Spellbook CEO: Early AI Companies Use Fake ARR to Create Illusory Growth
Scott Stevenson, co-founder and CEO of contract AI product Spellbook, pointed out in interviews and social media statements that an increasing number of enterprise AI and SaaS startups are "beautifying" their Annual Recurring Revenue (ARR) data. "Some count three-month pilot projects as ARR, and they are even free pilots," he said. Investors have reported to him that many companies emerging from accelerators claim to have "1 million dollars in ARR," but upon inspection, it turns out to be all unconverted pilot projects. He specifically mentioned the concept of CARR (Contracted ARR), stating, "This is a very easily manipulated metric." Many companies include contracts that have not yet gone live for billing, or even those tied to future development commitments, in their ARR, severely exaggerating the true scale of their business.
Stevenson cited common practices, including: signing complex implementation projects with major clients that state "develop features first, start billing after delivery," yet counting the entire contract commitment amount as ARR in advance; using "tiered pricing contracts" to count the high price of the third year as current ARR, rather than calculating based on actual payment progress or averages; and even stacking numerous unfulfilled contracts with 12-month exit clauses as so-called "three-year contracts," inflating the reported ARR to 2-3 times the actual recurring revenue on the books. He emphasized that Spellbook only counts "invoiced, effective, net discounted recurring revenue," excluding all pilot projects, free usage, and unbilled commitment revenue from ARR, calling for the industry to return to a more conservative and comparable definition.
Source: Public Information
ABAB AI Insight
Stevenson's description is not just a technical debate over a single metric, but a familiar cyclical "accounting trick" rehash: each wave of technology and capital frenzy invents a seemingly reasonable, yet actually vague, metric that blurs the boundaries between future and present—last cycle it was "community EBITDA," this time it's "Contracted ARR." Essentially, CARR packages commitments that are "signed but not yet fulfilled and billed" (including parts that clients can exit at any time) into current "recurring revenue," compressing the time dimension into a current figure, making it convenient to tell a "explosive growth" story in financing, mergers, and media narratives.
In the enterprise AI scenario, this deviation is more easily magnified: complex projects often go through multiple stages of prior POC, integration development, change management, and internal promotion, with real cash flow and usage intensity lagging behind the signed contracts by months or even longer. If this entire uncertain future path—including undelivered features and unverified usage scenarios—is counted as "today's ARR," early companies may appear as mature SaaS that have crossed the risk threshold, while investors, secondary markets, and employees receive an idealized series that is "undiscounted and unaccounted for failure rates."
Stevenson's criticism of "counting the third-year amount in ARR for tiered pricing contracts" reveals another structural issue: in an environment of expected long-term high interest rates and slowing growth, management is incentivized to raise future expectations through one-time boosts to achieve higher valuations or larger financing amounts, even if it means exchanging accounting illusions for future execution pressures. When the first-year customer only pays one-third of the price and has a 12-month exit right, the company is economically facing a fragile one-year contract, but narratively it is packaged as a large three-year confirmed order—this mismatch, once systematically used, will lead to an overestimation of the revenue quality across the entire sector.
From a longer financial and historical perspective, after each round of bubble bursts, the market tends to reactively correct the "metric innovations" of the previous cycle: after the internet bubble, there was a collective reckoning of "pro forma profits," and after the sharing economy, there was a discount on "adjusted EBITDA." This round of enterprise AI cycles is likely to treat "Contracted ARR" and various "adjusted ARR" in the same way. Stevenson’s public exposure of these practices at this time is, to some extent, a forewarning: when real renewal rates, feature delivery rates, and project suspension ratios begin to surface, those companies relying on "stacking CARR to tell stories" will be forced to compress their inflated revenues back to reality within one or two financial reporting cycles, at which point the repricing of valuations and credit will no longer be a localized phenomenon, but a collective check-up for the entire enterprise AI sector.