Annual recurring revenue used to be one of the more honest numbers in software. It described the annualised value of contracts with live, paying customers — income you could bank on repeating for as long as the customer stayed on board. No auditor ever signed it off, but it had a settled meaning, and investors trusted it enough that a whole generation of SaaS companies was priced on the strength of it. That trust is now being spent, because a growing number of AI startups are announcing “ARR” figures that describe something much softer than recurring revenue.

I started my first business in the wreckage of the dot-com crash, and every boom since (social media, big data, now AI) has taught the same dependable lesson: the scoreboard gets redefined right when the game becomes hardest to win honestly. Eyeballs in the dot-com years, gross billings in the daily-deals era, community-adjusted EBITDA at WeWork — each was invented because the real numbers couldn’t carry the story on their own. So when the most trusted metric in software starts being stretched to breaking point by the hottest sector in a decade, that tells me more about the stage this cycle has reached than any forecast could.

Three different numbers, one acronym

Scott Stevenson, co-founder of the legal AI firm Spellbook, went public in April, first on X and then on the tech podcast TBPN, with what he called a “huge scam”: AI startups reporting inflated revenue figures with the support of some of the world’s biggest funds. His specific complaint was the blurring of ARR with “contracted ARR”, a measure that counts signed contracts whose customers aren’t yet live, aren’t yet paying, and in some cases never will be. Stevenson said the confirmed cases he knew of showed gaps of three to five times between the announced number and the money actually flowing. When TechCrunch followed the thread and spoke to more than a dozen founders, investors and startup finance people, they confirmed the practice was a common one: one VC described companies whose contracted figure ran 70 per cent ahead of genuine recurring revenue, a former employee described a year-long free pilot being counted as ARR with the board’s knowledge, and several investors pointed to a high-profile startup celebrating $100 million in “ARR” while only a fraction of it came from paying customers.

The acronym is doing a second job as well. Plenty of AI companies now use ARR to mean “annualised run rate”: take a strong recent month, multiply by twelve, announce the result. For a usage-based product with no long contracts underneath it, that’s not recurring revenue in any meaningful sense — it’s a snapshot with a megaphone. Picture a vendor with $1 million actually being paid by live customers, another $3 million signed but not yet deployed, and one strong month that annualises to $5 million. All three figures can be announced as “ARR”, only the first is money in the bank — and you’re rarely told which one you’re being quoted.

The oldest trick in the boom

None of this is new, which is rather the point. In 2011, Groupon filed for its IPO claiming $713 million of 2010 revenue; by the time the SEC had finished asking questions, that same year’s revenue had been restated to $313 million, because the company had been booking the gross value of its vouchers rather than the slice it kept. In 2018, WeWork wanted to sell bonds while losing $933 million on $866 million of revenue, so it offered investors “community-adjusted EBITDA” — profitability once you excluded most of the costs of running the business, and a phrase the founder of one bond research firm said he’d never seen in his life. The offering raised $702 million anyway. These weren’t the same manoeuvre — Groupon’s was a revenue-recognition failure, WeWork’s an invented profitability measure, and contracted ARR is a forecast wearing a metric’s clothes — but the motive behind them is common: change the scoreboard when the ordinary score stops flattering you.

The pattern always has two authors. Founders stretch the metric because the incentives reward the stretch: in a market that prices an AI startup at a large multiple of whatever revenue figure it announces, a few extra million of claimed ARR converts into tens of millions of valuation. And investors let the stretch stand because their portfolios benefit from it. Stevenson’s blunt assessment was that funds are incentivised to create the narrative that they hold runaway winners, and TechCrunch’s sources agreed, with one admitting that investors can’t call the practice out because everyone has a company doing it. The people best placed to police the number are the people paid to promote it. I’ve written before about how much of the AI era’s public story is performance — executives blaming AI for layoffs they’d planned anyway — and this is the same theatre playing to the investor gallery.

Why this lands on your desk

You may not much care whether venture capitalists mislead one another; they can look after themselves. But these numbers travel, and if you run a business they reach you through more doors than you might think. The most direct is vendor selection: the “fastest-growing AI company in its category” pitch works on buyers just as it works on the press, because momentum looks like proof that a product is safe to choose. If the momentum is partly fictional, some of that safety is fictional too. There’s also the plainer risk of durability, since a vendor whose collected revenue is a third of its announced revenue is far more likely to put its prices up sharply or disappear altogether once its funding environment cools, taking your workflows and integrations with it. And the number reaches your boardroom too. When every AI company in the headlines is apparently tripling revenue, “why haven’t we transformed yet?” becomes a harder question to answer calmly, even when the fair answer is that the comparison was rigged from the start. A strong AI business and a strongly promoted one can look identical in a headline; the difference only shows up in the numbers nobody is required to publish.

Whether this ends in a Groupon-style public reckoning or the sector simply grows into its exaggerations, I can’t say — booms have gone both ways before. What history does say clearly is that when a correction comes, the funds hedge across a portfolio and move on, and the lasting damage lands on the customers and employees who committed to a flattering number.

Questions worth asking before you buy

The practical response costs you perhaps ten minutes per vendor, and it belongs with whoever signs off your software spend: the finance director weighing a contract renewal, the marketing lead being shown a category leaderboard, the founder about to bet a workflow on a startup’s longevity. Every one of these questions can be asked politely and directly:

  1. Ask which ARR you’re being quoted: Is the figure built from live, paying customers, or does it include signed-but-not-deployed contracts and unconverted pilots? Ask for the split in writing: a vendor with clean numbers can produce it, and a salesperson who can’t should be happy to get it from finance.
  2. Ask about the annualisation window: If the number is a run rate, what period was multiplied to produce it? A best-week-times-52 figure and a trailing-twelve-months figure are entirely different animals.
  3. Ask about renewals and real prices: How many pilots have converted to paying contracts, what are the cancellation terms, and do existing customers renew at the price you’re being quoted? A product subsidised below cost by venture funding shows up in those answers long before it shows up in a price rise.
  4. Judge the evidence, not the headline: The same discipline I suggested in the AI pivot test applies to vendors — genuine traction shows up in reference customers, retention and delivered results, none of which can be manufactured with an acronym.

None of this requires cynicism about AI itself — a fair few of today’s startups will earn every pound of their valuations. But a sector’s willingness to redefine revenue tells you something important about the rest of it, and being the buyer who asks what the number actually means costs you nothing beyond a moment of awkwardness in a sales call. It may turn out to be the cheapest due diligence you ever do.