Product, Distribution, and the New AI Growth Equation

AI has made it dramatically easier to build a product. You need fewer people to get started, prototypes can become working products much faster, and entire parts of development and marketing can now be handled with AI.

The uncomfortable part is that everyone else got the same advantage.

That tension runs through our latest conversation with Misha Galian, CEO of Boosters, the Ukrainian product company and venture builder behind JustDone and other AI-enabled products. Boosters has already built and spun off products including Promova and Avrora, while its current team is focused on developing a new generation of AI products.

Misha has seen several waves of the app economy firsthand. He joined Genesis in 2014 as a junior marketing analyst, moved into leading mobile products, later became CPO at BetterMe, and founded Boosters in 2018. That history makes his perspective on the current AI boom particularly useful. For all the talk about completely new rules, his argument is almost the opposite: AI has changed how quickly companies can execute, but the fundamentals of building a business remain surprisingly familiar.

The hard parts are still understanding demand, solving a real problem, finding distribution, and being willing to change direction when the market tells you that your original idea was wrong.

AI changed the speed, not the fundamentals

There is an understandable temptation to treat an AI-native company as a completely different type of business. Misha doesn’t see it that way.

The product cycle is still recognizable: understand demand, identify customer needs, create a solution, deliver value, and find a way to monetize it. AI can make the solution considerably better and compress the time required to build it, but it doesn’t remove any of those steps.

What it does remove is some of the cost of getting started. A small team can now launch something without the development and design resources that would have been necessary a few years ago. That creates more room for experimentation and makes previously unrealistic ideas testable.

But there is a catch. Building became easier for everyone. Distribution did not.

Advertising is getting more competitive, platforms are raising their standards, and thousands of new products can enter a category faster than before. The result is an odd inversion of the traditional startup problem: shipping software may no longer be the biggest barrier to entry, while getting that software in front of the right customers becomes increasingly difficult.

As Misha puts it, you can ship faster. That doesn’t mean you can get users faster.

For app companies, this shifts even more weight toward product-channel fit. A strong product with the wrong acquisition model can struggle just as much as a weak product with great marketing. AI accelerates the experiment, but it doesn’t tell you which market or channel will work.

JustDone found growth by abandoning its original audience

JustDone is a useful example because the product that scaled is not exactly the product Boosters initially planned to build.

The original audience was marketers. It seemed logical: the team understood marketing, generative AI was already changing content production, and there was an obvious set of text-related use cases to address.

The market disagreed.

Boosters started seeing much stronger demand among students, academics, and people working intensively with information and text. Rather than continuing to optimize the original positioning, the team changed direction.

And they did it quickly.

According to Misha, repositioning the product, marketing, and much of the surrounding experience took roughly three weeks. The team also discovered that its original distribution assumption was wrong. They expected Meta to be the primary growth channel, but JustDone behaved much more like an intent-driven product. Google and Bing search became more important, supported by email remarketing.

That combination — changing both the customer and the channel — is more significant than a normal messaging adjustment. It is a reminder that product-market fit and product-channel fit often develop together.

A user who is actively searching for help with an academic task behaves differently from someone encountering an AI writing ad in a social feed. The same product can look weak in one channel and compelling in another because the intent behind the interaction is different.

The lesson isn’t that search is better than paid social. It is that distribution should follow how the customer experiences the problem, not how the company originally imagined acquiring them.

Agility is mostly about being willing to abandon good ideas

“Move fast” sounds easy when a company is small and has little to lose. It becomes much harder once there is revenue, a growing team, existing customers, and several plausible directions competing for resources.

For Misha, agility has two less glamorous components. First, the company needs enough data to see what is changing. Second, leadership needs to be willing to act on that information even when doing so means abandoning previous assumptions.

That can mean pivoting an existing product, narrowing its positioning, starting another product for an audience you already serve, or splitting a strong team across several bets.

Boosters is still doing this today. JustDone is becoming more focused, while the company is simultaneously developing more specialized products around adjacent customer needs. The objective is not to chase every opportunity. It is to create dedicated teams that can stay focused on individual products while the company remains flexible at the portfolio level.

That distinction matters for multi-product companies. Agility doesn’t have to mean every team changes direction every week. It can mean building an organization where individual teams have clear focus while the business itself can move resources and create new bets quickly.

A thin AI wrapper is a starting point, not a moat

One of Boosters’ early fears was straightforward: what happens when ChatGPT or another frontier model simply adds the same capabilities?

In 2023, that concern was reasonable. Misha describes the early JustDone as a relatively thin layer on top of existing models. The team even expected that frontier LLM providers might make the product irrelevant within months.

That didn’t happen. But Boosters also didn’t keep building the same product.

As the team learned more about its users, it identified jobs that general-purpose models did not handle deeply enough, including plagiarism checking, AI detection, and academic research. Using specialized third-party providers could cover some of those needs, but at significant cost. Misha says those services could have cost the company around $2 million per year.

So Boosters began developing its own technology. Today, Misha says the company has around six proprietary engines and models trained for specific use cases.

This changes the relationship with frontier models. Instead of trying to beat OpenAI, Google, Anthropic, or other model providers at general intelligence, Boosters can use those models where they make sense and build deeper technology where the customer problem requires it.

That is a much more defensible position than simply putting a new interface around the same underlying model.

General AI gives you a city bike. Vertical AI needs to build the road bike

Misha uses a useful analogy to explain where specialized AI products can still compete.

A general-purpose model is like a good city bike. It works for many people and can handle a wide variety of everyday journeys. Most users don’t need anything else.

But someone training for a triathlon needs a very different machine.

That is where specialized AI products have room to exist. They need to serve customers whose workflows are specific enough that a general-purpose interface becomes inconvenient, incomplete, or too shallow.

For Boosters, academia is one of those areas. A student, researcher, professor, or PhD candidate isn’t simply asking an LLM to produce text. The workflow can include discovering academic sources, checking citations, researching papers, structuring arguments, validating originality, and moving between several tools.

A specialized product can bring more of that workflow into one environment and design the interface, models, and functionality around those particular jobs.

This also provides a useful filter for founders considering a vertical AI product. If ChatGPT, Gemini, or Claude can already solve the customer’s problem comfortably in a general-purpose interface, there may not be enough room for another product. The opportunity becomes more interesting when customers need the equivalent of a road bike: something designed around a specialized workflow rather than general capability.

Customers still want solutions, not AI

The AI industry often assumes that users are becoming increasingly sophisticated in how they interact with models. Some are. Most customers, however, are considerably more pragmatic.

They have a problem and want it solved.

That creates a gap between what technically advanced users imagine AI adoption looks like and what mass-market adoption actually requires. A blank prompt box can be extremely powerful, but it also asks the user to understand what the system can do, how to instruct it, how to iterate, and how to judge the output.

A specialized tool removes some of that cognitive work. Instead of asking users to design their own workflow through prompts, it packages intelligence around a recognizable job.

This is relevant beyond AI products. Lifecycle marketing has dealt with the same principle for years: more options do not automatically create a better customer experience. Good onboarding, segmentation, triggered communication, and contextual guidance exist partly because customers should not have to discover every useful action themselves.

AI may make the underlying system dramatically more capable. The product still has to turn that capability into something people understand and use.

AI transforms a category when it can replace the workflow

Not every industry is being changed by AI at the same depth.

Misha’s distinction is useful: the biggest transformation happens where AI can replace a meaningful part of the workflow rather than simply improve an individual task.

Software development is an obvious example. AI can now participate across coding, testing, documentation, prototyping, and quality assurance. It is not just suggesting a better sentence or saving a few clicks; it can take over substantial parts of how the work gets done.

The same potential exists in fields with large amounts of structured information, established rules, and repeatable knowledge work.

Other categories are more exposed to commoditization. Copywriting is one example discussed in the episode. AI has made content teams far more productive, and Boosters itself now needs fewer people to produce the same volume of certain types of content. But that doesn’t necessarily create a strong standalone AI business. If a general model can already perform most of the task, a new product that only improves the same workflow slightly has a very thin moat.

That is the difference between AI as a feature and AI as a structural change. One improves productivity. The other changes how the job gets done.

The model is becoming less important than the customer

When every AI company talks about models, it is easy to assume that model choice determines who wins.

At the application layer, Misha sees two factors as more important: understanding the customer and building distribution.

Frontier models are already extremely capable for many common tasks. Application companies can switch between providers, combine models, use open-source alternatives, or train specialized technology where necessary. That makes the model an important part of the infrastructure, but not necessarily the reason a customer chooses one product over another.

The harder advantage comes from understanding a specific group of users deeply enough to build a better solution around their workflow.

Then comes distribution.

Misha argues that even strong AI products cannot assume organic demand will carry them. Competition for attention is expensive, and companies need acquisition expertise, analytics, automation, and enough data to understand where growth is actually coming from.

AI hasn’t removed the classic business problem of reaching customers. If anything, easier product creation has made that problem more visible.

Marketing may be the first department AI really rewrites

Inside Boosters, marketing is one of the areas where AI has already changed the workflow most significantly.

The company has experimented with a user acquisition management agent that can handle much of the operational cycle around paid campaigns: taking creatives, launching campaigns, adjusting settings based on performance data, and changing budgets. Similar automation can be applied to search campaigns, from managing keywords to creating landing pages around specific keyword groups.

This moves AI beyond content generation.

The more interesting change is decision automation. Instead of helping a marketer write five headlines faster, the system begins participating in the continuous loop of launching, measuring, adjusting, and stopping campaigns.

That has implications for lifecycle and CRM teams as well. The same direction is already visible in messaging automation: systems increasingly help teams move from manually building individual campaigns toward defining goals, rules, audiences, and constraints while automation handles more of the execution.

The marketer’s role moves upward. Less time goes into repetitive configuration; more goes into deciding what should be tested, how success should be measured, and whether the machine is making sensible decisions.

AI is spreading through the company, not just the product

Marketing is only the beginning. Boosters also uses AI across software development, QA, documentation, product requirements, prototyping, and recruitment. Product managers can create working prototypes before handing features to engineering, while recruiting teams can summarize candidate interactions across different stages and use that context to make decisions faster.

The company is also looking at how AI can support internal HR processes such as feedback collection, performance reviews, and development planning.

Still, Misha is cautious about turning AI adoption itself into a KPI. Tracking how many AI credits a developer consumes, for example, says little about whether those credits created better work. Such metrics are easy to optimize without improving the outcome.

The more useful question is whether the process itself is getting better.

That is an important distinction for companies trying to become “AI-first.” Adoption is not the objective. Faster learning, better decisions, lower operational friction, and stronger output are.

Smaller teams can do more, but people are still the decision bottleneck

Boosters operates with roughly 70 people across several businesses and products. Misha sees that efficiency as part of the company’s operating model rather than something created by AI.

This shapes his view of the current debate around AI and headcount. In companies that were already operating with lean teams, there may be less obvious work to eliminate. AI gives those teams additional execution capacity, but it doesn’t solve the harder constraint.

“The problem is not about execution,” Misha argues. The real bottlenecks are ideation, hypothesis generation, and decision-making.

AI can give a strong operator more “hands.” It can produce variations, process information, create assets, generate code, and run pieces of a workflow. But someone still needs to decide which problem deserves attention, which hypothesis is worth testing, and which result is good enough to scale.

That is why the one-person billion-dollar company remains more useful as an extreme example than an operating plan. Smaller teams will undoubtedly be able to build more than before, but competition moves at the same time. If every company gains more execution capacity, the advantage shifts toward the quality and speed of decisions.

In a Nutshell

AI has changed the economics of building products, but it has not made the old rules irrelevant. Boosters’ experience with JustDone is almost a case study in that contradiction. The team could build and pivot faster because of AI, yet growth still depended on recognizing that the original audience was wrong, finding a better product-channel fit, understanding specific customer workflows, and investing in distribution.

The same logic applies to the next generation of AI products. General-purpose models will keep getting better, which makes shallow application layers increasingly difficult to defend. Specialized products need to go deeper: into customer context, proprietary technology, domain-specific workflows, or distribution advantages that frontier models are unlikely to prioritize.

Inside companies, meanwhile, AI is moving from assistant to operator. It can already participate in campaign management, coding, QA, prototyping, research, and other workflows. But faster execution only shifts the bottleneck. When teams can build almost anything faster, deciding what is worth building becomes more important.

That may be the most useful way to think about the AI race. The winners will not necessarily be the companies with the most AI, the biggest teams, or even the most sophisticated model. They will be the ones that notice changes early, understand their customers better, distribute effectively, and are willing to change direction before the market forces them to.