AI made building and testing cheap. Deciding what’s worth building, and whether it’s right for the customer, is the whole job now. A conversation with Yaroslava Mazepina across the full product-quality remit, from the first customer insight to the last regression, on where AI genuinely lands and what the industry is overselling. Yaroslava has spent 14 years at the seam where product management meets quality assurance, owning the arc from what gets built and why, to whether it holds. That double vantage point produces an unusually clear read on the AI trend.
Before we dig in, tell us who you are, and where you start when you think about product and quality.
Thanks, I’m glad to share it. I’ve been curious since I was a child, always chasing the “why” behind everything, and that never really left me. I studied applied mathematics at Donetsk National University, and QA was the natural next step 14 years ago. At the time it was the perfect fit: I got to dig into every problem and use root-cause analysis to work out why something was broken. Then I became a QA lead and something shifted. I started seeing quality not only as whether we built the thing correctly, but through a wider product lens, what is actually good for the customer. That took me into product management, first in consumer products, then in fintech, and now the influencer-marketing space. Each experience has helped me hone my technical and product skills and deepened my understanding of different market dynamics and cultural approaches.
My professional path took me from Ukraine to the UK. Across all of it, my “is this good for the customer” has usually landed on one of two things, people’s money or their safety, two of the most important parts of everyday life. While working for a Big Four firm, I delivered a computer vision AI product that dramatically enhanced customer experience and significantly improved day to day operations. Later on I was invited by eToro, award-winning social trading platform in Europe, where I redesigned tax-reporting system for our customers: it lifted efficiency and satisfaction, saved real cost across several departments, and became the best tax season in the company’s history, which helped the company open new countries and tiers. Then Pairfect, an NVIDIA-recognised startup, approached me to head their product direction. The mission there is protecting businesses from fraud and wasted spend, and giving them clarity on the advertising strategies that are genuinely right for them.
So where I start is by joining the two things people keep separating. Product management asks “is this the right thing to build for this customer.” Quality assurance asks “did we build it right and does it hold.” For years those were different teams with different meetings. AI has just made both of them the same problem, because it made the middle, the implementation, almost free. When building and testing are cheap, the only thing left that is scarce is intent: knowing what is worth building, for whom, and what “good” actually means for them. That is a product question and a quality question at once, and it is exactly the seam I work at.
So where is AI actually helping you, in product and in quality?
Honestly, it’s a real acceleration, and I won’t pretend otherwise. On the product side, in my own teams, research synthesis that used to eat two or three weeks now lands in days, a first-draft PRD is minutes rather than a day, and it will monitor competitors and let you interrogate product analytics in plain language instead of queuing behind an analyst. On the quality side, same story, different tools: the biggest gift is the time we stopped losing to redundant, boilerplate scripts, plus generating edge cases a tired human misses and self-healing a script when a selector moves. For the mechanical weight of both jobs, the drafting, the summarising, the first-pass analysis, it is a serious accelerant.
And where do you think it’s being oversold?
I would say money. The vendors put a number on it, 40 to 50 percent cost reduction, and in its pure form that saving does not exist. It relocates. You draft the spec faster, then spend the savings discovering it was built on a shaky customer assumption. You generate the tests faster, then spend it chasing what slipped through. I’ve watched a team celebrate a velocity number while its escaped-defect rate quietly climbed. Google’s 2025 DORA research names the mechanism: AI lifts throughput and hurts delivery stability unless strong control systems sit underneath. It’s an amplifier, on the product side too. Point it at a team that understands its customer and it compounds that. Point it at a team that doesn’t and it just helps them build the wrong thing faster.
From your point of view, what is actively going wrong with AI adoption right now?
You know, the thing that genuinely worries me is the “vibe coding” enthusiasm. There is a whole movement celebrating shipping software you did not read, generated by a model and merged because it looked right. As a way to prototype or learn, fine. As an engineering practice at any scale it is the single most irresponsible trend in the field, because it treats the plausible-looking answer as the finished answer, which is precisely the thing my whole discipline exists to distrust. And it is not only a code problem. The product equivalent is shipping a roadmap the model generated because the reasoning read well, without anyone checking it against a real customer. Same failure, one layer up. The teams indulging either are running up a debt they have not measured yet.
You mentioned the output can be confidently wrong. What does that actually look like?
Take the hardest number we have, because it generalises. Veracode tested AI generation across more than 150 models: 45 percent of the output carried a real security flaw while the syntax stayed near-perfect. Confident, fluent, and wrong is a worse failure mode than obviously broken, because it passes the eye test. Lift that from code to a PRD and it is the same: an AI-drafted spec reads authoritatively and can rest on a customer belief nobody validated. The model is fluent at the artefact, code or requirements doc, and blind to whether the intent behind it is correct.
And where does that actually bite, day to day?
The failure I see is quieter than a crash, and it’s a product story and a quality story at once. A model makes a small, clean, well-commented change that passes the tests and the review, because it reads like it knows what it’s doing, and it breaks something unrelated in a way only someone who knew the product would have predicted. A tidy-up of a shared helper that turns out to be load-bearing in one market. Nobody, human or machine, connected “this small change” to “what this market’s customers actually see and expect.” The AI knew the diff. It did not know the product. And the tests miss it because they were written to the same narrow context the code was.
So what’s left for the human, then?
Owning the product decision, which is the thing AI cannot touch. What are we trying to do, who is the customer, what does quality mean for them rather than in the abstract. That is where product management and quality assurance are actually the same discipline: both are judgements about the customer, one made before we build and one made to check we served them. AI will generate a hundred plausible tests and three plausible roadmaps. Knowing which tests protect something a customer cares about, and which roadmap is worth a quarter of the company’s time, is the same underlying skill, and it’s the one the model does not have.
From your point of view, what can AI still not do?
I would say business context, critical evaluation, ethical responsibility, and collaboration across teams. AI is a brilliant contributor to each of those, but I don’t think it can own any of them, because you need a truth the model doesn’t hold, and real understanding and disagreement rather than a polite “you’re right.” And accountability cannot sit with something that cannot be accountable. AI is a tool. The product and quality professional is not replaced, but developed, and the whole team has to adapt around the tool rather than defer to it.
Isn’t a product quality manager saying “the human is irreplaceable” just defending her own patch?
Fair, and I’m aware how convenient it sounds. So here’s the version that isn’t self-serving. Most of the loop should be automated and I’d automate more of my own function tomorrow. I’m not arguing for humans everywhere. I’m arguing for one accountable human at the point of decision with a view of the whole product, and that’s a role, not a title. If a machine could genuinely hold the customer context and carry the accountability, I’d hand it over. But right now it can’t. It just makes the confident-wrong answer, in a spec or in code, more convincing, which makes the review harder, not easier. Capability going up raises the value of judgement. It doesn’t retire it.
What’s your take on all the new AI regulation?
I think it’s the right direction. We now have the EU AI Act, which governs AI systems, and the Cyber Resilience Act, which governs the security of products with digital elements placed on the EU market. Under the CRA, from 11 September manufacturers must report an actively exploited vulnerability within 24 hours, and non-compliance runs to 15 million euros or 2.5 percent of global turnover. Now, the statutes don’t literally say “know what your coding assistant generated,” that part is my inference. But the quality and compliance functions are converging, and I’d bet AI won’t shrink QA so much as regulation will grow it. And it reaches the product side too, because data governance and provenance, what data trained or informed the product, are product decisions as much as testing ones. The remit I see expanding is being able to evaluate what the AI produced rather than wave it through, documenting the strategy as an auditable artefact, and treating regulatory change as a core duty rather than the legal team’s problem.
If this isn’t mainly about speed, where’s the upside the industry is underrating?
Coverage and confidence, not velocity, and it applies to both halves. On quality, AI makes work that used to be economically impossible affordable for the first time, every locale, the full accessibility surface, adversarial cases, the messy edge of real data. On product, it makes discovery and validation cheap enough to explore ten customer hypotheses where you used to afford one. The real prize is not shipping the same narrow thing faster. It’s raising the floor of what you can understand about your customer and verify about your product, across the board. The teams that see that will build measurably better products. The teams chasing the cost cut will build the same products slightly cheaper and considerably more brittle.
If you had to leave people with one piece of advice, what would it be?
Know what “good” means for your customer and what’s worth building. If you invested in the people who understand the customer and can stand behind a decision, product and quality alike, AI just multiplied your advantage. If you thinned them out to bank a saving, you’ve automated your way to shipping the wrong thing faster, and soon you’ll have to prove to a regulator that you didn’t. Everyone has capable models now. That’s table stakes. The edge is the professional who knows when the fluent, plausible answer, in a roadmap or a release, is wrong. That was always the job. AI just removed every excuse for not being excellent at it.
The views expressed here are Yaroslava’s own and do not represent any current or former employer.
