As AI expands what product managers can research, prototype, and implement, businesses face a consequential choice about how to use that capacity. They can give product leaders more ways to investigate customer problems or increase their delivery responsibilities until there is little time left to understand those problems. Yaroslava Mazepina sees that choice as one of the defining questions in the next phase of AI adoption.
“I am a strong supporter of AI because it already helps us move faster,” she says. “What concerns me is the expectation that one person can absorb product management, design, and even development responsibilities while continuing to deliver features nonstop. Where does that leave the time to understand customers and the market?”
Her assessment draws on 14 years of experience across product management and quality in fintech and AI-powered marketing. She previously led the transformation of eToro’s tax-reporting system across more than 20 countries and now leads products at Pairfect. As a Women in Tech mentor and finalist in the Women in STEM category at the 2025 Women’s Business Awards, she brings both delivery experience and practical use of AI to her evaluation of the changing profession.
The expanding product role needs a clearer purpose
The broadening of responsibilities is already visible in product teams. Figma’s September 2025 research found that 55% of surveyed product builders were taking on tasks previously performed by others. Its report describes product managers prototyping concepts and engineers contributing to early design work.
Mazepina welcomes the ability to explore ideas more directly. An early prototype can make an assumption easier to examine, while a small implementation change can help a product manager understand technical dependencies. Her concern is what that expansion could displace. In vacancies she has reviewed for AI-enabled product roles,she has seen responsibilities from several disciplines combined in one role. She reads those descriptions as a signal of changing employer expectations. “The ability to build something yourself can improve the conversation with your team,” Mazepina says. “But if every new capability becomes another delivery obligation, you can end up spending less time on the decisions that determine whether the work is worth doing.”
Her assessment is that broader roles need explicit priorities. Organisations should establish which additional responsibilities improve the product leader’s contribution, which remain shared with specialists, and how time for customer discovery and market analysis will be protected.
Faster preparation can move work onto someone else
As more people contribute to implementation, Mazepina identifies a separate question: who inherits the work of checking what they produce?
A product manager can generate a detailed specification or prototype quickly, while designers and engineers must separate useful decisions from unsupported assumptions. The apparent completeness of the output may make that work harder to recognise.
“If my prototype saves me a day but costs the designer two days of clarification, the team has not gained,” she says. “The handover needs to make the next person’s work clearer.”
“The test is whether AI-assisted preparation reduces clarification and correction, or just transfers them to colleagues,” she says. “An exploration, a proposal, and an approved specification need distinct expectations, even when they look equally polished. The opportunity is earlier collaboration. The risk is a growing review queue that stays invisible when teams measure only the time saved by the person who created the first draft.”
Competitive urgency can obscure the case for AI
Mazepina also questions the pressure to introduce AI features before establishing the problem they should solve.
She sees businesses responding to genuine technological opportunity, but also to the fear of appearing slower than competitors. In that environment, the availability of AI can become the starting point for a roadmap.
“There is a pattern of ‘we have AI, we need to use AI, and we need to do it quickly,’” she says. “The missing questions are where it will add value for this business and its customers, and whether it will be cost-effective. We are entering an era in which the cost of service is measured not only in staff time, but also in tokens.”
Her objection concerns the sequence of decisions. A company can commit to an AI implementation and then ask the product team to justify it, reducing discovery to a search for supporting arguments.“Being enthusiastic about AI does not mean approving every proposed use of it,” Mazepina says. “You need enough understanding of the customer to decide where it belongs.”
She believes competitive urgency should sharpen the search for a valuable application, rather than settle the implementation before that search has begun.
More options do not necessarily make decisions easier
Another change Mazepina is watching concerns the growing ease of generating information, recommendations, and alternatives. Her assessment is that this capability can create additional decision work for customers. Producing a larger selection is relatively straightforward; helping someone judge which option fits their needs requires a deeper understanding of the decision.
She shares an example from Pairfect, where brands received five influencer recommendations matched to their campaign requirements. Customers repeatedly asked for more candidates, yet interviews revealed how difficult they found choosing whom to approach. Mazepina developed several hypotheses and used AI to critique them. She rejected both a “load more” option and a swipe-based interface because, in her judgment, they would preserve the expectation that customers needed to continue searching. She combined one of her ideas with an AI suggestion and added concise explanations drawn from the existing matching criteria.
The recommendations themselves remained unchanged. The explanations translated their relevance into the customer’s business context. “In the two weeks following release, the proportion of brands contacting an influencer rose to almost three times its previous level of 26%,” Mazepina says. The request was for more options, but the difficulty was choosing. We needed to make the reasons behind the selection understandable. I use AI to challenge my hypotheses and widen the options I consider,” she says. “Customer knowledge is what decides which one we test.”
Agent adoption raises the threshold for delegation
The move towards AI agents raises a further question: when should assistance become authority to act? In research published in February 2026, Microsoft reported that about four in five surveyed organisations were at or beyond the agent pilot stage. The survey covered 500 enterprise decision-makers across 13 countries and 16 industries; 32% reported preparing to scale, while 15% were revisiting their strategy after early tests. These findings describe self-reported deployment stages among large enterprises, rather than independently verified operational reliability.
Against that backdrop, Mazepina argues that expanding delegation requires evidence about how a workflow behaves outside its expected conditions. “Expanding autonomy requires evidence about what happens when something unexpected occurs,” she says. “Successful preparation does not automatically establish readiness to take the next action.”
She expects delegation to advance fastest where objectives are clear, errors can be detected promptly, and actions can be reversed. Decisions involving competing priorities or external commitments present a higher threshold. As an illustration, she distinguishes between a system narrowing a set of candidates, drafting an outreach message, and sending that message on a company’s behalf. Each step requires a different level of permission and confidence.
“A customer accepting a recommendation does not automatically mean they want the system to act for them next time,” she says. “Product leaders need to establish what has actually been delegated. Useful autonomy will develop task by task. The scope of authority should expand as the system demonstrates reliability, not before”
Productivity needs to be measured across the workflow
For Mazepina, the commercial test of AI is the total cost and quality of completing a useful task. That calculation includes staff time, model usage, review, retries, and corrections. Token expenditure is one component: a lower model bill does not establish efficiency if more human intervention is required, just as faster drafting does not establish a saving if rework increases.
“The question is what it costs to reach a result we can actually use, at the quality we need,” she says. “The first draft is only part of that calculation.”
She would compare complete workflows against their previous performance, examining both operating costs and whether the intended outcome is achieved. Customer value remains a separate requirement: an inexpensive process can still produce something customers do not need.
“This fuller accounting becomes more important as AI moves from experimentation into routine operations,” Mazepina adds. “That is where usage volumes and repeated interventions decide whether an apparent saving persists.”
The risk of becoming an approver
“The risk I see is workflows that turn the product professional into an approver of generated conclusions,” Mazepina says. “A coherent analysis can still misinterpret a customer, and recognising that error requires continued contact with the people the product serves.”
She ties that risk to where the role is heading. Over the next 12 to 24 months, she expects AI assistance to move from an optional individual practice towards an expected part of the product role.
“Documentation drafts, first-pass interview analysis, organising customer needs into jobs to be done: more of that will begin with AI,” she says. “Product managers will then be expected to check the interpretation, resolve ambiguity, and refine the material for decisions and delivery.”
She also expects coding agents to become a more routine way for product professionals to address small changes and bugs isolated from core functionality, subject to appropriate technical review. “AI can do more of the preparation and some of the implementation,” she says. “The product manager needs to cross-check the result and understand whether it moves the product in the right direction.”
Her prediction concerns a change in professional expectations, rather than the arrival of entirely new capabilities. “Knowing how to use AI will become part of the role,” she says. “Knowing when to challenge its output will remain essential to performing it well.”
For Mazepina, the leadership choice is how to turn expanded capability into better decisions: allocating some of the time saved to delivery, while preserving the investigation and customer understanding that give delivery its purpose.
