The Failure-Mode Test: The Fastest Way to Separate AI Hype From Reality

Every week someone asks me some version of the same question: with all the noise, how do you actually tell which AI is real and which is marketing? After 35 years in software and security, and after making AI buy-or-skip decisions inside my own companies every week, I use a filter that costs nothing and takes ten seconds.

Ask: where does this break?

I call it the failure-mode test, and it outperforms every benchmark, demo, and analyst report I have ever read.

Why one question beats a hundred demos

A demo shows you the happy path. A benchmark shows you performance on someone else's data. Neither tells you what happens on the ugly Tuesday when your real inputs, your real customers, and your real edge cases show up.

But the answer to "where does this break?" tells you something a demo never can: whether the people behind the system have actually done the work.

Real engineers light up at the question. They will tell you about the inputs that confuse the model, the edge cases they found in testing, the guardrails they built, and the monitoring that catches the rest. They know the limits because they have hit them, repeatedly, on the way to shipping something dependable.

Marketers change the subject. They pivot to the roadmap, the vision, the latest benchmark score. Not because they are dishonest, but because nobody handed them a failure-mode list. And if that list does not exist, the product is not finished. Someone will write it eventually. Usually your customers.

Real systems ship with their limits attached

Here is the pattern I have watched across every technology wave: capability gets you the pilot, but known limits get you to production. You cannot build guardrails around failure modes nobody has named.

I learned this the hard way building AI for sports video at Egoli. We used facial recognition to find marketable moments in football matches. Trained against every Premier League player, the system produced garbage; at one point it confidently tagged a player who had been dead for twenty years. The fix was not more AI. We constrained the search to the 22 players actually on the pitch, and accuracy soared. The system only became valuable once we understood exactly where and why it failed, and engineered around it.

That is what a working AI deployment looks like from the inside. Not a system that never fails, but a system whose failures are known, priced, and contained.

The word that signals reality: "except"

The AI pitches worth taking almost always include the word "except." It works, except when the input looks like this. It saves time, except for these cases, which still need a human.

"Except" is not weakness. It is evidence of stress-testing. Hype has no exceptions. Reality is made of them.

So when a vendor, an AI consultant, or your own internal team presents a plan with no downsides, that should worry you more than a plan with a page of caveats. The caveats are what let you design around the limits. A plan without them is a plan nobody has pressure-tested.

How to run the test yourself

You do not need a technical background to apply this. Before your next AI purchase or internal build, ask three things:

1. Where does it break? If the answer is vague, hand-wavy, or pivots to the roadmap, stop. If it is specific, keep going.

2. What does a wrong answer cost? Every process in your business has a price for being wrong. Know that number before you let AI touch the process.

3. How do we find out it failed? Silent failure is the expensive kind. A system that fails loudly, where a human can catch it, is worth far more than one that fails quietly at scale.

If you can answer all three, you are having a reality-based conversation about where AI fits in your business. If you cannot, you are shopping for a story.

This is the actual job

People sometimes assume the job of an AI advisor or fractional CTO is finding the most powerful model. It is not. The job is mapping the failure modes of both the technology and the business, then putting AI where its failures are survivable and its wins compound.

That is judgment work, and it is the part the hype never mentions, because judgment does not demo well.

If you are staring at an AI decision, whether a vendor pitch, an internal project, or the broader question of where AI fits in your business at all, I am happy to be a second set of eyes. I take a few clients at a time, and the first conversation is free.

Book your free 15-minute AI Fit Call →