The Darden School of Business just published a new teaching case on Theranos, and they called it “Fake It Till You Fail.” Not “Lessons in Scaling.” Not “A Cautionary Tale.” Darden’s framing treats the company as a study in the risks that ride along with entrepreneurship and new ideas, which is a polite way of saying: this is what happens when the story outruns the machine.
I read that title and immediately thought about AI agents. Stay with me.
Why a blood-testing scandal keeps showing up on my radar
Elizabeth Holmes is having another moment in 2026. A24 has a documentary coming this fall called “You Can See Everything,” with interviews conducted by Nathan Fielder, offering what ABC News describes as an intimate look at Holmes before her imprisonment. There’s also talk of a possible move to a halfway house in 2027, though the details are thin.
So we get a documentary, a business school case, and a news cycle, all at once. And in the same stretch of weeks, I’m watching broadcast segments where a host is handed an AI agent and told it can “go off and do stuff for you.” Organize your calendar. Handle your errands. Think on your behalf.
Those two things feel unrelated. They are not. Both run on the same fuel: a convincing demonstration of something you can’t personally inspect.
The demo is not the product
Here’s what I keep coming back to when I explain AI agents to people who don’t write code. An agent is software that takes a goal, breaks it into steps, and acts on those steps using tools: browsers, email, calendars, payment systems, your files. When it works, it looks like magic. When it fails, it often fails quietly, in the middle, in a way that looks finished.
That gap between “looks finished” and “is finished” is the whole ballgame. A polished agent demo on stage is a performance with a known script, chosen inputs, and a safety net off camera. Your Tuesday afternoon is not a known script.
The Theranos lesson that stuck with business schools wasn’t that the technology was hard. Hard is fine. Hard is normal. The lesson was about the distance between what people were told was happening inside the box and what was actually happening inside the box, and how long that distance went unexamined because the presentation was so good.
AI agents are nowhere near that scale of deception. Most of the people building them are earnest, and the underlying tech genuinely does useful work. But the structural setup is similar enough that it’s worth borrowing the skepticism. You are being shown an output and asked to trust a process you cannot see.
Questions to ask before you hand over the keys
I’m not telling you to avoid agents. I use them. I’m telling you to ask the kind of questions a Darden case study would want you to have asked earlier:
- What specifically did it do? A good agent shows its steps. If you can’t see the trail, you can’t audit the result.
- What happens when it’s wrong? Ask the vendor to show you a failure, not a success. How does the tool signal uncertainty instead of inventing a confident answer?
- What can it touch? Email, bank accounts, and file deletion are not the same risk tier as drafting a summary. Start agents in low-stakes places.
- Who checks the work? If the answer is “nobody, it’s automated,” you’ve just built a system where errors compound silently.
- Is this a product or a pitch? Shipping software has documentation, pricing, limits, and known bugs. Pitches have sizzle reels.
Intimacy is not transparency
The thing that fascinates me about the documentary’s title, “You Can See Everything,” is how it plays on the promise Theranos made in the first place. Full visibility from one small sample. Total clarity, minimal effort. That’s also the pitch for AI agents: tell it your goal, see everything handled.
Access to a person’s story isn’t the same as understanding what they did. And a chat window that narrates its reasoning in friendly prose isn’t the same as a system you can verify. Both give you the feeling of seeing everything while the mechanism stays out of frame.
The practical takeaway is unglamorous. Treat agent output like a draft from a fast, tireless, occasionally confused intern. Check the parts that matter. Keep humans in the loop where money, health, legal exposure, or someone’s reputation is involved. Build the verification habit now, while the stakes are small and the tools are new.
Because the uncomfortable part of “fake it till you fail” isn’t the faking. It’s how many smart, well-intentioned people watched the demo and decided they didn’t need to look inside the box.
🕒 Published: