AI startups' annual revenue has never been so fragile, according to a new study

The article starts from a paradox that is redefining the enterprise artificial intelligence business: never has so much money been spent on technology, and never has that spending offered so few guarantees of continuity for those collecting it. Consulting firm IDC forecasts that companies will devote $4.25 trillion to technology in 2026, spending driven almost entirely by AI. A new report from venture capital firm Madrona, based on a survey of 150 enterprise IT professionals, confirms that upward trend: 74% of respondents plan to expand their AI budget over the next 12 months, and the rest will keep it steady. In short, nobody is planning to cut.

Yet that budgetary enthusiasm coexists with a still very low success rate. According to the same Madrona report, fewer than half of AI pilot projects actually make it into production. The figure sounds disappointing, but it represents a notable improvement: last year, an MIT report that became famous within the sector claimed that 95% of enterprise AI projects had failed in terms of return on investment. Going from a 5% success rate to fewer than half of pilots is, in that context, considerable progress, even if the starting bar was extremely low.

The study's most relevant finding, however, has nothing to do with whether pilots reach production, but with what happens afterwards. Even when a company successfully deploys an AI tool, it does not commit to it over the long term: 77% of companies reassess their AI vendors every six months, or even on an ongoing basis. Madrona describes this as a 'fast in, fast out' dynamic that breaks with the traditional logic of enterprise software (SaaS), where multi-year contracts acted as an inertia barrier protecting vendors' revenue. In the case of enterprise AI, the firm notes, the costs of switching vendors are lower and the pace of reassessment is relentless.

This instability has direct consequences for one of the startup ecosystem's most closely watched metrics: annual recurring revenue (ARR). Large companies' trial budgets were the fuel for AI's initial boom in 2025, and 2026 was supposed to be the year those same customers settled in and signed lasting commitments with AI startups. It is precisely those enterprise contracts that have allowed some companies to boast meteoric revenue growth, such as the phenomenon of startups going from $0 to $10 million in revenue in barely three months. The problem is that, for the first time, that enterprise revenue remains insecure even after the product clears the pilot phase and is formally adopted by the client company.

Part of the problem, according to the article, is that many AI startups have yet to find a suitable way to price their enterprise products. A separate survey, conducted by venture capital firm Andreessen Horowitz (a16z) among 50 technical AI buyers, reveals that more than half of them prefer AI pricing tied to work performed or to concrete outcomes, rather than to consumption, as happens with per-token billing models. Charging by usage, note a16z partners Tugce Erten and Sarah Wang, quoted in the report, is in fact a model inherited from the SaaS era: once a company knows it needs email, HR software or cloud storage, all that is left is to decide how many employees or how much data it has to pay for. With AI, something different happens: pricing 'around recognizable work' — for example, by the number of reports processed, tickets closed or leads generated — is what allows the startup to prove its real value to the client, making the product 'economically valuable for both sides'.

Taken together, these figures suggest that AI has ushered in a new era of enterprise experimentation, with upsides and downsides for startups. On the one hand, large companies are more willing to try new technology, which opens doors for young companies that previously found it harder to break into big corporate accounts. On the other, that same openness means that signing an enterprise contract no longer guarantees stable long-term revenue, as it did in the classic software-as-a-service era. The article closes without resolving the central question: whether companies will at some point return to their traditional buying habits, more committed and longer in duration, or whether this volatility in the vendor-client relationship is here to stay as a structural feature of the enterprise AI market.