There's a quiet accounting scandal running through the AI startup ecosystem, and it's been hiding in plain sight. Founders are announcing eye-popping revenue milestones, VCs are amplifying those numbers to journalists, and somewhere between the press release and reality, the definition of "revenue" quietly shapeshifted into something substantially less impressive.

Scott Stevenson, the CEO of legal AI company Spellbook, called it publicly last month. His post on X didn't mince words—he called it a scam, pointed at the biggest funds in venture capital as knowing participants, and accused the whole operation of misleading the press. The response was visceral: hundreds of reshares, commentary from prominent investors, and enough founder pile-ons to suggest he'd hit a nerve that a lot of people had been quietly nursing.

So what's actually going on? Let's pull back the curtain on the metric manipulation that's inflating AI valuations and obscuring who's actually building a real business.

ARR: A Metric That Used to Mean Something

Annual Recurring Revenue—ARR—earned its credibility during the SaaS era. The concept was straightforward: add up the annualized value of all your active, paying customers under contract. Not future customers. Not signed-but-not-yet-started customers. Paying. Active. Now.

It was a useful shorthand precisely because it captured momentum without the noise of one-time sales. Accountants don't formally audit it under GAAP (which only cares about revenue you've already collected), but the definition was understood well enough that the metric carried real informational weight. When a company said they hit $50M ARR, you knew what that meant.

Past tense. Because now? The definition has been quietly stretched until it barely resembles the original.

Enter CARR: The Squishier Cousin

The main sleight of hand is replacing ARR with CARR—Contracted or Committed ARR—and then just... not mentioning the swap.

CARR counts revenue from customers who have signed contracts but haven't actually started using the product yet. They're not onboarded. They haven't paid. The product might not even be fully deployed. But their contract value gets added to the revenue figure that ends up in press releases and pitch decks.

Here's where this gets materially problematic: enterprise AI implementations are notoriously messy. Integration timelines slip. IT departments drag their feet. The proof-of-concept that looked clean in the demo hits organizational friction the moment real data and real workflows get involved. Clients who signed in Q3 might not go live until Q2 of the following year—if they go live at all.

One VC told reporters that they've seen companies where CARR runs 70% higher than actual ARR, with a meaningful chunk of that contracted revenue unlikely to ever materialize.

Bessemer Venture Partners laid out the right way to use CARR back in 2021: you're supposed to discount it for expected churn and "downsell" (customers who eventually buy less than they contracted for). Done honestly, CARR can be a useful leading indicator of pipeline health. Done dishonestly—or just done sloppily—it becomes a number that exists primarily to impress people who won't dig into the methodology.

The Specific Ways This Gets Gamed

Let's be concrete about the mechanics, because "revenue inflation" sounds vague and the reality is quite specific:

• Free pilots counted as ARR: At least one startup reportedly booked a year-long free pilot as part of its ARR figure. No money changed hands. The customer was still evaluating. The board—including representation from a major VC fund—knew this was happening and knew the customer could walk away before paying a dollar.
• Multi-year contracts front-loaded: A startup can offer steep discounts in years one and two of a three-year contract, then count the undiscounted year-three value in their ARR announcement. The headline number looks huge; the actual near-term cash flow is a fraction of it.
• No churn adjustment: Startups reporting CARR as ARR frequently skip the step where you discount for expected customer losses. Which means every signed contract, regardless of likelihood of completion, gets counted at full value.
• "$100M ARR" with minority paying customers: Multiple investors confirmed awareness of at least one high-profile enterprise AI startup that crossed the $100M ARR announcement threshold with only a fraction of that figure coming from customers who were actually deployed and actively paying.

Why VCs Go Along With It

This is the part that makes the whole thing more troubling than ordinary startup optimism. Founders stretch metrics—that's practically in the job description. But investors are supposed to be the adults in the room.

The competitive dynamics explain a lot of it. When one startup in a category announces $50M ARR (really CARR), their competitors face an uncomfortable choice: report honestly and look like they're losing, or adopt the same methodology and keep pace. The race to the top of the revenue leaderboard creates pressure to match whatever accounting creativity the category leader is deploying.

Investors also have asymmetric incentives. A VC whose portfolio company announces a milestone that drives up the valuation—even if the underlying metric is mushy—benefits from the markup on their fund's performance. The cost of the inflation is borne later, often by the next round of investors or by public market shareholders if the company eventually lists.

And frankly, some of the largest funds are sophisticated enough to know exactly what they're looking at. When a fund is told their portfolio company hit $100M ARR and the fund's partner knows that figure includes undeployed contracts and optimistic churn assumptions, the decision to amplify that number in press coverage is a choice, not a mistake.

What Actually Matters (And What to Ask)

If you're building with AI vendors, evaluating competitors, or deciding where to deploy capital, here's what the ARR number in a press release doesn't tell you:

• Net Revenue Retention (NRR): Are existing customers expanding or contracting? NRR above 120% is a genuine signal of product-market fit. NRR below 100% means the company is losing ground with its own customers even as it signs new ones.
• Implementation success rate: For enterprise AI, what percentage of signed contracts successfully complete implementation and go live? This is the number that exposes whether CARR is optimistic or delusional.
• Revenue collected vs. contracted: GAAP revenue—money actually received—is auditable and standardized. If a company is reporting ARR figures dramatically higher than their GAAP revenue, you know the gap is filled with something softer.
• Logo churn: How many customers are leaving? A startup can grow headline ARR while quietly losing customers if the new contracts are larger than the churned ones.

The Longer-Term Problem

The ARR inflation game has a shelf life. Eventually, contracted revenue that doesn't convert to collected revenue shows up as a gap in the financials. Implementation failures become public. Customers who signed during the AI enthusiasm wave of 2024-2025 start demanding ROI evidence that the technology often can't yet provide at enterprise scale.

When that reckoning comes—and for a meaningful subset of these companies, it will—the headline ARR figures will look less like milestones and more like the wishful thinking they always were.

The companies worth watching are the ones reporting NRR, deployment rates, and actual collected revenue alongside their ARR figures. That transparency is uncomfortable when the numbers are modest. It's also the only version of the metric that tells you whether there's a real business underneath the press release.

Everything else is just sophisticated storytelling with a spreadsheet attached.