arato.ai

It took us three turns to get out of a parking ticket, no hacking required. We tried talking to an AI agent one of our municipalities deployed, to cancel a ticket we'd gotten. Nothing sophisticated about it, just a few turns of conversation and making something up, and the ticket was canceled. We tried the same thing on two other municipalities' agents and it worked just as easily both times. I'm sure these teams thought they'd tested their agents properly before deploying them. They probably ran a few datasets and called it done. But they never actually tested the agent the way a full system needs to be tested. That's one of the biggest gaps we keep running into. Organizations deploy these agents assuming they work, without ever really knowing that they do.

For thirty years, it didn't matter who sat in front of the software. Now it does. I've been in software testing for about three decades. Traditionally, it didn't matter who was using it, male or female, any location, everyone got the exact same experience. Once you move to an AI-based experience, that stops being true. Who you are, your context, your history, all of it shapes the experience you get. Male or female. Europe, the Middle East, the US. The language you speak, your tone of voice, your history. One bank insisted we test dyslexic users specifically, because they get a completely different experience once AI is involved. Most people still don't realize this. The only way to actually know your AI holds up across that many contexts, behaviors, languages, religions, and intents, is to simulate thousands of real user interactions and see how the system actually handles the variety.

"You told me it works." That's what the CIO of one of the largest banks in Spain said to his own head of AI, an hour before they pulled their agent off the website. Here's what happened. We ran a simulation on their AI agent, live on their site, without being asked. We emailed them: we found these things, do you want to talk? They got on a call. The head of AI said the issues were probably old, everything works now. We told them we'd run the simulation the night before. The CIO stopped the conversation. Looked at his own AI lead. "You told me it works. How can they find all these issues when you promised me it works? I can't believe the team promised me this thing's working and it's not." An hour later, the agent was gone. This is what we keep seeing. Teams assume it works. Someone said it would. It got tested, technically. It's not ready for production. They assume 90%. It's running under 40%. If you want to know your agent works, you have to actually test it. Simulate it, thoroughly.

One of Spain’s largest banks thought its AI was performing as expected. An Arato simulation showed otherwise, and the agent was taken offline within the hour.

A municipal AI agent canceled a parking ticket after a user simply made something up. That’s the kind of failure standard testing can miss.

One bank spotted a gap in its AI testing: dyslexic users. It’s a good example of why production readiness depends on testing far beyond the expected user.