AI-Native Software Grows at 34% and Earns Minus 12%
KeyBanc Capital Markets has released its seventeenth annual AI x Software Benchmark, formerly the SaaS Survey, expanded this year to measure how AI is reshaping product strategy, monetisation, unit economics and operating models. The central finding is a split between growth and economics. Companies selling AI as their primary product reported median ARR growth of 34 per cent in 2025, against 22 per cent for companies embedding AI in broader products and 9 per cent for those still developing customer-facing AI. Those same core AI companies reported a median EBITDA margin of negative 12 per cent, against negative 1 per cent for embedded AI. ARR per full-time employee was $201,000 at embedded AI companies and $117,000 at core AI companies. The release gives no sample size, no respondent count and no definition of either category.
The trade is roughly one for one
The two headline figures are worth putting side by side because the symmetry is the finding. Core AI companies grow twelve points faster and earn eleven points worse. That is not a coincidence in search of an explanation; it describes an exchange rate.
A software company can buy growth by spending on inference, and for AI-native software the benchmark suggests the market has settled at roughly a point of margin for a point of growth. That is a rational trade for a company raising capital and a poor one for a company that needs to fund itself, and the difference between those two positions explains a great deal of current behaviour in the sector.
The productivity figure sharpens it. Embedded AI companies generate $201,000 of ARR per employee and core AI companies generate $117,000, which means the AI-native model produces 58 per cent of the revenue per head. A company growing faster with less revenue per person and a worse margin is a company buying market position rather than building an efficient business, and the benchmark does not say how long AI-native software can continue that trade.
The gross margin gap is the mechanism
The most illuminating number in the release is not the EBITDA figure. It is the gross margin split: companies spending more than 5 per cent of revenue on AI reported a median gross margin of 66 per cent in 2025, against 78 per cent for companies spending less than 2 per cent.
That twelve-point gap sits in cost of goods, not in operating expense, and the distinction is the whole argument. Traditional software has near-zero marginal cost: serving one more user costs almost nothing, which is why software gross margins run in the high seventies and eighties and why the industry has commanded those multiples for two decades. AI-native software has a real marginal cost per request, paid in tokens, and that cost lands above the gross margin line.
This is the structural change the benchmark documents for AI-native software, and it is larger than a cycle. A business model whose marginal cost approaches zero can absorb almost any amount of growth profitably. A model with a per-request cost has to earn a spread on every interaction, and if the pricing does not cover the inference, growth makes the loss larger rather than smaller. Sixty-six per cent is a respectable gross margin in most industries and it is a warning sign in software.
The token bill nobody budgeted for
The benchmark gives direct evidence that inference costs have outrun the assumptions built into pricing. Forty-four per cent of respondents report AI spend over budget year to date, and roughly a quarter are over by more than 20 per cent. Higher-than-expected token usage was the most cited driver of the variance.
That phrase is doing a lot of work. Token usage running above expectation means either that customers use the product more than the model assumed, which is a good problem in a business with no marginal cost and a difficult one in a business with a marginal cost, or that the pricing was set before anyone knew what a typical workload would consume. Both are plausible and both point the same way: the industry priced the unit economics of AI products in an environment of uncertainty, and the bill has arrived.
Scott Peterson, managing director of software investment banking at KeyBanc, describes AI spend as running ahead of budget with the economics still catching up, and calls 2026 a proving year for AI-driven revenue. That is a fair reading of the data and it is a more careful statement than the growth figures alone would support.
The productivity gain is being spent, not saved
One finding cuts against the expectation that AI adoption would show up as cost reduction, and it is the most human part of the dataset. Ninety-two per cent of respondents report that AI is shortening product development cycles and 91 per cent deploy AI internally in engineering, yet 65 per cent report no AI-related headcount reduction or reallocation in 2025.
Peterson’s explanation is that companies are reinvesting the productivity gain to ship more product faster rather than letting it reach the bottom line. That is a rational choice in a market that rewards growth, and it explains why margins have not improved despite a widely reported productivity gain. The gain is real and it is going into output.
The forward-looking figure is where the change sits. Looking out 12 to 24 months, the proportion expecting no AI-related headcount effect falls from 65 per cent to 34 per cent, with research and development at 45 per cent and customer service at 37 per cent the functions most exposed. Two-thirds of respondents therefore expect AI to touch headcount within two years, and the functions named are the two where AI capability is most developed.
Customers are building with AI faster than vendors
Two figures, read together, describe a competitive risk that the growth numbers do not capture. Fifty-seven per cent of respondents say their customers are building alternative products with AI coding tools, and only 20 per cent offer fully autonomous agents to their customers.
The first figure means more than half of these software companies are watching their own customers use AI to build something that competes with what they sell. The second means four in five have not yet shipped the agentic capability that would make their product harder to replace. A supplier whose customers can build around it, while it has not yet automated the part of the workflow that would justify its price, sits in an exposed position regardless of its ARR growth rate.
The benchmark also finds that companies charging separately for AI products reported 27 per cent median ARR growth against 17 per cent for those that do not, and that the growth advantage has not yet translated into higher net retention. Charging separately wins new business and does not yet keep it, which is consistent with products whose value is not yet proven enough for a customer to stop looking for alternatives.
The revenue is growing fast off a small base
One set of figures keeps the whole picture in proportion. AI-driven ARR is projected to grow 128 per cent in 2026, up from 74 per cent in 2025, and will still represent only 12 per cent of total ARR at that rate.
Overall median ARR growth was 24 per cent in 2025, projected at 23 per cent in 2026, with participants planning for a broad-based acceleration to 40 per cent in 2027. Participants expect the overall median EBITDA margin to improve from negative 3 per cent to positive 1 per cent over the same period.
A 128 per cent growth rate on a base that is one-eighth of revenue describes a segment that is scaling quickly and has not yet become the business. The plan for 40 per cent overall growth in 2027 is a substantial step up from a projected 23 per cent, and the benchmark reports it as an intention rather than a forecast it endorses. Whether the margin turns positive on schedule is the test the 2026 edition will settle.
What the benchmark does not say
Three things are missing, and the first limits how far a reader should push any of the figures.
KeyBanc does not disclose the sample size. The release gives no respondent count, no revenue range for participating companies, no geography and no statement of how the survey was fielded. For a benchmark in its seventeenth year that is a notable omission, because the credibility of a median depends entirely on the population it is the median of.
The category definitions are absent. Core AI product, embedded AI and still developing customer-facing AI are the three groups the entire analysis turns on, and the release does not say how a company lands in one. A company that sells an AI feature as an add-on could plausibly fall into two of them.
And the commercial interest deserves stating. KeyBanc Capital Markets is an investment bank whose clients include software companies and private equity firms, and a benchmark that establishes AI-native companies as the growth story is a useful document for a bank advising on that sector. That does not make the figures wrong, and it is the reason a reader should know who produced them.

What to watch
Whether the gross margin gap narrows in the 2027 edition, because that is the single number that would show AI-native companies have found pricing that covers inference rather than one that subsidises it. Whether the EBITDA margin for core AI companies improves from negative 12 per cent as projected, since a second year of losses at that level would suggest the trade is structural rather than transitional. And, whether the headcount expectations convert into actual reductions over the next two years, because the survey records an intention that has not yet appeared in the data. And whether the proportion offering autonomous agents rises from 20 per cent, because that is the figure that determines whether these companies can hold customers who are learning to build without them.

