Theranos Film Sparks Warning: Audit Every AI Claim Before You Buy

Thirty-four days before reporting to federal prison, Elizabeth Holmes opened her private life to a film crew. Nathan Fielder and Lance Oppenheim kept filming for roughly three years, and the result, You Can See Everything, opened in theaters this month to strong box office and wide acclaim. The title quotes a scene in which Holmes and her partner, Billy Evans, promise the camera it can indeed see everything.

The film invites a true-crime reading. Business author and Apply executive who builds agentic customer experiences reads it differently. The piece argues the story is about how trust is won and what happens when the product cannot justify it. Theranos was a brand long before it had a working product, and every signal it sent said the same thing: trust us. The product could not back any of it up. That should worry anyone writing a brand promise right now, especially one about artificial intelligence.

How did Theranos borrow credibility?

Theranos collected trust the way some brands collect logos. In 2014, Fortune described its board as three former cabinet secretaries, two former senators, a retired Navy admiral, and a retired Marine general. Holmes wore the black turtleneck and called her invention the iPod of health care. The Steve Jobs comparisons followed.

Each of those signals lent credibility. None of them created it. The author's frame is blunt: borrowed trust works like a loan, and the product eventually has to repay it. When it cannot, the brand pays. Any firm leaning on a famous partner, endorser, or technology name should ask what happens the day the product gets examined on its own.

The author applies the test to his own firm. Apply holds partnerships with Google Cloud, Anthropic, Amazon, and OpenAI. But a partner's name is a foundation, not proof, he writes. Make every agency, including mine, prove what it can deliver.

Is access the same as transparency?

No, and the distinction is the article's core test. An open door differs from openness. Anyone can offer to show everything and still control what gets seen. Brands treat transparency as performance: the trust page, the responsible AI statement, the demo built to be watched.

The author offers a quick check: a responsible AI page that lists principles but names no owner is decoration. If no one can say who answers for the claim, the page proves nothing. Buyers should ask to see everything and to choose what they see — unscripted, on their own data, including the failure log. Real transparency shows where the product falls short as clearly as where it shines.

What does real AI delivery look like?

In many agencies, the author says, using AI means people run ChatGPT or Claude on top of the same delivery workflows as before. That is a faster tool inside an old process, worth a 20%-30% efficiency gain. Calling it AI-native or agentic makes the claim bigger than the change.

At Apply, by contrast, the delivery model was rebuilt to be agentic-native on a proprietary platform, with the entire workforce trained to work that way. The author's benchmark: reimagining the model, not just the tools, is where 20%-30% becomes 20x-30x. He also concedes the industry's broader problem: brands are making AI claims their products cannot support yet, and agencies like his are helping write them. "Autonomous," "agentic," and "AI-powered" are easy to write and hard to back up.

How do claims drift from reality?

The article borrows a concept from sociologist Diane Vaughan, who after the space shuttle Challenger disaster described the normalization of deviance: small departures from what is safe get accepted one at a time until they are simply how the work is done.

AI claims drift the same way. First, the copy says "autonomous" while a person still handles the exceptions, and nothing goes wrong. The next campaign borrows the wording. By the third, the gap between claim and product is simply house style. The people closest to the work are often the first to see the gap and the least likely to say so.

Which three questions should buyers ask?

Before any AI claim goes into a campaign, the author says, pressure-test the agency with three questions:

• Can you show it working in production, failures included? Demand evidence from real data and real workflows, including what went wrong.
• Whose name is on it? Put one person's name behind the claim and the supporting results. People ask harder questions when they know they will answer for the outcome.
• What happens when it fails? State where the system needs human help, and make the promise reflect that limit. Test the handoff before launch, or customers will find where it breaks first.

Apply holds itself to the same questions, the author writes. Its teams are about half people and half agents. The agents extend the people, and every deliverable has a named human owner — accountability assigned before the work begins, a step beyond a human in the loop. "We're still learning," he adds.

The Theranos parallel closes the argument. At Theranos, the hard questions were ignored, and whistleblower Tyler Shultz, who raised concerns, faced legal pressure to stay quiet. The lesson for AI buyers: a brand earns trust by making the evidence shape the promise, rather than letting the promise outrun the evidence.