Before Daniel Vassilev had an AI company, he had an annoyance. As a school student, he wanted to keep listening to YouTube music while using something else on his phone. So he built an app in about a week. The job was plain, the timing was right, and the audience arrived: the app reached roughly a million users. A later companion for Pokémon Go found several million more. Neither project required a manifesto. People understood what the software did because the problem was already sitting in their hands.
That early run is the useful preface to Relevance AI, the company Vassilev co-founded in Sydney in 2020 with Jacky Koh and Daniel Palmer. Today its vocabulary is agents, tools, models, memory and multi-agent systems. Underneath the vocabulary is the same product instinct. Find a job people are already trying to complete. Give the software enough context to help. Then keep removing the moments when the user has to push it along.
The scale has changed. A teenager’s consumer app could win with one sharp feature. Enterprise software has to survive permissions, handoffs, existing systems and the local customs that never appear in an employee handbook. Vassilev’s career has been a journey from making one obvious task easier to asking whether software can accept responsibility for an entire stretch of work.
A co-founder relationship before the company
Vassilev was born in the United Kingdom and moved to Australia. His father worked in information technology, giving him an early view of computers as things that could be understood and manipulated, not sealed boxes. He began by automating repetitive actions in video games. Soon he was making products for other people.
He met Koh at school. A conversation at a high-school party helped set off a collaboration that would continue through several projects. Their skills were complementary: Vassilev leaned toward product and engineering; Koh toward artificial intelligence and machine learning. Vassilev later studied advanced computer science at the University of Sydney, where he received a Microsoft Research Prize in 2017 for top performance in a final-year computer-science course.
The friendship matters because Relevance AI has required more than one reinvention. A co-founder pairing built around agreement alone can become brittle. Vassilev has spoken instead about the value of being challenged, dividing responsibilities and staying aligned on the larger direction. The product could move because the partnership had room for argument.
“The simple way to think about agentic AI for us is whether it can now make decisions that are dynamic.”Daniel Vassilev, RevOps FM, 2025
The nouns kept changing
Relevance AI did not begin as the home of an “AI workforce.” Its early public identity was a developer-first vector platform. The problem was unstructured data: text, images, audio and other material that companies accumulated but struggled to search or use. Embeddings could turn that material into mathematical representations, yet the infrastructure for applying them was awkward. The founders built the missing layer.
It was a technically serious starting point, but customers rarely wake up wanting better vector experimentation. They want an answer, a decision or a process completed. Relevance moved steadily closer to that result. Infrastructure became no-code chains and workflows. The arrival of capable transformer models opened another step. Software could now interpret instructions and choose among possible actions instead of following only rigid if-this-then-that rules.
This progression is easy to narrate in retrospect. Living through it is less neat. Each move asks a startup to explain itself again, rebuild parts of the product and risk confusing the people who understood the previous version. The constant was a belief that machine learning should automate business processes, not remain a component that technical teams admire from a distance.
The expensive choice to stay broad
After ChatGPT appeared, Relevance AI’s founders heard a familiar piece of startup advice: pick one use case. An agent that writes cold emails is simple to describe. A platform for creating many kinds of agents asks the buyer to imagine more. Vassilev has acknowledged that the broader demos were harder to explain and the ideal customer was less precise. For roughly two years, he wrote, it was unclear whether declining the advice had been correct.
Their case for staying horizontal came from an observation about jobs. Put two people with the same title into two different companies and their onboarding may barely overlap. The title is standardized; the work is not. Each organization carries its own tools, thresholds, language, history and judgment. A narrow product with configuration toggles can cover a common slice. It cannot easily absorb all of that context.
So Relevance AI chose a metaphor closer to hiring and training. A company should be able to give an agent instructions, access to the right tools and the context for a particular role. Several specialized agents could then pass work among themselves. The horizontal platform was not breadth for its own sake. It was an attempt to make company-specific context part of the product.
Autopilot, with a return address
Vassilev draws his most important distinction between copilot and autopilot. A copilot remains beside a person, waiting for each prompt and helping with the next step. Autopilot accepts a delegation. It can make bounded decisions, use tools and return when the task is complete or when it needs judgment. In his definition, autonomy is not the removal of people. It is a different shape of collaboration.
“Autopilot does not mean without humans,” he said in a 2025 conversation. The practical comparisons come from ordinary management. Engineers use pull requests and code review. Sales teams hold deal reviews. Organizations already build permissions, checks and escalation paths around human work. Agent systems need their own versions of those controls.
This makes his AI argument less theatrical than its “workforce” label can sound. The core interface is a handoff. Can the system understand the brief? Does it have the correct access? Can it recognize a boundary? Will it leave an observable trail? Does it know where the work goes next? A useful agent is measured after the demo, when a process changes or an edge case arrives.
“Autopilot does not mean without humans.”Daniel Vassilev, RevOps FM, 2025
Funding the next layer
The market eventually moved toward Relevance AI’s language. In January 2025, users created 40,000 agents on the platform. That May, the company announced a $24 million Series B led by Bessemer Venture Partners, with King River Capital, Insight Partners and Peak XV participating. Reported total funding reached $37 million. Customers named publicly included Qualified, Activision and SafetyCulture.
The round arrived alongside two products that made the long thesis visible. Workforce offered a visual canvas for arranging agents into collaborating teams. Invent let a user describe an agent in text and generate a starting point. Both aimed to move the builder beyond engineers and into the hands of subject-matter experts, the people who already know the process and can recognize when the output is wrong.
Vassilev also relocated to San Francisco to open a U.S. office and work closer to customers and partners, while Relevance AI retained its Sydney base. It was a geographic version of the product instinct: shorten the distance to the people using the thing. The company had begun as an Australian infrastructure startup. It was becoming a cross-Pacific enterprise software company.
The old product lesson inside the new category
AI agents invite abstract forecasts about labor, productivity and the future of the firm. Vassilev participates in that conversation. He wants organizations to be limited by their ideas rather than their headcount, and he has described delegation as a skill that will become more valuable. Yet the more transferable lesson in his story is grounded in product work.
The background-music app succeeded because its job was instantly legible. Relevance AI took longer because its job had to be discovered layer by layer. First, make messy data usable. Then connect operations. Then let models choose. Then coordinate specialized agents. At every stage, the company moved from capability toward consequence.
There is also restraint inside the ambition. Delegated software needs an address to return to. Context has to be taught. Tools need boundaries. People remain responsible for designing the system and reviewing important decisions. The handoff only works when both sides understand where it ends.
Vassilev’s earliest apps waited for taps. The software he is building now is supposed to accept a brief and come back with completed work. Between those two interfaces sits his whole career so far: a school friend who became a co-founder, millions of consumer users, several product identities, and a long bet that the best software will be judged by what a person no longer has to shepherd minute by minute.