Knowing what a product computes tells you whether the price is for the math or for the label.
Every security vendor now sells anomaly detection with AI in the name. Some of it is new. A good deal of it is a statistical baseline with better marketing, and telling the two apart is worth money to anyone signing the contract.
The hype is inflated, because I helped build one of these systems and I know how little magic sits under the hood. The buyers who get hurt are small companies without a security team, paying a premium for a word and signing contracts they can’t evaluate. The vendors know that, and plenty of them price for it.
What does AI anomaly detection compute?
Nearly all of it computes the same three things.
First, what normal looks like for this customer, user, or device. Second, how far the current behavior sits from that normal. Third, whether the distance is far enough to act on.
That’s a baseline, a deviation, and a threshold. Modern systems may build the baseline in more dimensions and update it faster than older ones did. The shape of the answer hasn’t changed.
Ask a vendor to describe it in those terms and watch what happens. A good vendor will explain how their baseline is built and where the threshold comes from.
A weak one will say the model learns. I treat that answer as a warning, because a vendor who can’t explain the math either doesn’t understand it or doesn’t want you to.
Watch: Fraud Detection with AI: Ensemble of AI Models Improve Precision & Speed, IBM Technology
What is a behavioral baseline in fraud detection?
In the late 1980s I worked at a consulting firm that built a fraud detection system for phone cards. The client sold the cards, people used them, and some of those people weren’t the people who bought them. My idea was to give every customer a profile. The profile described what normal looked like for that person. Where they called from. When they called. How often, and for how long. Then every call got measured against that profile using standard deviation. A blanket rule saying calls from Guatemala are suspicious would have punished plenty of legitimate customers who call from there. We measured how unusual this call was for this particular person.
A man who lives in Florida makes several calls from Guatemala at two in the morning. That sits outside his profile. It’s not proof of anything, because he might be on vacation.
Then a call comes in from New Orleans. Now the pattern has gone from unusual to impossible, because a person can’t be in both places. The flags accumulate and the card gets cut off.
Now the pattern isn’t just unusual, it’s impossible, because a person can’t be in both places.Share on X
Can statistics do what machine learning claims to do?
For this class of problem, yes, and it did.
The system had no model and no training run. It had a profile per customer, a deviation measure, and a threshold. It caught fraud, the client was happy, and the firm turned it into a product. I’m not arguing that nothing has improved since. Modern systems handle far more signals, adapt faster, and find patterns a person wouldn’t think to look for. Those are real gains. The argument is narrower. A great deal of what gets sold as AI detection is the older method with more dimensions, and it’s priced as though it were new. I think that’s a quiet kind of dishonesty, and it falls hardest on the buyers least equipped to spot it. Knowing which one you’re buying changes the negotiation.
How do you set the threshold for a fraud alert?
The threshold is the entire product and it’s where most systems fail.
Set it tight and you block real customers. Every legitimate person who travels, changes habits, or has an unusual week becomes a support call and an angry email. Set it loose and fraud walks through. Our design handled this by requiring accumulation instead of a single event. One odd call moved a number. It didn’t cut anyone off. The card stopped only when the pattern had built past the point where an innocent explanation still fit.
A threshold you can’t adjust is somebody else’s risk appetite running inside your business.
That approach is available to any small company running any kind of detection today, and it’s nothing to do with the technology. Decide what a single anomaly is worth, decide what the total has to reach, and write both numbers down.
If a vendor can’t tell you where their threshold sits and who can change it, that’s the answer to whether you should buy it.
What questions should you ask an AI security vendor?
Four, and none of them are technical.
What does the system compute, described without using the words AI or learning. How’s normal established for a new customer who has no history yet. What happens when the system is wrong, and who finds out. Who can change the threshold, and does that require the vendor. A vendor with a real product answers all four in a few minutes. A vendor selling a label will move to case studies. The last question is more important than it looks. A threshold you can’t adjust is somebody else’s risk appetite running inside your business.
What happens when one person is the only one who understands the system?
A man named Tom built the code base, and Tom was a genius. I came up with the profile and deviation concept, and he built the thing. He probably did the majority of the design as well. Tom was Croatian, and during that period he’d occasionally leave for a week to help with a freedom movement back home. Little notice. He’d tell us he was going, and then he was gone. Managing a project where the one person who understood the code could vanish for a week wasn’t the disaster it looks like from outside. It happened twice during that project, we built the time into the schedule, and the work got done.
What made it survivable was that the absences were known and not surprises. A dependency you’ve planned for is a scheduling fact. The same dependency unplanned is an emergency. Most small companies have a Tom, and pretending you don’t is the mistake. An owner who won’t plan around his Tom has picked the emergency over a fact he could have scheduled for, and his staff are the ones who pay for that choice.
My take
Make the vendor tell you what the system computes, without saying AI. A baseline and a deviation from it is a method we used on phone cards in the late 1980s. It works fine, and it isn’t new. Then ask where the threshold sits and who can change it. A vendor with a real product answers in minutes. One selling a label changes the subject to case studies, and you can walk away.
What did the phone card project prove about new technology?
The phone card system came out well. Good product, happy client, and the firm sold it afterward.
What stayed with me was how ordinary the core idea turned out to be. Know what normal looks like for each individual. Measure how far the current behavior sits from it. Act when the distance is far enough. Nobody called that intelligence in 1990 and nobody would have paid extra for the word.
When a vendor describes something you can’t follow, ask what it computes. Sometimes the answer is new, and often it’s a baseline, a deviation, and a threshold. That’s worth buying at the right price. I’d walk away from any vendor who charges a premium for the word AI stacked on top of math we were running on phone cards decades ago.
