4 min. read
An interview with Antonina Vodovytska, Salesforce Product Owner at OLX Romania — on building Olly, the AI that handles customer conversations, and the unexpected lesson about what makes AI trustworthy.
Antonina, where did Olly come from? What was the problem you were trying to solve?
We’ve always been looking for a better way. Over the years, like many companies, we tested various chatbot and automation tools. But traditional bots often felt rigid — they could only follow strict decision trees, which left users feeling stuck and frustrated.
The moment everything shifted was when we realised that to genuinely improve the customer experience, we needed something that could hold a natural conversation — and something completely under our own control. Not a third-party black box. Ours.
Olly was born out of that: a move away from rigid scripts, toward conversations that actually help people. OLX gave us the space and the tools to build it ourselves rather than buy something off the shelf. That decision shaped everything that followed.
Olly knows when to hand over to a human. Was that a deliberate design choice?
Completely deliberate. And it comes from hard-won experience.
Our previous work with automation taught us a lesson we’ve never forgotten: your goal should never be to close a ticket at any cost. Trust is incredibly hard to build and very easy to lose. An AI that pushes through when it shouldn’t — on a payment issue, a sensitive account problem — doesn’t just fail that one customer. It damages something that takes months to rebuild.
So we designed Olly to be honest. If a query involves something complex or sensitive, Olly hands it to a human expert — gracefully, without making the user feel like they’ve hit a wall. Being genuinely helpful sometimes means knowing when to step aside. That principle is built into Olly at the foundation, not added on as an afterthought.
What surprised you most during the build? What didn’t go as expected?
We assumed that if the information was in our Help Center, Olly would find it and explain it perfectly. We were completely wrong.
AI interprets information differently than humans do. If we had two articles with overlapping content, or a slightly outdated policy still live on the site, Olly would get confused — and that confusion showed up in customer conversations in ways we hadn’t anticipated. We had to completely rethink how we structured our knowledge base.
We started calling it “knowledge hygiene.” We learned to write articles in a highly structured, modular way that large language models can actually digest cleanly. No redundancy. No ambiguity. Clear hierarchy.
And here’s the unexpected part: teaching Olly to read made us better at writing for humans. By structuring our knowledge base, we unlocked the clearest, most readable customer support we’ve ever had. We didn’t expect that side effect. It’s become one of the quiet wins of the whole project.
What’s surprised you most about how customers interact with Olly?
The politeness. I wasn’t expecting it at all.
Even though users know they’re interacting with an AI, so many of them still open the conversation with “Hello” or “Good morning,” and close with “Thank you.” It moved me, honestly. It tells you something about human nature — and about what good AI design can do.
If an AI treats people with respect and clarity, they naturally mirror that warmth back. People aren’t cold toward Olly because Olly isn’t cold toward them. That’s not an accident — it’s a design outcome. And it’s one I’m genuinely proud of.
How do you measure whether Olly is actually working?
Two numbers: resolution rate — the percentage of conversations Olly successfully resolves on its own — and customer CSAT, how satisfied users were with the interaction.
A perfect day is when both rise together. High resolution alone isn’t success. If Olly is closing conversations quickly but customers leave frustrated, that’s the worst of both worlds — automation without care. What we’re after is resolution with satisfaction. Those two things have to move together, or something is wrong. Ultimately, success for us means proven efficiency that customers actually enjoy using – and those numbers tell us we’re right on track.
What would you tell a team at another company starting a project like this?
Don’t treat it as a “set it and forget it” project. Launch day is just Day 1.
The real work happens in the weeks after launch — analyzing real conversations, identifying where the AI is stumbling, refining the knowledge base, improving the responses. It’s an ongoing cycle, not a one-time build. The teams who treat it like a product — something that evolves continuously — are the ones who see it pay off in customer satisfaction.
The payoff is real. But it requires patience, iteration, and a willingness to keep learning. That’s true of most things worth building.
Anonina’s story is part of our “AI Innovators” series. Across OLX, people are finding new ways to use AI — to create, simplify, learn, and move faster. This series celebrates the colleagues who are sharing their discoveries and inspiring others to experiment along the way.
