What AI-Powered Due Diligence Still Cannot Detect: Dishonesty.

AI startups are inflating their ARR numbers. Their investors know it. And the whole ecosystem is nodding along like this is just how the game works now.

The Wall Street Journal reported it plainly: founders and VCs are stretching the definition of Annual Recurring Revenue to make AI companies look bigger, faster, and more inevitable than they actually are. Usage-based fees getting annualized. One-time pilots dressed up as recurring contracts. Consumption revenue wearing a subscription’s clothes. The numbers look clean on a slide deck. The underlying reality is considerably messier.

This is not a footnote story. This is a judgment story.

**The Metric Isn’t Broken. The Judgment Is.**

ARR is not a complicated concept. It measures predictable, recurring revenue. That’s it. But somewhere between the whiteboard and the pitch meeting, founders and their backers decided that “predictable” was negotiable. That “recurring” could mean “we think they’ll come back.” That the spirit of the metric mattered less than the number it produced.

Software veterans have been measuring ARR the same way for twenty years because consistency is what makes the number useful. The moment you let everyone define it differently, you don’t have a metric anymore. You have a story. And stories, however compelling, are not the same as data.

What failed here was not the spreadsheet. It was human judgment. Somewhere in every one of these inflated numbers, a person made a call. They decided the risk of overstating was worth taking. They decided the investor across the table either wouldn’t notice or wouldn’t care. They decided the short-term headline was worth more than long-term credibility.

AI didn’t make that decision. A human did.

**Pattern Recognition Is Not Wisdom**

Here is what AI does extraordinarily well. It finds patterns in large datasets faster than any human analyst alive. It can scan a thousand pitch decks, identify the common signals of early traction, and rank them by likelihood of success. It can automate the grunt work of due diligence in ways that would have seemed like science fiction a decade ago.

Here is what AI cannot do. It cannot sit across a table from a founder who is technically telling the truth while spiritually lying through their teeth and know the difference.

Experienced investors talk about this all the time, though rarely in print. They describe a feeling. A moment in a meeting where the numbers add up but something doesn’t. Where the founder’s confidence reads slightly too rehearsed. Where the answer to a hard question comes just a little too fast. That read — built from years of watching people perform certainty they don’t have — is not a formula. It is judgment. It is irreducibly human.

No language model trained on pitch decks and funding announcements can replicate that. It can tell you the ARR number. It cannot tell you what the founder’s eyes did when you asked about churn.

**The AI Hype Cycle Is Specifically Designed to Outrun Scrutiny**

There is a reason inflated metrics survive in AI more than in other sectors right now. The technology is genuinely novel enough that most people in the room aren’t sure what normal looks like. When nobody has a clear benchmark, the founder with the most confident benchmark wins. Ambiguity is a feature, not a bug.

This is the exact environment where human judgment becomes most valuable and most endangered at the same time. Endangered because social proof is powerful. When a16z is in the round, the pressure to defer is enormous. Valuable because someone still has to ask the question nobody wants to ask. Someone has to look at the annualized pilot revenue and say: that’s not ARR, that’s a forecast wearing a costume.

That someone has to be a person. Specifically, a person with enough independence of mind to be unpopular in a room full of believers.

**What Gets Lost When Judgment Stops**

Markets built on inflated metrics do not correct gracefully. They correct catastrophically. We know this because we have watched it happen before — in dot-com valuations, in SPACs, in crypto. Every cycle has its preferred unit of fiction. This one’s is ARR.

The damage is not just financial. When the correction comes, it discredits legitimate AI companies alongside the fraudulent ones. It makes the real breakthroughs harder to fund. It poisons the well for founders who were actually telling the truth.

That damage was preventable. It required exactly one thing: human beings willing to exercise judgment rather than defer to the room.

AI can optimize your pipeline. It can summarize your board deck. It can do a hundred things faster and cheaper than a human analyst.

It cannot be the person who says no when everyone else is saying yes.

That job still belongs to us. We should probably start doing it.