The way companies buy AI technologies is changing rapidly, so moving a product from the experimentation stage to full production no longer necessarily means that a startup will secure stable revenue for years. According to research conducted by venture capital firm Madrona, 74% of the 150 IT professionals at companies surveyed intend to increase their AI budgets over the next 12 months, while the rest plan to keep spending at its current level.
But continued spending does not mean vendor stability. Participants say that fewer than half of experimental AI projects reach full production. Although this percentage appears low, it represents an improvement over a result reported by MIT last year, which found that 95% of enterprise AI projects failed in terms of return on investment.
From Experimentation to Continuous Reassessment
The most significant finding in Madrona’s research is not how many projects succeed in reaching production, but what happens afterward. Some 77% of companies reassess their AI vendors every six months or continuously. Madrona sees this as creating a “get in fast, get out fast” dynamic, fundamentally different from traditional enterprise software, where multiyear contracts gave vendors considerable stability.
In the AI market, the costs of switching vendors appear lower, while the pace of comparison and reassessment has become more intense. As a result, a product can secure an enterprise contract and move beyond the pilot-project stage, yet remain at risk of losing the customer in a later review cycle.
What Is Changing in Practice for Startups?
Funding enterprise experiments was one of the primary drivers of the initial AI boom in 2025. The expectation was that 2026 would become the stage for solidifying these relationships and converting them into long-term commitments. Enterprise contracts typically allow startups to report exceptional growth rates in annual recurring revenue, including cases of going from zero to $10 million in three months.
However, these figures now need to be read more cautiously. Revenue generated by an enterprise contract does not necessarily equal revenue guaranteed over the long term, even after the product has been adopted within the company. This puts additional pressure on startups to continually prove their value rather than simply completing the initial sale.
Outcome-Based Pricing
Part of the problem is tied to the instability of pricing models for AI products. Another study conducted by venture capital firm Andreessen Horowitz, involving 50 technical buyers of AI products, found that more than half preferred tying fees to work completed or outcomes achieved rather than to usage, such as the number of tokens processed.
Usage-based pricing relies on a logic common in the software-as-a-service era: a company knows it needs email, human resources software, or cloud storage, and then determines the amount of usage it will pay for. But AI products are not always measured this way. Tying the price to the number of reports processed, support requests closed, or sales opportunities generated may make the product’s economic value clearer to both parties, according to what Andreessen Horowitz partners Tugce Erten and Sarah Wang wrote.
Why Does This News Matter?
The new purchasing pattern opens the door to more experimentation, because companies appear more willing to test AI tools. At the same time, however, it weakens the relationship between signing a contract and revenue stability. For startups, this means that rapid ARR indicators may conceal a higher level of uncertainty than traditional software contracts did.
The research does not determine when, or whether, companies will return to long-term purchasing habits. The open questions therefore concern AI products’ ability to prove measurable results and startups’ ability to build pricing models that reflect those results. This is an analytical reading based on Madrona and Andreessen Horowitz research as presented by Julie Bort in TechCrunch, and it is not evidence that all enterprise contracts are short-term or that all current ARR models are unreliable.