial in this day and age, but again, these decision makers are not technologists. And even trivial though it is, it adds more to the workload.
Nobody else is doing this!
This is, of course, not true — Twitter is stuffed to the gills with people trying — but these well-adjusted, not chronically online non-technologist leaders don’t know that. They haven’t read of it in their business sources yet. And nobody wants to be the first to do something new in case it fails catastrophically; it’s much easier to sell something with industry standards and historical case studies. Our industry standard is… /skill. We’ll get there.
Agents are not reliable enough
Some complain that agents just aren’t reliable enough. Maybe egoism has deluded leadership into thinking what they do is harder for an agent than it really is. Maybe the thing in question is genuinely complicated and requires a lot of human-in-the-loop steps. Maybe some combination of both. Regardless, it’s a fair objection, as agents can be hilariously unreliable under certain conditions. But going AI-first doesn't mean handing an agent the keys tomorrow. It means building the architecture so that when agents are reliable enough, the wiring is already there for you. You keep every human checkpoint you have now; you just stop assuming the human is the one doing the work between them.
What if it all goes horribly wrong
Remember what happened at Blockbuster. The executive who bet on digital got pushed out, and the job went to the man who promised to protect the stores. Now put yourself in the shoes of whoever has to champion the AI-first project. They're standing in front of a budget committee, proposing to spend two years' profit rebuilding systems that currently work, for a payoff they can't fully describe on a timeline they can't promise. If they're right, the benefits land long after they've moved on. If they're wrong, it's their name on the write-off.
Here's the part that should unsettle the risk-mitigators. It is precisely because established leaders won't take this bet that the door swings open for entrepreneurs building AI-first from scratch, with no stores to protect. The incumbents' escape hatch is the one they always reach for — wait, then buy the winners. Which is exactly the option Blockbuster had, at $50 million.
If you're a small operator rather than a Fortune 100 company, the door is already open for you. Please, for the love of God, walk through it.
The Netflixes That Will Kill You are Already Here
These AI-first entrants aren't hypothetical. The same MIT report that counted the failing pilots found AI-native startups, some run by founders barely out of their teens, going from zero to $20 million in revenue in a year, mostly by picking one narrow workflow and executing it well.
My favorite example is Medvi, a telehealth company selling GLP-1 weight-loss drugs. Its founder stood it up in about two months with $20,000 and a stack of AI tools. Agents wrote the platform code, generated the ads, and answered the support tickets, in a market already crowded with Hims & Hers and Ro. Total headcount: two. In its first full year, Medvi booked $401 million in revenue, according to financials reviewed by the New York Times.
Medvi is far from perfect. It’s support chatbot has hallucinated drug prices, which they then honored, and the FDA has already sent it a warning letter over its marketing claims. But that's exactly my point! A duct-taped, two-person, agent-operated business took four hundred million dollars out of a regulated market while 95% of big pharma enterprise pilots couldn't move the P&L at all.
The careful version is starting to appear inside big companies, one function at a time. Klarna's AI assistant took on two-thirds of its customer-service chats in its first month, doing the work of 700 human agents. When quality slipped, Klarna brought humans back into the loop rather than scrapping the system. That stumble gets reported as an AI failure, but it wasn’t one. It’s the system working exactly as designed.
Stop Building a Better Blockbuster
Blockbuster didn't need a crystal ball. It needed to stop treating nine thousand leases, $145 million in late fees, and "we have always been a stores company" as the fixed premises everything else was built on.
Every time you catch someone asking — or catch yourself asking — "How can AI improve this process?" stop. That's the Blockbuster question. The one that opens the ground is "If we were founding this company from scratch using AI, how would we do it?"
AI was used in the writing of this article.
Enjoy this kind of writing?
I write about AI, intentional living, and doing meaningful work in a world that won't stop changing. Join 700 other subscribers.
Keep Reading
Studies Show People Can’t Tell AI Art From Human Art
They even show that sometimes, we prefer the AI’s work
Your Clothes Use More Water Than Your AI
The environmental case against AI doesn't survive contact with the data
Apps Will Soon be Replaced by AI
The first new computing interface in sixty years doesn't need them.