A mid-market vertical SaaS product — practice management, field service dispatch, restaurant back-office — commonly lists somewhere between $100 and $300 per seat per month. That number is not arbitrary and it is not greed. It is arithmetic, and the arithmetic was correct when it was done.
Work backwards from it. Take a product at $200 per seat with 2,000 seats: roughly $4.8M a year. In a healthy pre-2023 SaaS business, something like 20–25% of revenue goes to R&D, which funds eight to ten engineers once you include benefits, tooling, and management overhead. Another 40–50% goes to sales and marketing, because a $200 seat does not sell itself — it needs demos, an SDR pipeline, and a quarterly close. Support, compliance, finance, and leadership absorb most of what remains. The price is close to the smallest number that keeps that organization solvent.
These ratios are industry rules of thumb, not measurements of any specific company, and any individual business will land somewhere different. But the shape holds well enough to make the point: the price encodes the cost of the organization required to build and sell the product. It does not encode the cost of running the software, which for most of these products is a rounding error. It does not encode the value delivered, which is whatever the market will bear above the price.
What actually changed
The honest version of the AI argument is narrower than the one usually made. Models did not make software free to build. They compressed one specific category of work — producing and modifying large volumes of correct, conventional code — from the dominant cost of a product to something closer to a variable input.
That category used to be most of the R&D line. A four-person team in 2019 could not credibly maintain a scheduling engine, a billing system with local tax rules, a mobile client, an admin console, and an integration surface. Not because any one of those is intellectually hard, but because each one is a large volume of careful, conventional work, and careful conventional work was the binding constraint.
It isn't anymore. The binding constraint has moved to judgment: deciding what to build, reviewing what was produced, owning the parts where being wrong is expensive. That work does not compress, but it was never the part that required forty people.
The part that did not change
This is where the argument usually overreaches, so it is worth being blunt about the limits.
Distribution did not get cheaper. A better product at a quarter of the price still has to be found, trusted, and switched to, and switching costs in vertical software are brutal — data migration, retraining, the operational risk of moving a business onto something new. Compliance did not get cheaper. Neither did the decade of continuous operation a customer is implicitly buying, nor the support obligation that comes with real users in production. If anything, the collapse in build cost makes these harder, because they are now a larger share of what separates a product from a demo.
So no, you cannot rebuild Salesforce for $20 a seat. The correct claim is much more specific: for a well-scoped vertical product, the overhead that justified the price is no longer structurally necessary. A small team that owns its own architecture can operate at a cost base that makes a $40 seat sustainable where a $200 seat used to be the floor — and can do it without the growth obligations that would force the price back up.
Why the price is sticky anyway
Incumbents are not going to follow the cost curve down. Not because they are foolish, but because the ones who could are the ones least able to: a public company cannot halve its revenue to match a cost structure it does not have, and a venture-backed company at a $4.8M run rate has commitments that assume the old arithmetic. Repricing downward means firing the organization the price exists to fund. Almost nobody does that voluntarily.
What happens instead is slower and less dramatic. The old price holds for existing customers, who mostly stay, while new entrants take the segments that were always priced out — the solo practitioner, the two-location restaurant, the clinic that could never justify $200 a seat and has been running on spreadsheets. That market is larger than the one being defended, and it is not being defended at all.
That is the opportunity, and it is worth being precise about what kind of opportunity it is. It is not "AI lets you build a competitor cheaply." It is that a large category of businesses was priced out of decent software by a cost structure that was real, was rational, and no longer has to exist. Building for them used to be economically irrational. It isn't now.
At MTM we build and operate our own products, so this is the arithmetic we run on rather than a thesis we sell. It is also why we take the second half of this post seriously: a cost base that makes a $40 seat work only matters if the software underneath it is worth trusting for ten years. That is a harder problem than building it, and it's the subject of the next post.