The Procurement Scissors Are Closing

The inflated-ARR story TechCrunch ran this week reads like a VC accountability piece. Anonymous founders. Anonymous investors. The familiar choreography of a sector caught counting its own pipeline as revenue. It's a fine read on its own terms.

It's also the less important half of the article that should have been written.

Because the same week, Fortune surfaced internal Microsoft data showing that AI tooling, used heavily, costs more per employee than the employee. Two pieces, two desks, no apparent overlap. Put them next to each other and what you have is a closing scissor on every enterprise AI business case written in the past eighteen months. One blade is the revenue number the vendor reports. The other is the cost number the buyer models. Both are wrong, and they're wrong in the same direction.

I'll call it the procurement scissors. The buyer overpays against a vendor metric that's been padded, then pays again at deployment because the cost model that justified the deal never anticipated how the tool actually gets used. CFOs holding both numbers are holding positions that don't survive a year of production.

The first blade

The TechCrunch reporting describes founders and their investors counting things as ARR that any auditor would refuse. Month-to-month consumption revenue annualized after a single high-usage week. Pilot contracts with explicit termination clauses treated as recurring. Bookings from related-party investors counted twice if you squint. The story sources from anonymous founders and VCs who, even granting the obvious motivation to vent, describe practices specific enough to be checkable. The reporting doesn't claim universality, and I won't either. Some AI companies are doing this. The structural pressure that makes them do it is universal.

Here's the pressure. A Series B in AI right now requires growth metrics that real revenue won't produce on the timeline the round demands. The investor needs a number for the next markup. The founder needs a number for the next round. Both parties know what ARR used to mean, and both parties know the audience for the new definition isn't an auditor. It's the next investor, who has the same incentive to wave the number through.

The enterprise procurement team reading the pitch deck has no current framework for asking which definition of ARR is in use. They're looking at growth curves. The growth curves are the product.

The second blade

The Fortune piece centers on Microsoft's internal cost data, specifically that employees using AI tools heavily are generating per-seat costs above what those employees earn. That framing is attention-grabbing and it needs scoping. This isn't average enterprise deployment, it's high-usage roles inside the company whose AI stack runs at margin. Not every Copilot seat costs more than the salary it sits next to. Yet.

Worth noting: Microsoft has no commercial reason to surface this kind of data. They sell the tools whose costs the data exposes. The signal is leaking out in spite of the incentive, not because of it. That makes it more credible, not less.

The directional point doesn't depend on the framing being literally true for every role. Token-metered tools have a cost curve that pilot programs are structurally unable to discover. Pilots cap usage. Production doesn't. A six-week pilot with twelve users running constrained workflows produces a cost profile that has no relationship to a year-two deployment with eight hundred users running agentic workflows they invented themselves. Anyone who's modeled SaaS-to-consumption transitions knows the shape of this curve. The AI version is steeper.

This is what no buyer modeled. The pilot's per-seat economics looked beautiful because usage was throttled by novelty. The org figured out what the tool was good for. Usage went up by an order of magnitude. The invoice arrived.

A brief detour through 2014

The SaaS era ran this play already, in slower motion. ARR was a fuzzier number in 2014 than the industry pretended, and consumption-based pricing surprised more than one CFO who'd signed for a fixed seat license and discovered halfway through year two that the platform had upsold the engineering team into a usage tier nobody approved. The defenses procurement eventually built — contract language around true-ups, audit rights, usage caps, committed spend ceilings — were specific to SaaS economics and took roughly a decade to standardize.

The AI version of this problem has been running for about two years. The defenses don't exist yet. The contract templates most enterprises are using were written for a software model where marginal cost was near zero. AI inference is not that model.

(If you're a GC reading this and your AI vendor contracts don't have a usage-tier change-of-control clause, you have an interesting conversation ahead of you this quarter.)

After the pilot

Here's where I genuinely don't know how this resolves. The defenses against blade one, definitional scrutiny on vendor ARR, require procurement teams to ask vendors questions that even sophisticated buyers don't currently know how to phrase. The defenses against blade two, usage-cost modeling, require finance teams to model demand curves they have no data on, because the only data that exists comes from the vendor whose contract they're trying to evaluate.

The closest analog is the way enterprise SaaS contracts evolved between roughly 2015 and 2020, when buyers finally got serious about consumption telemetry, true-up windows, and the right to audit. That took five years and a lot of bad renewals. The AI version compresses on both sides. Vendors are growing faster. Costs scale faster. Procurement adaptation is moving at the same speed as last time.

So the next eighteen months produce a predictable sequence. A wave of AI vendor bankruptcies whose ARR turned out to be one large customer paying month-to-month. A wave of enterprise budget overruns that finance teams will retroactively attribute to scope creep but which are actually the pilot-to-production cost cliff. And a smaller, quieter wave of buyers who priced both blades of the procurement scissors into their first contract and are running a useful program at a cost they can defend.

Which group your company ends up in depends on whether the person negotiating your next AI contract has heard the phrase before reading the term sheet, or after.

So here's the question I actually want answered. When the first enterprise AI vendor blows up because its ARR was a mirage and its anchor customer churned the same month, do the buyers who signed multi-year contracts off that ARR have any contractual recourse — or did they buy the number?