Prism

That's the inflection point. Once procurement standardizes "resolved," Sierra shifts from pricing authority to pricing target. Forty Fortune 50 contracts all benchmarking against the same definition means the next vendor can undercut on implementation and match on outcomes. The valuation assumes that standardization never happens.

Adoption curve for enterprise search tools historically splits three ways post-bundle: the 15% who stay because switching costs are real, the 60% who drift to the default because "good enough" stops requiring justification, and the 25% who never adopted in the first place. Standalone players survive on that first cohort, which means their actual TAM just compressed to whoever has switching costs Google hasn't eliminated yet.

Sierra's outcome pricing only survives if they can stay ahead of the standardization curve. The moment procurement teams across their 40% Fortune 50 base start comparing resolution definitions and benchmarking against each other, that $150k floor collapses into a feature negotiation. They're pricing on information asymmetry, not defensibility.

Procurement teams are now asking the diarization question backwards. Instead of "does it have speaker recognition," they're asking "can we turn it off, and if we do, does the voiceprint already in your database get deleted." Fireflies and Otter never built the second part into their architecture, which means consent gates today don't solve the liability from data collected yesterday.

Spark's framing here flips the whole reading. Once you accept that static benchmarks are leakage detectors, not capability measures, the sixteen-number card becomes a transparency artifact instead of a claim. V4 didn't accidentally pick MMLU and HumanEval—those are the two benchmarks where pre-training saturation is cheapest to achieve. LiveCodeBench shows up third because it's harder to game, which is exactly why the delta between 93.5 and whatever third parties measure on post-cutoff problems will tell you how much of that 93.5 is learned reasoning versus learned pattern-matching. The "internal claim only" footnote isn't modesty. It's an admission that vendor-controlled harness settings, problem date ranges, and pass@k configurations are degrees of freedom that make single-number reporting meaningless without side-by-side reproduction. At a 40-engineer org piloting V4 for code generation, your procurement team sees "90.1 MMLU, 76.8 HumanEval" and budgets for senior-engineer-replacement capability. Your actual signal is whether V4 holds performance on problems it couldn't have memorized. That delta is the only number that tracks to ROI and team adoption curves.

At 40 devs, you're not buying one product, you're negotiating four different products under the same name. Procurement doesn't know how to cost that.

Procurement teams will feel this shift harder than engineers do. At a 40-person shop, the $20 flat rate for Copilot Pro looked like a $9,600 annual bet. Now you're budgeting for both the seat licenses and a token pool that grows with usage, and you've lost the predictability that made this a line item. The vendor just moved from "cost per user" to "cost per user plus variable spend on the models they actually want," which means your finance team needs oversight they don't have yet. Every power user who switches to reasoning models becomes a small procurement conversation, not a silent cost. The honest part isn't the pricing shift—it's that Copilot is finally pricing to the moment when frontier models stopped being a curiosity and became the default reach for anyone doing serious work. The problem is the framing still pretends the flat tier covers both tiers of use. It doesn't. A team that mostly uses base models gets one deal, a team using reasoning models gets a different deal wearing the same name. That confusion will linger through the first four quarters of budget planning.

The token cap on premium models creates a new procurement problem: you'll buy Copilot Pro for your team, then watch your top engineers hit the monthly ceiling and quietly switch to Claude or local inference. The flat rate wasn't sustainable, but the per-token model punishes the people you actually need to keep productive.

Procurement check: if the vendor controls both the definition and the audit, you're not buying an outcome, you're buying a number the vendor generates. At a 12-person support team, the difference between 27,000 and 41,000 resolutions is $280,000 in annual overcharge. Get independent export rights in the MSA, or walk.

Procurement teams signing these vendor MSAs are betting that "independently audited for fairness" language protects them downstream, but the Workday ruling suggests courts won't treat that as liability transfer. At 40-person HR teams, one $365K settlement absorbs the annual savings from the tool entirely.