Here is a sentence that would have sounded absurd in 2019 and sounds merely ambitious in 2026: a company in Sydney would like to sell you employees who do not exist. Not staffing. Not consultants. Software that behaves enough like a worker that you manage it like one, assign it tasks like one, and - this is the part that makes the finance people lean forward - pay for it without adding a line to the org chart.
That company is Relevance AI, and its pitch is unusually clean for an industry that specializes in fog. Most AI products promise to make your existing people faster. Relevance AI proposes something structurally different: that the work itself can be handed off. The distinction sounds small. It is not. One version makes a worker more productive. The other quietly changes what a worker is.
A pivot with good timing
The founding is a tidy little parable about being in the right place when the ground moves. Daniel Vassilev, Jacky Koh and Daniel Palmer started the business in 2020 - legally, it is still OnSearch Pty Ltd - as a data and semantic-search tool. This was a perfectly reasonable thing to build. It was also, in hindsight, the wrong thing to build, or rather the right thing to build in the year before the interesting thing became buildable.
When large language models arrived in force, the founders did what founders are supposed to do and are usually too attached to their old idea to actually do: they pivoted, hard, into AI agents. The data-tool DNA turned out to be useful. Agents, after all, are mostly a problem of getting the right information to the right reasoning step at the right moment, which is a search problem wearing a costume.
Pivots are easy to admire in retrospect and painful to execute in the moment, because they require throwing away work that is functioning fine. The tell that this one was real, rather than a rebrand, is the growth that followed: a company does not accidentally go from a data tool to 40,000 agents a month. It has to rebuild the product around the new idea and mean it.
The no-code bet
The strategic choice that defines Relevance AI is about who holds the tools. The conventional assumption is that building AI agents is an engineering task, which means it lives behind a ticket queue and a six-week wait. Relevance AI rejected that. Its platform is designed so that the person who actually understands the workflow - the ops manager, the sales leader, the support lead - can build the agent themselves, on a visual canvas, without writing code.
This is a bet with a clear logic. The bottleneck in deploying AI inside a company is rarely the model. It is the translation between the person who knows what needs doing and the person who knows how to make software do it. Remove the translation layer and you remove the wait. "We enable training the agent to specialize in niche workflows," co-CEO Daniel Vassilev has said. "We're tool- and model-agnostic" - which is a polite way of saying they would prefer not to bet the company on any single foundation model, a sensible posture in a market where today's best model is next quarter's runner-up.
Workforce, Invent, and a rep named Bosh
The product line reads like an attempt to make an abstract idea concrete. Workforce is the flagship: a no-code environment for assembling multiple agents into a team that collaborates on a process from start to finish - the multi-agent system, in the jargon. Invent is the on-ramp, a text-to-agent generator that turns a plain-English description into a working agent, which is roughly as close to magic as enterprise software gets.
And then there is Bosh, which is the company at its most legible. Bosh is an AI sales rep - a business-development agent, co-designed with actual sales leaders - that researches prospects, writes personalized outreach, chases follow-ups, books meetings and updates the CRM. It does this around the clock, because it is software and software does not sleep, take PTO, or get discouraged after the fortieth ignored email. You can argue about whether that is a good thing. You cannot argue that it is not a clear thing.
The number that does the talking
If you want the whole thesis in one statistic, here it is: roughly 40,000 agents were created on the platform in January 2025 alone - reportedly a fortyfold jump year over year. Over the same stretch the company grew from 19 employees to more than 80 across San Francisco and Sydney. Small team, exploding output. That gap is not an accident; it is the entire argument. Leverage, in the Relevance AI worldview, is no longer a headcount question.
Investors noticed. In May 2025 the company closed a $24 million Series B led by Bessemer Venture Partners, with Insight Partners, King River Capital and Peak XV returning, bringing total funding to roughly $37 million. The money is going where the growth is - product depth and a US go-to-market push, with Vassilev relocating to San Francisco to open the office. Customers, meanwhile, span the range from scaleups to the Fortune 500, and include names you have heard of: Activision, Qualified, SafetyCulture.
The obligatory caveat
None of this is a guarantee. "Build teams of AI agents that do real work" is a sentence a great many well-funded companies are currently saying, and the category - Stack AI, Lindy, CrewAI, Sierra, and the model providers themselves - is crowded and loud. Agents that demo beautifully do not always survive contact with a messy CRM and an angry customer. The founders' forecast that every team will hire an AI agent, then full AI teams by 2030, is the kind of prediction that is either visionary or a slide, and you usually cannot tell which until later.
But the useful thing about Relevance AI is that its bet is falsifiable and specific. Either domain experts can build reliable agents without engineers, or they cannot. Either the work delegates cleanly, or it comes boomeranging back with a human attached. The company has planted its flag on the optimistic side of both questions - and, for now, the adoption curve is pointing its way.
Why "workforce" is the load-bearing word
It is worth dwelling on the vocabulary, because the vocabulary is the strategy. Plenty of companies sell "AI automation." Relevance AI sells an "AI workforce," and the noun does a lot of work. Automation implies a fixed script - a thing that runs when triggered and stops when finished. A workforce implies judgment, delegation, and a manager. The company leans into the second frame deliberately: you do not configure a macro, you hire an agent; you do not maintain a workflow, you manage a team. This is partly marketing and partly a genuine design philosophy, and the honest read is that it is both at once, which is usually how the good ones work.
The frame also solves a practical adoption problem. Buyers already know how to think about employees - how to onboard them, assign them scope, review their output, and fire them when they underperform. By dressing agents in that familiar costume, Relevance AI shortens the distance between "interesting demo" and "thing I actually deploy." The mental model ships with the product.
The unglamorous part: keeping agents on the rails
An agent that can act autonomously is also an agent that can act wrongly, at scale, without a human noticing until the CRM is full of nonsense. This is the quiet, unsexy engineering problem underneath the whole category, and it is where model-agnosticism earns its keep. By refusing to marry a single model, Relevance AI can route different steps to whatever performs best, swap providers as the frontier moves, and avoid the failure mode of a company whose entire product degrades the day its one vendor has a bad week. It is a hedge dressed as a feature, and a reasonable one.
There is also a governance angle that enterprise buyers care about even when they are too polite to lead with it: who watches the agents. A workforce you cannot audit is a liability, not an asset. The multi-agent framing helps here too, because a system of narrow, specialized agents is easier to reason about than one sprawling generalist trying to do everything and explaining none of it. Narrow agents fail in narrow, legible ways. That is not a slogan; it is just easier to debug.
The bottom line
Strip away the category noise and Relevance AI is making a concentrated wager on a single idea: that the scarcest input in a modern company is not capital or compute but the ability to get qualified work done, and that software can now supply some of that work directly. If the wager pays off, the interesting consequence is not that companies get more efficient. It is that the relationship between a company's output and its headcount - a relationship so stable we treat it as a law of nature - starts to loosen. A five-year-old company from Sydney does not get to decide whether that happens. But it has built a fairly specific instrument for finding out, and a lot of people are already using it.