AI-native software companies are often measured by the same metrics as traditional ones, but that is exactly the problem. These modern companies may be built on models they rent, at margins that may not widen, with customers who have less reason to stay.
This month, in collaboration with Commonplace, we look at what that means for the investors pricing them.
Some of the most impressive companies in venture today are built on top of artificial intelligence models they do not own. They grow fast, sign real customers and raise at prices set years ahead of their traction. Whether they are good businesses is not the question; many are. The key question is which of them will deliver the return you are after. Two of the numbers traditional software investing leans on, recurring revenue and gross margin, answer that less well than they used to and the thing that would answer it best, whether the company can defend its position, was never really a number at all. Read the data the old way and you can overpay for a company with little underneath, or pass on one of the few that would have delivered it.
Early-stage numbers have always been full of assumptions and investors know when to adjust for that. For AI-native software companies, the adjustment may need to be much larger, for two reasons that compound. The first is speed. A company can reach a large revenue figure in a year or two, where traditional software companies would take five or six. So its valuation is set before there is real history to judge it, before a cohort of customers has been through a renewal, before a product has survived the next foundation model release. It can be that while new customers pile in, the ones dropping away barely show, so heavy churn can hide inside healthy growth. The second is that the company does not own what it is built on. The profit margin that should widen with scale stays thin, because the model is rented from one of a few frontier suppliers and the cost climbs with every additional query rather than staying fixed. Rivals often rent the same model, so brand becomes ever more important. Revenue that should harden into habit over time instead churns, because little holds a customer who can buy the same thing elsewhere. The numbers still describe something. But the one that sets the price, annual recurring revenue, is the least reliable of them, and the ones that would show whether the business lasts are either overlooked or not yet there.
This is not the argument everyone is having about AI. One of the most prominent is whether the megacaps can earn back what they spend: Sequoia's David Cahn puts this year's AI infrastructure bill at $1.5tn, against the $3tn of revenue the industry would need to justify it. This is a narrower one, about the early-stage companies that fill most venture portfolios. It is about how to evaluate these companies when the economics underneath them work differently.
Revenue that looks recurring but isn't
Companies built on rented models keep far less of their revenue than ordinary software does. ChartMogul's 2025 study of ~3,500 software companies found median net revenue retention of 48% and gross revenue retention of 40% among the ~200 AI-native firms, versus a median NRR of 82% for B2B SaaS. In plain terms, a year on, a typical AI-native company is still collecting about $48 of every $100 it was earning, against $82 at a comparable SaaS business.
That average hides a wide spread, and price is the clearest divider. AI-native companies charging more than $250 a month held 70% gross and 85% net retention, at or above the B2B SaaS median. Those charging under $50 kept 23% gross. Most of the churn sits at the consumer end, where a subscription can be cancelled in a click, while enterprise revenue behaves much more like traditional software. It carries a different risk instead: a good deal of it is pilot and trial spend that never converts once the trial ends. Jamin Ball of Altimeter has a name for it: experimental run-rate revenue.
Spotting it means reading the deal terms rather than the headline: how long the contract actually runs, whether the customer can leave without penalty, and whether the money comes from a discretionary experimentation budget or a committed operating line. The practical risk is concentration, since a company with a handful of large pilot accounts can post impressive run-rate growth that collapses when two or three of them do not renew. Renewal risk is high at the moment partly because the results are not yet there. MIT's Project NANDA found in 2025 that 95% of organisations were seeing no measurable profit-and-loss return from their generative-AI pilots, and a tool that does not pay for itself does not get bought again. So net and gross retention, and how much of the current book has been through a renewal, are critical metrics to review.
Margins, and why the Rule of 40 requires a different perspective
Traditional SaaS typically carries gross margins of around 70% to 80% because serving one more customer costs almost nothing: the software is already built, so an extra login barely moves the cost base. AI-native companies do not always benefit from that. Every extra query often calls the foundation model provider again, and that inference cost sits inside cost of goods sold, rising with usage instead of staying fixed. Bessemer's State of AI 2025 report put a number on the gap: its fastest-scaling "Supernovas" ran gross margins of around 25%, while the more capital-efficient "Shooting Stars" managed closer to 60%, still well below the SaaS norm. Even those figures are not fixed, since the price these companies pay for model access is set by a small number of suppliers and can move before the next funding round.
That compression flows straight through to traditional investor metrics. The Rule of 40 adds growth rate to profit margin, usually EBITDA or free cash flow. But gross margin sets the ceiling for that profit margin: sales, engineering and support all get paid out of it, so whatever is left over is capped by how much gross margin there was to begin with. That does not make a 25% margin bad on its own; it makes it incomparable to a 2022 SaaS benchmark built on 80% margins. An investor comparing the two directly, without adjusting for that gap, is scoring two different kinds of business against the same standard.
Moats you cannot see
Bain & Company now rebuilds a takeover target's product before buying it. Its consultants vibe-code a rough replica to test whether the code is really the defensible part of the business. For a growing number of AI-native software companies, it is not. Treat every fast-growing AI company as a thin wrapper and you might overpay for none of them, but also could pass on the few that actually pay off. Perplexity was dismissed that way so often that its founder was publicly rejecting the label in early 2024, when the company was worth around $520m. It is roughly $23bn today. When software can be rebuilt this quickly, the work is to find where the defensibility actually sits, and Bain's test can be borrowed for that: the replica tells you where to stop looking.
What usually lasts is more ordinary: distribution, a trusted brand, a product wired so far into a workflow that leaving hurts. Legora is the clearest example. It reached $100m of ARR in about eighteen months, a mark the average Cloud 100 company takes seven and a half years to hit. The product is good: lawyers review hundreds of documents in a single grid and draft in Word against the firm's own playbooks. But the growth came from how it sold. Legora put its own legal engineers inside client firms to build workflows practice by practice, and firms like Cleary Gottlieb co-designed the product rather than buying it off the shelf. A rival can copy the software. Copying eighteen months of being wired into how those firms work is harder.
The underlying questions haven't changed
For AI-native software companies the familiar numbers obviously still matter. What has changed is what they represent to an investor. A gross margin may now carry the cost of every query to a model. A retention rate means one thing at enterprise prices and another at consumer ones. A multiple drawn from a thin peer set may not describe the company at all. The work is to read each number in that context, and to be clear about what none of them shows: whether the team can build the advantage the technology may no longer supply.
The questions are the same whether you are pricing a round or reviewing a mark: how much of a company's revenue has been through a renewal, what the gross margin looks like once inference is counted, what the mark would be against AI-native peers rather than a software comparable from 2022. Underneath them sits the question investors have always asked. Given the risk and the likely exit, is this cheap, fair, or expensive.
How Allermuir approaches it
Hebrides is the software we built for private-market investors. For valuations, its job is to do the science so the investor can supply the art.
Hebrides starts by classifying the company, which sets the methods and the comparable unit. An AI-native software company is not put through the template built for ordinary software, where the margin and multiple assumptions mislead. It flags revenue that may not last, a run-rate flattered by pilots or a book resting on one large account. The technology is valued from what it would cost to rebuild, setting a floor above which any premium has to be earned. Comparables are scored on how alike each peer really is, so the closest fits stay in and the weak ones drop out, and which peers belong stays a judgement the investor can change. The valuation comes out as a range with a confidence score beside it, both shown rather than buried in a single number. This method produces better due diligence, faster.
Commonplace is the private, curated community for LPs, built around the idea that the best conversations in private markets happen off the record. We're grateful to Jocke and the team for the collaboration.