Harish Chidambaran sat in a federal courtroom in Brooklyn, facing a 10-count indictment. He built iLearningEngines to a $1.5 billion valuation on AI-powered learning tools and $421 million in reported revenue. Federal prosecutors say more than 90% of that revenue was fabricated. The fraud was not in the technology. It was in the language around the technology.
You have been in that room. A vendor across the table, using words you could not quite pin down to describe a product you could not quite see. The part of this story that touches your balance sheet is not the indictment. It is the pitch that came before it.
The Pitch You Have Heard Before
"AI does the work so you don't have to." You have seen it in pitch decks. You have heard it on sales calls. It sounds like progress. It is not progress. It is a claim, and in the past year and a half, three separate federal agencies have filed cases against companies that made that exact claim while humans did the work by hand.
The Evidence
The record on this is worth reading.
iLearningEngines reported $421 million in revenue for 2023. Then an investment research firm did something simple. It pulled the company's Indian subsidiary filings. The subsidiary reported $853,000 in revenue. The parent claimed a $216 million run rate in India. That is a 99.4% gap, found by reading a public document anyone could access.
Shares dropped more than 50% in a single day after the report came out. The company filed for bankruptcy by the end of that year.
Then came Nate Inc. Albert Saniger, the founder, told investors his shopping app ran at 93% to 97% automation. The SEC and DOJ filed parallel actions against him in early 2025. The actual automation rate was not low. It was zero. The orders were filled by contract workers in the Philippines and Romania. Saniger raised $42 million on that claim.
Then Presto Automation. The company told investors its drive-thru AI "eliminated the need for human order taking." Over 70% of orders still needed a human. At some locations, the rate was 100%. The SEC brought its first AI-washing enforcement action on that case.
And the FTC settled with a startup called Air AI, which promised small business owners its product would earn them "tens of thousands of dollars within 30 days." The company is now banned from selling business tools of any kind.
Four companies. Three agencies. One pattern. The pitch did not fail the people who heard it. It failed the system they were trying to run.
The Flaw Under the Pattern
The problem is not that AI vendors lie. Some do. Most do not. The problem is that technical language replaced the buyer's habit of checking whether the product actually does what the sales team said it does.
The vendor knows the product. The buyer does not. Jargon widened that gap instead of closing it. The seller holds the proof. The buyer holds the faith. Every case above turned on that distance between what was claimed and what was real.
Most people treat this as a technology problem. It is a verification problem.
Five Checks Before You Sign
The fix is not a degree in machine learning. It is five old checks that cost nothing and take an afternoon.
Check 1: Ask for the Human Rate
Every AI tool falls back to a human at some point. Ask the vendor what percent of tasks still need a person. Get the number in writing. If they will not give you one, that is your answer. This is the question Presto's investors never asked. It would have ended the pitch in five minutes.
That single question requires you to sit in the pause after you ask it. The way the vendor fills that pause tells you more than the demo ever will.
Check 2: Read the Filing, Not the Deck
If the vendor is a registered company, its filings are public. Match the revenue on the pitch to the revenue in the filing. The 99.4% gap at iLearningEngines sat in a document anyone could pull up. No one pulled it up.
Check 3: Call a Client the Vendor Did Not Give You
Not a reference from the sales team. A current user you found on your own. Ask one thing: does the tool do what the sales team said it would? A vendor with real clients will not mind. A vendor who minds is giving you a data point.
Check 4: Run a Paid Pilot with a Kill Clause
Do not sign a full contract on a demo. Pay for 30 days. Write in a clause that lets you walk if the tool does not hit a stated number. Air AI promised returns in 30 days. A 30-day trial with one hard metric would have caught the gap before the check cleared.
Check 5: Ask Who Does the Work
If the product claims full automation, ask where the tasks are handled and who touches them. Nate's app claimed machines processed the orders. Humans in Manila processed the orders. One question, one answer, $42 million in fraud exposed.
What the Checks Reveal
Running these across your next three vendor meetings does something the pitch deck never did.
You see which vendors welcome the questions and which ones shift the topic. You see the distance between the demo and the filing. You notice how fast a "fully automated" product needs a human when the use case gets specific. And you find that the vendors worth paying are the ones who already have the numbers ready before you ask.
The deflection is the signal.
At the End of the Quarter, Ask Three Things
→ Which vendor gave you a human rate without being pressed?
→ Which tool looked like progress but left no mark on your numbers?
→ Which sales call got worse when you asked for a filed document?
That is the gap between a pitch that sounds right and a system that proves itself.
Where You Stand
Chidambaran's company carried a $1.5 billion valuation. The check that would have caught it was free. It was a public filing anyone could read. The instinct to look was never missing. The habit of acting on it was. Start with what is true. Then build from there.
