Fiverr Goes All-In on AI: Empowering Creators, Not Replacing Them, with Micha Kaufman, CEO of Fiverr
Summary
Kaufman’s labor-market call is that AI has shifted the baseline one rung down. Easy tasks disappear, hard tasks become easy, and formerly impossible tasks become the new hard; entry-level, mediocre, and even average performance therefore becomes vulnerable. He hedges the analogy, but expects a dot-com-like shakeout in which “999 companies out of a thousand” doing interchangeable AI work disappear and layoffs potentially snowball.
The decisive divide is AI mastery, not humans versus machines or junior versus senior workers. Kaufman expects both output per unit of time and quality to rise, arguing that “freelancers that are going to master AI…are going to replace those who don’t.” Juniors can still win because they often have current training and stronger hacker instincts; a seasoned candidate who still works “like it’s 2024” may not clear Fiverr’s hiring bar.
Freelancers may be better positioned than salaried employees because market incentives have already trained them to adapt. They sell outputs, feel productivity gains directly, continually learn new tools without an employer’s safety net, and are practiced at finding the next customer or reinventing their offer. Kaufman says Fiverr’s freelancers adopted AI faster than its own team, while displaced full-time employees may find rebuilding “mentally harder.”
Fiverr’s horizontal marketplace gives it a portfolio hedge against category destruction. The platform expanded from roughly 6 categories in 2010 to nearly 800, while moving from bite-sized microbusiness work to multidisciplinary enterprise projects, including transactions around $1 million lasting 6–12 months; Kaufman says perhaps 80%–90% of the Fortune 500 are customers. His bet is that newly invented jobs will appear on Fiverr first because supply can rotate rapidly as old categories vanish.
Fiverr Go is designed to scale creators rather than substitute for them. Its personalized creation models use an individual freelancer’s body of work, preserve control and compensation, and are released only where Fiverr seeks at least 90% fidelity to that creator’s signature; its personal assistant handles intake, qualification, offers, availability, and sales around the clock. Kaufman calls the creation model “a conversion tool” and says the assistant is already converting substantially better than sellers working alone.
Kaufman’s deeper concern is that uncredited model training could destroy the incentive system behind cultural and technological progress. He traces the acceleration of human advancement to the printing press, wider knowledge sharing, recognition, and commercialization, then asks what happens when AI “eats up everything” while returning neither attribution nor reward. His governing question is “who’s in the center?”—humans using AI as a tool, or humans becoming training material for an increasingly autonomous entity.
Human services retain demand through time arbitrage, competitive differentiation, and complexity—not merely solidarity with creators. AI gives everyone superpowers and therefore becomes “ground zero,” leaving taste, judgment, orchestration, and unusually resourceful hacker behavior as the scarce layer. Kaufman expects repetitive operation to disappear naturally, while mission-critical work and distinctive brands may continue paying for human craft because generic AI output is available to everyone.
Inside Fiverr, Kaufman is using direct experimentation to reset velocity expectations. When a team estimated three weeks for a build, he gave himself three hours and returned with a working result despite calling himself a rusty, mediocre developer; he now sees potential to move “five times faster” or pursue five times as much. The memo frightened some employees, but its first 36 hours also triggered demonstrations, KPI discussions, and requests for guidance—evidence, in his view, that radical candor can mobilize an aligned organization.
Deep dive
1. AI has moved the difficulty ladder one rung down
Kaufman’s wake-up call begins with a new baseline: “things that were considered easy tasks will no longer exist,” hard work becomes the new easy, and previously impossible work becomes the new hard. Anyone operating at entry-level, mediocre, or average quality now “has an issue.”
This is not presented as a Fiverr-specific threat. AI is “coming for all of us,” including Kaufman’s own CEO role, and people unable to increase both output per unit of time and quality may find themselves “out of the industry.”
His deliberately hedged macro analogy is the dot-com crash: pets.com has become “pets.ai,” trivial applications are proliferating, and perhaps “999 companies out of a thousand” will vanish because substitutes become cheaper and faster. If layoffs spread simultaneously, displaced workers may find few companies still hiring.
Kaufman calls the memo an act of “radical candor” rooted in care: people deserve the truth early enough to respond. His operational test is blunt—“if you’re still working like it’s 2024, you’re doing something wrong.”
2. Freelancers may be more resilient than full-time employees
Labenz initially frames marketplace work as more exposed because short engagements often carry less client context. Kaufman pushes back: freelancing has evolved since 2010 from an activity between jobs into a legitimate career, and Fiverr relationships can persist for years.
Freelancers lack organizational career development and therefore develop “a hunger to teach themselves.” They track current software, techniques, and practices because their income depends directly on delivering a competitive output.
Kaufman’s strongest proof point is internal: freelancers adapted to AI “much faster than our team.” They can reserve their own time for learning and already know how to reinvent an offer, acquire another customer, and move to the next project.
Full-time employees possess more support but may also be more brittle. If displaced, Kaufman argues, rebuilding from the ground could be “mentally harder” than it is for freelancers accustomed to recurring uncertainty.
3. Fiverr’s horizontal supply is a hedge against category destruction
Fiverr launched with roughly 6 categories and now has close to 800, overwhelmingly digital and often creative: coding, design, music and audio, data, AI, writing, and marketing. Categories have disappeared along the way, but hundreds more have emerged.
The platform began with microservices for microbusinesses, then expanded through SMBs into enterprise work. Kaufman estimates that 80%–90% of Fortune 500 companies are customers, with engagements ranging from cheap, bite-sized services to approximately $1 million projects spanning six months or sometimes a year.
Fiverr is consequently “more than a marketplace”: complex work may require business tools, agencies, multiple specialists, and orchestration across disciplines. Kaufman’s objective was to extend the platform from entry-level supply to sophisticated needs without abandoning either end.
Category churn is normal; only its velocity is changing. Kaufman cites an Economist study saying that, in 2018, 60% of existing jobs had not existed 50 years earlier, then argues that any job “invented a minute ago” will probably appear on Fiverr first.
4. The scarce AI worker will be a hacker, not merely a tool user
Labenz proposes “AI scout” as an emerging profession: someone who knows current capabilities, limits, tools, and comparable solutions. Kaufman agrees there is demand but notes that an agent could also perform scouting; his broader call is that “the future belongs to the hackers.”
His hacker is relentlessly current, pushes new technology to its boundaries, breaks it, connects mismatched tools, and discovers use cases the developers missed. Those people were not the audience for his warning memo because “they’re there already,” and he expects them to be paid generously.
Fiverr is already seeing work in model training, fine-tuning, bespoke GenAI applications, prompting, and experiments using AI-capable mobile hardware. These sit alongside continuing demand for ordinary design, writing, and marketing.
Kaufman views foundation AI as a commoditizing substrate: it launched free, cannot get cheaper than free, and is racing toward tiny prices for GPU cycles. The “golden era” should arrive in applications built above it; today’s market remains pre-WhatsApp, pre-Instagram, and pre-Uber.
5. AI gives everyone superpowers—and therefore no durable edge
Kaufman rejects the idea that merely using AI differentiates a buyer: “all of us” receive the same superpower. If both Nathan and Micha can generate a design, AI has simply elevated “ground zero.”
The first enduring reason to hire another person is time arbitrage. A buyer might be capable of learning a task over two days, but paying a specialist $200 to deliver it in two hours can still be economically obvious.
The second is competitive edge: a master may possess better taste, deeper experience, and command of better tools. Buyers seeking something superior to what everyone else can generate still need differentiated skill.
The third is complexity—work may be multidisciplinary, difficult to orchestrate, or simply not worth internalizing. AI changes what counts as easy or complex, but Kaufman argues that these three motivations “haven’t changed.”
6. Creator incentives are infrastructure for human progress
Kaufman’s philosophical premise starts with a time scale: Homo sapiens have existed for about 300,000 years, but advancement began accelerating exponentially only around 500 years ago. He attributes that inflection to the printing press and wider access to knowledge.
Sharing works because contributors receive something back. Open-source programmers may appear altruistic, yet gain recognition, opportunities, and the satisfaction of seeing others extend their work; Britain’s 1710 copyright framework similarly sought to encourage learning by attaching identity and commercialization rights to creation.
Generative AI disrupts that bargain by consuming “everything everyone ever created,” regenerating it, and returning neither attribution nor compensation. If a work has “zero value” the moment it is published, Kaufman fears creators may stop sharing freely.
The governing question is “who’s in the center?” Kaufman wants humans at the center and technology serving them; today, free AI makes humans training material—“if you’re getting it for free, you’re the product”—raising the question of whether the user or the model improves more.
7. Fiverr Go turns each creator into a production house
Fiverr’s alternative is a model trained on one creator’s own corpus rather than an effectively infinite pool. The resulting output should carry that person’s signature, give customers a predictable preview, and leave control and economic reward with the creator.
Kaufman’s analogy is Michelangelo’s workshop: having nine or ten assistants did not make its output less Michelangelo-like. Personalized generation can similarly “turn one person into a production house” without claiming that the machine replaces the originator.
Fiverr Go’s assistant is salesperson, intake desk, and business partner. It answers repetitive questions, assesses fit and availability, gathers requirements, helps price offers, researches positioning, and can wake a freelancer only when a promising customer merits personal attention.
The assistant is trained on the freelancer’s profile, offers, orders, and sometimes tens of thousands of customer conversations. Fiverr openly identifies it as AI but deliberately humanizes the interaction; Kaufman says buyers sometimes ask whether it is really AI, while conversion already exceeds that of sellers responding alone.
8. Product scope is governed by fidelity, not AI novelty
The assistant was the obvious first product because it removes work creators dislike and gives them more time to create. Its constant availability also lets freelancers sleep without leaving potential buyers unattended.
Fiverr identified about 60 service categories where GenAI might apply, but Kaufman set a threshold: if the system cannot achieve roughly a “90% score of accuracy” to a creator’s specific signature, it should not represent that creator.
Initial creation models therefore concentrate on large categories where both marketplace demand and model quality are meaningful: graphic design, voice-over, music production, and complex text such as business planning. Kaufman stresses that AI still cannot serve every category reliably.
Go is managed like an app, with updates effectively every day: new versions, capabilities, bug fixes, lower costs, and simpler workflows. Kaufman repeatedly refuses to call the present design final—community acceptance, not the launch keynote, will determine when Fiverr has it right.
9. Personalized models inherit AI’s versioning and control vulnerabilities
Labenz’s Replika example supplies the key product risk: an upgraded foundation model can become smarter while ceasing to feel like the same personalized entity. A creator who approved one version may find that its successor no longer represents them.
Kaufman accepts this as “a big vulnerability of AI in general.” Fiverr takes responsibility for managing the underlying technology and builds additional technologies above base models to reduce exposure; it must re-run quality and stress tests whenever a provider updates one.
Capability is only half the problem; obedience also changes. Kaufman contrasts ChatGPT 3, which could be “insubordinate,” with the more compliant version 4, while warning that LLM boundaries remain “embarrassingly easy to break.”
At Fiverr’s potential scale, every creator may impose distinct prohibitions—no political work, no material that is not safe for work—and the AI must honor each one. Labenz notes that model builders themselves often discover new capabilities alongside users; Kaufman agrees that users are simultaneously exposing hacks and informing better successors.
10. No one yet knows whether creator-owned models survive AI parity
Labenz asks the hard commercial question: if generic AI eventually matches top human quality, will buyers keep paying creator-owned models out of solidarity, or will law need to preserve human centrality? Kaufman’s answer is an explicit “I don’t know.”
His analogy is digital media’s path from Napster and unauthorized YouTube uploads to Spotify, Apple Music, and YouTube Premium. Consumers pay when legitimate access becomes easy and at least as pleasant as infringement, but he will not assume creator models follow the same trajectory.
Two months after launch, Fiverr was already changing its model. One learning Kaufman highlights is that people may enjoy generating samples themselves yet still want the opportunity to talk to a human for business needs requiring professional advice, a good eye, and a good hand.
Kaufman expects regulation eventually to update outdated copyright constructs, but distrusts confident long-range predictions: anyone certain about ten years from now is a “charlatan” when even two years is hard to extrapolate. His preferred response is first principles plus rapid iteration.
11. Repetitive operation should disappear while judgment remains scarce
Asked what human work deserves protection, Kaufman rejects the premise of preserving tasks for their own sake. If a computer performs repetitive work better, humans should stop doing it, just as accountants moved from pencil-and-paper balance sheets to spreadsheets and sophisticated software.
Human creativity requires no artificial shield, in his view. Better taste, deeper knowledge, and genuine originality will continue attracting buyers; if someone’s only advantage is operating a tool that everyone can now operate, “you don’t need people for that.”
High-end brands may even move toward more bespoke work because generic AI is ubiquitous. Kaufman cites Coca-Cola’s AI advertisement: the headline centered AI, but the production reportedly involved roughly 500 people, three of the best agencies in the world, and legal and compliance teams.
Labenz pushes back that advanced systems may possess meaningful understanding. Kaufman sees better algorithms, patterns, zeros, and ones rather than human-like comprehension; machines can generate impressive work without knowing what an image is, leaving humans to judge whether outputs are good for humans.
12. AI should raise marketplace liquidity and organizational velocity
On marketplace efficiency, Labenz asks whether AI can reduce search, discovery, matching, negotiation, trust, and conversion friction; Kaufman gives his least hedged answer: “yes.” Fiverr is already using it deeply in matchmaking, though he declines competitive detail, and expects Amazon and other platforms to attempt bolder redesigns.
The internal memo’s reception varied during its first roughly 36 hours: confirmation for advanced users, validation of existing fear for some, and a true wake-up call for others. Employees brought demos, asked how to measure progress, and sought KPIs through Kaufman’s open-door discussions.
Kaufman also leads through direct experimentation. When a team estimated three weeks for a build, he gave himself three hours, returned with a result despite describing himself as rusty and mediocre, and reset the discussion; he now imagines Fiverr moving five times faster or attempting five times as much.
Hiring follows the same test across engineering, marketing, and even legal: current technical fluency matters more than whether someone is seasoned or junior. University may still offer social life and an alumni network, but obsessively curious people can learn freely online; Kaufman closes not with doomsday, but optimism—“I’m having fun” because the upheaval is intellectually challenging.