Vivek Thomas is the CEO of AISensum.
A bakery in the Netherlands trains its staff on a standard operating procedure (SOP): Welcome the customer. Offer one relevant upsell. Say “thank you” on the way out.
A bike shop two streets over trains on something almost identical: Understand the problem. Offer a real solution. Suggest one accessory if it fits. Close warmly.
Every frontline business runs on some version of this SOP. It’s short, simple and almost nobody actually measures whether it happens.
Leadership assumes the SOP is being followed because there is a training deck, a certificate and a manager who occasionally watches the floor. None of that tells you what happened during the other 350 transactions the manager didn’t see.
After nearly two years building AI measurement for frontline teams across Southeast Asia, here are the three blind spots I keep finding and what closes them:
1. The Sample Size Problem
Most companies measure frontline compliance through NPS scores and mystery shoppers. Both have the same flaw. They capture a tiny, self-selected slice of what actually happened. An NPS survey catches the customers who bothered to respond. A mystery shopper catches one visit, once a quarter and every experienced cashier can spot one coming.
None of this covers the bulk of the interactions where money was left on the table. A missed upsell doesn’t show up in a satisfaction score. It shows up nowhere.
2. The Self-Recognition Effect
I have asked cashiers directly how often they think they complete the upsell step. The answer is almost always the same. “Ninety percent of the time,” they say. Then, I show them a moment where it didn’t happen.
The reaction is instant. “Oh. Yeah. I missed that.”
What matters isn’t the gap between belief and reality. It’s what happens the second they hear the recording. They don’t need it explained. They remember the whole scene, the customer, the till, the exact second they moved on instead of asking one more question. Our brains hold onto such moments the way we remember an old film scene. The memory was already there, it just needed to be surfaced.
That moment of recognition does more for compliance than a training deck ever will. Nobody improves because a manager told them to. People improve when they see themselves miss something and feel the discomfort of knowing they could have done better. That is self-training, and it is quieter, cheaper and stickier than a classroom.
At one of my client’s retail entertainment companies in Southeast Asia, frontline staff who could see this kind of feedback on their own performance, with no manager pushing them, drove a 10% improvement in their performance index. That lifted sales by 3% and increased individual incentive payouts by 25%. Nobody was beaten with a stick. The data was theirs, the improvement was theirs and the reward was theirs.
3. The One-Size Training Waste
Most companies train every frontline worker on the entire SOP, regardless of where they’re actually falling short. A cashier who is excellent at greeting customers still sits through the greeting training, because nobody has isolated where their real gap is.
Once you can see performance by person and by step, training stops being a blanket exercise and becomes targeted. Some people need upsell coaching. Others need training on problem resolution. Spending equal hours on both wastes the hours that mattered.
How This Actually Gets Captured
Here’s a method worth understanding. A microphone near the collar captures the employee’s customer conversations. An AI engine listens to that audio and scores it against the SOP: Was the customer welcomed, was the upsell offered, was the “thank you” said? It surfaces specific instances, moments they did it well and moments they missed it. Because it’s an engine and not a person, it can run across every cashier, shift and store in a network, continuously.
That scale is exactly where a problem shows up. The customer is standing there, too, so some of their voice gets captured. That matters because a voice can qualify as biometric data, and the EU AI Act draws hard lines around what may be inferred from it.
Here’s how to deal with it. Filter the audio for amplitude before anything else. The employee’s voice is closer and louder, the customer’s is farther and fainter. Use that gap to isolate the employee’s side, then deliberately degrade whatever residual customer voice remains. The customer’s voice is never analysed, scored or used to identify anyone. What gets scored is the employee’s compliance with the SOP.
What To Watch Before You Build This
This isn’t free of complications. The EU AI Act flatly prohibits inferring emotions from workers on the job; it is one of the Act’s banned practices, and any serious operator has to respect that line rather than route around it.
There is a harder question underneath it, too. Does this become a surveillance tool leadership uses against people or a mirror an employee uses on themselves? The answer should be the latter. The individual feedback goes to the person first, privately, the way a coach talks to an athlete between plays. What goes to leadership is the aggregate, never the individual transcript.
Get that right, and the incentive structure does the rest. Nobody wants to sit in the bottom quartile once they can see where they stand. The old employee of the month was usually decided by one customer who happened to write a nice note. A fairer version rewards the person who was actually best, consistently, across every transaction—not the luckiest one.
The Real Shift
The technology here isn’t the interesting part. The interesting part is what happens to a person in the two seconds after they realize they missed something they were fully capable of doing. That reaction isn’t new. What’s new is the ability to show it to them, every day, without a manager standing over their shoulder.
Measurement was never the real problem. Getting someone to see themselves clearly was.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?