Home / Content Hub / Blog

When the AI Gets It Wrong, We Get Better

Read why data quality for AI agents matters and how clear business rules and human oversight make mistakes easier to spot and fix.

Renata Pavlovska Renata Pavlovska Published 19 November 2026 · 5 min read
Share
When the AI Gets It Wrong, We Get Better

How a data leader at IWConnect thinks about challenges, mistakes, and building AI people can actually trust.

Nobody gets everything right the first time, not people, not teams, and not AI. What really matters is how we respond when things go wrong. Over the years, I’ve learned that the moments when something breaks often teach us far more than the moments when everything works.

Data was always the common thread

Twenty-some years ago I was hired to build applications. Somehow, I kept getting pulled into the data instead. A broken import here, an urgent report for a manager there, a file that had to reach a partner’s system by morning.

Most of the applications I built back then have long been replaced. The data they produced is still around, moved, reshaped, and feeding systems that didn’t exist at the time. At some point I stopped seeing data as a side job and started seeing it as the part that lasts.

That realization led me from development into data engineering, then architecture, and eventually leadership. Today, as Director of Data at ⋮IWConnect, my role mixes shaping our data and AI strategy, exploring new technologies, and helping clients solve complex challenges, increasingly with AI at the center of them. Most importantly, it means building a team where people can grow, experiment, and keep learning.

The courage to try, the openness to fail

If I had to pick one of our six values, it would be Accepting Challenges. But the more I think about it, the more I see that Accepting Challenges and Accepting Mistakes are inseparable.

Every new challenge comes with uncertainty. Stepping outside your comfort zone means accepting that things won’t always go as planned. To me, both values are about the courage to try, the openness to admit when something doesn’t work, and the willingness to learn and improve.

After all, if we’re not making mistakes occasionally, are we really challenging ourselves enough?

Mistakes are steps forward, not failures

Mistakes are often where the real learning happens. In our industry, technology evolves so fast that we can’t always rely on what worked yesterday. Sometimes the only way to find a better solution is to try something that doesn’t work the first time. I’ve hit a wall hard a few times myself, and looking back, those are the moments that taught me the most.

The goal isn’t to avoid mistakes at all costs. It’s to recognize them early, take responsibility, and turn what we learned into something valuable.

As a leader, I want my team to feel confident enough to challenge ideas, experiment, and speak openly when something goes wrong. That’s how we grow together.

Everything was green, and still wrong

Application errors are loud. You click a button, nothing happens, and everyone knows something is broken. Within minutes there is a ticket, a fix, and a deployment.

Data errors are very, very quiet. The pipeline runs on time, every validation passes, every check is green. Nothing crashes and nobody gets an alert. Meanwhile, a misunderstood business rule or an assumption nobody questioned is slowly producing the wrong numbers. You usually find out too late, or worse, the business finds out for you, sitting in a meeting with a report that doesn’t add up.

And today, data doesn’t stop at a report. It feeds data products, machine learning models, and AI agents that make decisions on their own. A quiet error no longer sits in one dashboard waiting to be noticed. It travels, gets reused, and gets acted on, often at a scale and speed no person can follow. I’ve learned that a green check only tells you the data arrived. It doesn’t tell you whether it’s true, and the further data travels, the more that difference matters.

When the AI gets it wrong

If data errors are quiet, AI errors are quiet and confident. An agent gives you an answer that is well written, nicely explained, and completely wrong. It doesn’t hesitate, and it doesn’t warn you.

The first instinct is to fix the answer, tweak the prompt, and move on. I’ve learned to resist that. When an AI agent gets something wrong, the more useful question is why. More often than not, the model did exactly what it was asked. It just didn’t know what we knew. It saw tables and columns, but not the business behind them: that “revenue” means something different in finance than in sales, that a customer, a contract, and an invoice are connected in a very specific way, or that a certain rule only applies to one market. It filled the gap with a guess, and the guess sounded convincing.

That’s why I believe trustworthy AI starts with trustworthy data, and with meaning that lives outside people’s heads. For years, well-modeled tables were enough. They served BI perfectly, because there was always a person on the other side of the report who knew what the numbers meant. An AI agent doesn’t have that person in its head. An agent needs a step of abstraction above them: a shared picture of what things are and how they relate, the kind of picture ontologies and knowledge graphs quietly give it. With that, it reasons over the business instead of guessing from column names. It still needs us to keep checking its work, and a person close enough to question a decision before it becomes a problem. None of this makes AI perfect. It makes its mistakes visible, explainable, and fixable, and that’s a very different thing from a mistake nobody can see.

Trust doesn’t come from perfection

Trust in AI doesn’t come from expecting it to be perfect. It comes from knowing we can understand, challenge, and improve its decisions.

The same goes for people, teams, and data platforms. Every mistake an AI agent makes is a chance to sharpen our definitions, tighten our controls, and deepen our understanding of the business. When the AI gets it wrong, we get better.

Renata Pavlovska

Renata Pavlovska

The IWConnect team shares insights on enterprise IT operations, observability, automation, and digital transformation.

Curious how this applies to your numbers? Let's find out.

Share where things are getting stuck today and we will walk you through what a fix could look like.

Talk to our team