The Third Generation of Services Sourcing Companies
I already covered some of the history of eDesk, Elance, and Fiverr in a previous essay, and the benefits and challenges that their new business model brought with them. And I mentioned how the third generation of services companies was largely created to solve those challenges.
But first you’ll note that we’ve skipped the first generation entirely, so let’s back up a step.
Long before the internet, businesses bought human capability through professional services firms, staffing firms, recruiting firms, and agencies. David Ogilvy and those before him built advertising agencies because companies needed advertising expertise and were willing to hire another organization to provide it. Consulting, accounting, law, engineering, staffing, and hundreds of other industries were organized around variations of the same basic idea: somebody has expertise or labor that you need, and you pay them to provide it.
There is probably a much longer essay to write about how ancient this model really is. Human beings were selling services long before anybody was selling SaaS subscriptions, and history is full of considerably less pleasant examples of people making specialized capabilities available for hire. Mercenaries are an obvious one. In that sense, selling a service may be one of the oldest business models in existence.
The second generation took this old model and put it online. oDesk, Elance, Upwork, Fiverr, and similar companies gave customers direct access to workers at a scale that traditional services firms could not. oDesk’s own description of the model was straightforward: employers could find people based on skills, work history, ratings, and rates, then hire, manage, and pay them through the same system. The experience was a little like walking inside a professional services firm and being allowed to choose the precise people who would work for you on a self-service basis. That gave customers considerably more control over selection and price, but it also meant they had to do more themselves. If you want to compare the engineer, select the engineer, negotiate with the engineer, and manage the engineer, somebody has to give you enough information and infrastructure to make those decisions. And the risk of hiring someone in an area you do not have expertise is considerably higher when you are taking on the evaluation process yourself.
The third generation of services companies grew out of that problem. Companies like Toptal, Andela, Arc, Lemon.io, and Gun.io began adding curation back into the experience. Toptal developed a multi-stage screening process; Andela evaluates technical and communication skills before matching technologists to companies; Arc screens communication and domain expertise; and Lemon.io explicitly screens for commercial experience, communication, self-management, and technical problem-solving. At Gun.io, we vetted people before presenting them, developed different ways of determining quality, improved matching, and tried to make the experience easier for both sides. Over time, varying degrees of do-it-for-you-ness returned as well. The customer still benefited from the enormous global labor market that the second generation had opened, but increasingly did not have to navigate all of it alone.
At the same time, other parts of the problem began to splice off into businesses of their own. Technical assessment became a standalone category, while international payroll, compliance, employer-of-record, contractor-of-record, and related problems became enormous categories unto themselves. Deel now offers global payroll, contractor management, and EOR services, while Rippling offers global payroll, EOR, contractor management, and related compliance infrastructure. What had once existed inside the walls of a traditional firm was increasingly available as specialized infrastructure that other companies could consume.
This is where talking about “generations” can become misleading, because none of these generations actually replaced the previous one. A traditional professional services firm can use a second-generation online services firm to find talent. A second-generation firm can use capabilities developed during the third generation. A third-generation company can provide talent to a first-generation company. A staffing firm can use an assessment provider and an EOR while sourcing someone through an online network and still present the resulting service to its customer under its own brand. These are historical generations, not mutually exclusive categories. In fact, some of this recursion existed remarkably early: oDesk itself began as a high-touch staffing operation before becoming a more scalable online model.
The common thread is that each generation has made it easier for smaller organizations to command capabilities that once required much larger ones. The first generation packaged expertise inside the firm. The second generation gave the customer much more direct access to the people providing it. The third generation retained that access while trying to solve the resulting problems of curation, vetting, experience, and complexity. Around all three, specialized companies emerged to handle more and more of the administrative infrastructure required to make the system work.
The result is something that would have seemed bizarre not very long ago. A founder with a laptop can find highly specialized people almost anywhere in the world, have someone else evaluate them, engage them through infrastructure provided by another company, pay them across borders, and assemble a team that would once have required an international corporation to support. Deel and Rippling, for example, explicitly sell infrastructure for hiring and paying people internationally without requiring a company to establish its own local entity in every market. You can effectively build a multinational company before you have built much of a company at all.
That is roughly where the last decade brought us, which leaves the obvious question of what happens next. We are now living through another technological shift that could change the economics of services, products, and corporations themselves, but the signals are contradictory enough that I am reluctant to declare what it means while we are still in the middle of it. Talk to an engineer using AI every day and he may tell you he has never been more productive. There is real evidence behind that sentiment: Research from the Google Cloud’s DevOps Research and Assessment group (DORA) published in 2025 found that more than 80 percent of technology professionals surveyed believed AI had increased their productivity. Yet controlled research has produced much less intuitive results; one study published by Model Evaluation and Threat Research group (METR) found experienced open-source developers working in familiar codebases actually took 19 percent longer when allowed to use early-2025 AI tools. Talk to a CFO and the question is different again: the companies are paying for tokens, licenses, infrastructure, and experimentation, and want to know when individual productivity gains become measurable improvements in revenue, margins, headcount, or output. The engineer can point directly to the hours he saved; but the CFO is looking at the company’s financial statements and asking where those hours went. DORA’s research increasingly treats exactly this translation from local productivity to organizational performance as a systems problem rather than merely a tooling problem.
Not very long ago, some people jumped to predicting the end of the software engineer, but the future is obviously more complicated. Whether AI ultimately means fewer engineers, more engineers, different engineers, or simply more efficient engineers is still an open question.
The contradictions extend beyond labor. Frontier AI companies command extraordinary valuations at precisely the moment competition among models is driving inference prices lower and making sophisticated software capabilities cheaper to consume. Recent reporting, for example, has put Anthropic’s valuation near $1 trillion while describing intensifying competition from cheaper models and falling AI costs; OpenAI, meanwhile, has continued raising enormous amounts of capital while projecting extraordinary infrastructure spending. Depending on who you talk to, we are still extraordinarily early in the adoption curve and have barely begun to understand what these systems will make possible, or companies are already spending a shitload of money on tokens, licenses, infrastructure, and experimentation without always being able to articulate the economic value they are receiving in return. Some people expect AI to create extraordinary abundance, others expect it to eliminate enormous categories of work, and still others worry about considerably more literal forms of destruction, with frontier-model safety and control now a substantial area of research and investment in its own right.
All of these things have implications. What does a corporation look like when an individual can command dramatically more productive capacity? What happens to services when the amount of human labor required to produce them falls? What happens to software products when the cost of producing software falls too? And what happens to the boundary between products and services if bespoke software becomes dramatically cheaper to create? I have opinions, as does seemingly everyone else, but there is some risk in trying to write history before it has happened. I’m no Cassandra or Nostradamus. We’ll see how this all plays out. That’s an essay for another time.
