Services are the new software.
For every dollar the world spends on software, it spends roughly six on services. For two decades, technology only ever nibbled at the smaller number. AI is finally coming for the larger one, and in doing so it is quietly redefining what an AI company even is. This is a field guide to that shift, from selling tools to owning outcomes, and where a full-stack AI and data firm like Welzin fits in it.
The question changed from what can we sell you to can we just do the work.
For fifteen years the playbook was clear: find a business, build software it can rent, charge per seat. Then the models got good enough to do the work itself, and the most ambitious builders stopped asking what tool a law firm or an insurer might buy. They started asking whether they could simply be the law firm, or the insurer. Y Combinator now says it in plain language in its Requests for Startups.
AI models are improving really fast, and they are now able to do complex work far beyond engineering. What we are excited about now is the next step: AI-native companies that do not sell software, they sell the service.Y Combinator · Requests for Startups
Sequoia drew the cleanest line through it. There are copilots, which sell a tool to a professional who stays on the hook for the result, and there are autopilots, which sell the finished work itself. A copilot competes for the software budget. An autopilot competes for the far larger labor budget.
The copilot
Sells a tool. A human still does the job and owns the outcome.
- Priced per seat, per month
- Competes inside the software budget
- Value captured: roughly 1 to 5 percent of a worker's output
- Moat often leased from the model provider underneath
The autopilot
Sells the work. The system does the job and owns the result.
- Priced per outcome, per task, against a salary
- Competes inside the labor budget
- Value captured: 25 to 50 percent of a worker's output
- Moat owned: data, workflow, and the operating loop
This is not a thesis on a slide. It is already funded, shipped, and operating.
The temptation is to treat all of this as venture-capital theatre. The funding record says otherwise. In 2025 alone, dozens of vertical and services-led AI companies each raised nine figures, and the ones drawing the largest rounds are not selling tools. They are doing the work.
The ladder: what owning each layer makes you
valuation
Series E
per letter
valuation
pricing
valuation
Series C
CAC payback
valuation
seed
contracted ARR
per expert
market
Priced against a salary, not a seat.
The pricing model is where the whole thesis becomes real money. When the software does the work, the per-seat meter stops making sense. A support agent billed at $115 a month per seat is a strange unit when the AI resolves the ticket and there is no seat. So pricing slides along a spectrum, away from access and toward outcomes.
The trade is real and worth naming. AI inference is not free, so a services-as-software firm runs at gross margins nearer 65 to 80 percent than the 90 percent of pure SaaS. But it captures five to ten times the revenue per customer, because it is priced against the cost of the labor it replaces, not the price of a tool. As one investor put it, the question stops being "what is this software worth" and becomes "what is the work worth".
The skeptics are right about four things. Here is what they miss.
A thesis that cannot survive its strongest critics is marketing. The full-stack story has four serious objections, and the honest answer to each one is also the clearest map of where the durable value actually sits.
"There is no moat. The wrappers just resell OpenAI."
Practitioners on Hacker News and elsewhere are right that the model layer is commoditizing. But that is the argument for owning the data layer, not against it. As one builder put it, your real moat is how good your domain-specific ETL is. The clone can copy your interface overnight. It cannot copy the data and the feedback loop you accumulate by doing the work, the same reason Bing never caught Google in search quality.
"AI is killing consulting. Services are a dying business."
Consulting is not dying, it is repricing, from billing hours to billing for results. And the tell is who the labs depend on: OpenAI runs only around seventy forward-deployed engineers and formed a Frontier Alliance with McKinsey, BCG, Accenture and Capgemini to actually deploy its models. The frontier labs cannot staff the last mile themselves. The firm that owns that last mile owns the business.
"Outcome-based pricing is a trap. The metrics get gamed."
Sometimes, yes. When "resolved" has no shared definition, an operations manager ends up spending hours a week arguing with a vendor about what counts. The answer is not to abandon outcomes, it is to only guarantee outcomes that are standardizable and measurable, and to fall back to per-task or usage pricing where they are not. Honesty about which is which is itself a proof of seriousness.
"It is a margin mirage. The growth is just humans in a trench coat."
Emergence Capital named this exactly: Mirage PMF, revenue powered by human labor dressed up as AI leverage. The test is brutal and correct. If gross margin stays flat and revenue per employee never climbs, the AI is not really doing the work. A real full-stack firm scales non-linearly against its costs. We hold ourselves to that line, not to the press release.
Welzin owns the chain from data to outcome.
We did not arrive at this position to chase a headline. It is simply what a full-stack AI and data firm is. We are not a thin layer over a model we do not control, and we are not a slide deck that leaves once the workshop ends. We build the system and we run it, across four layers, and the one nearest the bottom is the one that compounds.
Data & ground truth
Pipelines, labeling, evaluation and the domain-specific data engineering in front of the model. This is the layer a competitor cannot clone, because they do not have your data or your loop.
Compounds with every job runModels & inference
Selection, fine-tuning, retrieval and orchestration, tuned to the task rather than rented wholesale. The model is an ingredient, not the dish.
Owned, not leasedProduct & workflow
The interface where the work actually happens, embedded so deeply into the customer's process that the workflow itself becomes the switching cost.
Embedded in the workOperation & outcome
Running the system in production with humans on the exceptions, accountable for the result, and increasingly paid for it. We deliver the outcome, not access to a dashboard.
Paid for resultsThe next great AI company will not look like a software company. It will look like a firm that does the work, and quietly runs on AI.
Notes & sources
- Y Combinator, Requests for Startups - "AI-native companies that do not sell software, they sell the service."
- Sequoia Capital, Services Are the New Software - copilots vs autopilots; the 6:1 ratio.
- a16z, AI and the shift to outcome-based pricing; The Palantirization of Everything.
- Foundation Capital, The $4.6T services-as-software opportunity - "integration is the product surface."
- Emergence Capital, The AI-Native Services Playbook - 50%+ margins, the "Mirage PMF" warning.
- Newfund, Full-stack AI in service industries; omnius, the agency business model; and the LinkedIn read on YC's full-stack AI companies.
- Funding and traction: TechCrunch (Harvey, Sierra, Abridge, EvenUp), Fortune on EvenUp, Mercor's growth.
- The counterargument: Hacker News on GPT-wrapper defensibility, Capgemini on consulting, Siena on outcome-pricing, NFX on AI defensibility.
Company names, valuations, funding rounds and figures are drawn from the public reporting linked above and reflect what was reported at the time of writing. They are cited as evidence of an industry shift and do not imply any affiliation with, or endorsement by, Welzin. Market-size and margin figures are estimates from the cited investors and analysts.
