The $15.8B company whose entire pitch is "pay only for what works" will not tell you what working costs. The per-seat incumbent everyone wrote off publishes real numbers, right up until AI enters the bundle, and then goes dark too. Underneath both sit a filing that says the quiet part out loud, a billing definition almost nobody has read, and a contract Five9 rebuilt before the threat it was built for arrived.

Somewhere in the risk factors of Five9's most recent annual report, filed under the signature of executives who go to prison if it's knowingly wrong, is a sentence describing the company's own product killing its own revenue.

Not "could." Five9 wrote "will likely," which makes displacement the base case rather than the tail risk, and it wrote it for the SEC.

Ten weeks later, on the Q1 earnings call, Five9's CFO told analysts the seat count "continues to grow at a healthy rate." One quarter after that, asked directly about compression, he said: "We really have not seen that seat compression, nor have our customers."

The obvious story writes itself. Legacy vendor sells AI that eats its own meter, tells the street everything is fine, gets disrupted by a $15.8B challenger charging per resolution instead of per human. Bret Taylor, who ran Salesforce and Facebook engineering and now chairs OpenAI's board, has been saying exactly that on every podcast that will have him.

I spent a week in both companies' filings, pricing pages, DNS records and job postings, and the obvious story is wrong in two directions at once.

Five9 is not lying, because Five9 does not bill per human being. And Sierra, the company that made "you only pay when it works" its entire identity, has never once published what working costs. No rate has ever appeared on its site, and no range or calculator either. I pulled its sitemap: 680 URLs, zero pricing pages. The only pricing content Sierra publishes is essays about pricing.

So this is a teardown about two companies with opposite pricing philosophies and one shared instinct, which is that the price of AI is something you discuss on a call.

The number I keep staring at is the profitability row. Five9 earned $39.4 million in net income in 2025, up from an $81.8 million loss two years earlier. It generates roughly $175 million in free cash flow, holds $724 million in cash, and is buying back its own stock in size: a $90 million accelerated repurchase plus a fresh $200 million authorization. Sierra, priced at six times Five9's market cap, discloses neither margin nor retention.

One of these companies is a going concern with an AI narrative problem. The other is an AI narrative with an unproven cost structure. Neither of those is the same as "disrupted incumbent."

Six dimensions, each graded on cited evidence, composite derived last. This grades the go-to-market motion, not the product.

Composite: Sierra B+, Five9 C+.

Sierra takes four of six rows. Five9 wins one outright, and closes to a single band on the row everyone assumes is a blowout. Sierra also collects the first sub-C grade an index challenger has ever received, on the dimension its own marketing is about. I'll walk the three rows that break the pattern.

Sierra's motion is genuinely good, and the good parts are not the ones that get written about.

Start with the pricing mechanism, because it is the real thing. Taylor's own description, from Stripe's Cheeky Pint podcast:

The line that matters more, from the same conversation, is the one about who carries the model bill:

That is a real structural commitment and it is the correct answer to the question this whole category is organized around. When AI eats the unit your revenue is priced on, you want to be priced on the AI's output, not on the humans it replaced. Sierra is the only company in this index whose revenue mechanism gets stronger as its product gets better.

The demand engine underneath it is more built out than its reputation suggests. Sierra University has run for a year and has graduated, in the company's own words, "hundreds of people from dozens of companies." Sierra Summit ran its first customer conference in November 2025. A Ghostwriter hackathon drew "30 people from 17 world-class companies, more than half with over $10B in annual revenue." And Sierra publishes its own agent-evaluation benchmarks, tau-bench and its successors, including a voice benchmark covering 278 grounded customer-service tasks. Authoring the measuring stick your category gets judged by is about as good as demand creation gets.

Activation is fast for enterprise software. Vivid Seats went live in four weeks, Nordstrom's voice agent in five, Cigna in eight, Singtel in under ten. WeightWatchers "contained nearly 70% of all cases within the first week." Sierra staffs this with Forward Deployed Infrastructure Engineers who "own the end-to-end lifecycle of customer deployments," and it staffs agent engineers by the customer's language rather than by Sierra's headquarters, which is why there are Cantonese, Thai and Korean-speaking agent-engineer postings in Singapore.

Then there is the part of Sierra's own marketing I did not expect to find, which is Sierra describing Five9's predicament more precisely than Five9 ever has:

That is a competitor's marketing page stating, in plainer English, the risk Five9 filed with the SEC. Sierra's own comparison table on that page grades traditional seat-based pricing as carrying a High potential for wasted spend, consumption as Medium, and outcome-based as Low.

And here is where the motion contradicts itself. Sierra's revenue organization is a conventional enterprise sales machine. Its LinkedIn department split runs Technical 156, Sales 127. It hires Enterprise Sales Directors segmented by Majors, Strategic, Financial Services, Healthcare and Retail, plus Enterprise Sales Engineers verticalized the same way, plus a Marketing Operations lead whose posting asks for "12+ years... attribution modeling, multi-touch funnel analysis... Marketo or HubSpot, Salesforce, lead routing." Sierra runs Salesforce, which is a fine choice and also a funny one for a company founded by Salesforce's former co-CEO.

The one genuine efficiency: there is no SDR layer. Every Sierra AE posting folds "prospecting and lead generation, research, networking, and cold calling" into the rep's own job. That is leaner than the industry default, and it is the only part of Sierra's revenue org that looks structurally different from 2015.

Five9's distribution is the most underrated asset in this teardown.

It has been a Gartner Magic Quadrant Leader for CCaaS across eight-plus consecutive cycles, which in enterprise procurement is not a trophy but a demand channel: it puts Five9 on the shortlist before anyone has spent a dollar on acquisition. It runs 100+ accredited marketplace integrations and a staffed channel organization managing "15+ National & Global partners," with dedicated channel hiring for DACH. AT&T's Cloud Contact Center is built on Five9. Call Tower was one of the partners Microsoft selected for native Teams Contact Center integration. And the $100 million total-contract-value new logo Five9 announced in Q2 2026 was closed through Google Marketplace.

Now the finding that reframes the entire bear case. Buried in Five9's revenue-recognition note:

Five9 does not sell seats. It sells peak concurrency.

A contact center can shed people, real headcount, and shed no licenses at all, because the meter is set by how many agents are logged in at once during the busiest hour. AI has to flatten the peak, not the payroll, before Five9's revenue moves. That is a materially harder thing for AI to do, and it is almost certainly why the CFO can say "we have not seen that seat compression" while the same company's 10-K warns about license revenue. Both statements are true. They are measuring different quantities, and every analyst modeling headcount reduction is modeling the wrong number.

Five9's own pricing page corroborates it in a footnote, which is a strange place to find the load-bearing detail of a business model: "WEM offered on a named-basis. When named seats required exceeds concurrent seats sold, additional sold as add-ons."

The second thing Five9 did is more deliberate. Its CEO, Amit Mathradas, on the Q1 2026 call:

Customers commit dollars for three to five years and flex the mix between human agents and AI agents inside that commitment. They can fire people without paying Five9 less. One Fortune 100 financial services customer is on this structure with a five-year deal ramping to roughly $25 million in ARR. And when I pulled Five9's DNS records, Zuora was sitting in the SPF chain, which is billing infrastructure built specifically for committed-spend and consumption contracts. The plumbing matches the story, which is rarer than it should be.

Five9's AI line was never on the seat meter either. The CFO, on the Q4 call: "our $100 million of enterprise AI revenue is all consumption or capacity based." That line grew 78% year over year in Q2 and now runs at 15% of subscription revenue, up from 9% a year ago.

Now the breaks, and there are three.

The first is that the ratchet is asymmetric. From the same filing: reductions in licenses require "30 days' notice," while increases "can be provisioned almost immediately." Down fast, up faster. Concurrency protects Five9 right up until the day it doesn't, and then it unwinds monthly.

The second is that Five9's AI story is currently being delivered by more humans, not fewer. Adjusted gross margin fell to 61% from 63%, and EBITDA margin to 22% from 24%. The CFO's explanation: "Both metrics were impacted by a temporary expansion of professional services capacity, enabling us to address customer demand to deploy their AI solutions earlier than anticipated." Five9's own filing notes professional services "typically have negative margins." The company raised its AI growth guidance in the same quarter it paid for that growth in headcount.

The third is the internal machine, and it is the weakest thing about Five9. Its mail routes through Microsoft 365 with Mimecast layered on top, the classic legacy security-gateway tell. It runs Salesforce plus Marketo plus Outreach plus 6sense, and three separate transactional email vendors (Mailjet, a dedicated Mailgun subdomain, and SendGrid) for one company. Its sales development function is a textbook BDR and LDR pyramid reporting into an "RVP, Sales Development" who leads "the front line of our well-established Sales Organization." And across the entire careers pull there is not one GTM-engineering, RevOps-AI, or growth-engineering role. Every AI job at Five9 builds the product Five9 sells. None of them touch how Five9 sells.

That is the shape of the C+. Five9 fixed its contract and never touched its motion.

Five9 tells the SEC one thing and the street another, and the reconciliation is real. The filing says license revenue will likely decline and "may not be offset." Management says compression isn't happening. As established above, concurrency billing reconciles them, which is the most interesting fact in this teardown and appears in approximately none of the coverage. But note what the reconciliation costs: it means Five9's defense is a billing definition most of its own investors have not read, and the company has never explained it on a call.

The analyst who wrote the bear case still hasn't taken it back. Piper Sandler's James Fish downgraded Five9 in January 2026 with language that named the mechanism precisely: "AI concerns continue to weigh on primarily seat-based models like Five9 (despite AI traction and shift towards more consumption). A new CEO has taken over, but the best path medium-term would be likely to go private at this stage." He then raised his price target twice through 2026, from $21 to $24 to $30, and never upgraded off Neutral. Truist went to Buy at $35 and Rosenblatt to Buy at $32 on the same numbers. The house that built the thesis is still not convinced by the rebuttal.

Both companies have quietly stopped selling against software budgets. Five9 now describes its addressable market as roughly $24 billion of CCaaS software expanding to about $234 billion once "contact center labor arbitrage enabled by AI agents" is counted. Sierra's whole ROI story anchors on replacing a $13 human interaction with a sub-$1 AI resolution. Both are pricing against payroll. That is the actual category shift, and it means the competitive set for both companies is BPO vendors and internal headcount, not each other.

And on Sierra's side, the multiple is priced on margins nobody has seen. At Taylor's own confirmed $150 million ARR figure, the May round values Sierra at roughly 105 times revenue. The only third-party gross-margin estimates that show their work, and both explicitly label themselves as inferred because Sierra discloses nothing, land at 55% to 70% against a 75% to 85% software band. Sierra's own blog concedes where that risk sits: "customers aren't beholden to our gross margin, we're beholden to their future success." That is admirable alignment and it is also a company absorbing model-cost volatility on behalf of its customers at a software multiple.

The demand-capture grade is a D and I want to defend the number, because it is the harshest mark this index has given a challenger.

Sierra has no pricing page. sierra.ai/pricing returns "Page not found." There is no Pricing item in the navigation. Across 680 sitemap URLs there is not one page with a price on it, only blog posts and a podcast episode about pricing philosophy. There is no trial, no calculator, no self-serve path, and two calls to action sitewide: "Learn more" and "Sign in." One buyer's guide put it as "you cannot find out what it does costs until you are three sales calls deep."

The usual defense is that nobody self-serves a $300,000 enterprise contract. That defense fails on the evidence, because Intercom's Fin publishes a real number for the identical unit: $0.99 per resolved conversation. That is the same category and the same meter, with the number on the page. Sierra's opacity is therefore a choice rather than a constraint of enterprise selling. And Sierra's own buyers name it: a G2 reviewer lists as their dislike "the limited transparency on technical details and pricing, which makes it harder to fully assess long-term costs and integration, and the fact that scalability and consistency at enterprise scale are still largely unproven."

Second, the hedge. Taylor says on podcasts that escalations are free. Sierra's own blog post says, twice, "in most cases, there's no charge." Those are different claims, and the gap between them is exactly where a procurement lawyer lives. Sierra writes the definition of "resolution" that the customer is billed against, which is the one contract term that determines whether outcome pricing is genuine alignment or a better-marketed meter. To be clear about the evidence: no named Sierra customer has publicly disputed a resolution count. This is a structural critique of a vendor-authored term, not a scandal.

Third, Sierra discloses nothing that would let anyone check the compounding story. It publishes no retention figure, no churn figure, and no usage or quality data of any kind. The sharpest version of this came from an independent reviewer: revenue growth "tells us nothing about whether the AI agents actually work well, whether customers retain them long-term, or what percentage of customer service work is genuinely automated." For a company whose entire pitch is measurable outcomes, publishing no outcome data is a strange place to end up.

And the lock-in is real. A former implementer on Hacker News: "Their implementation is rather cumbersome, requiring implementation fees and AI configuration that is rather bespoke to Sierra. Anyone rolling off of Sierra will find there is nothing they can take with them."

Trigger: sierra.ai/pricing 404s, 680 sitemap URLs contain zero prices, and Intercom's Fin publishes $0.99 per resolved conversation for the same unit.

Why it moves the needle: Sierra's demand-capture grade is the only sub-C row in its scorecard, and it sits on the exact claim its brand is built on. Publishing a floor rate for standard support resolutions, with enterprise and voice still quoted, converts the largest credibility gap in the company into its loudest proof point. The competitor already did it, so the "enterprise deals are bespoke" defense is spent.

Trigger: Sierra's own blog hedges "in most cases, there's no charge" twice, while Taylor says on the record that escalations are free, and independent buyer guides converge on the resolution definition as "the single most consequential clause in the contract."

Why it moves the needle: the vendor authoring the billing term is the structural objection to outcome pricing, and it is the objection every competitor will use as the category matures. A published, versioned definition of what counts as resolved, including the recurrence and abandonment cases, turns Sierra's weakest contract position into a category standard it authors. Sierra already did this once with tau-bench. Do it for the meter.

Trigger: Sierra discloses no retention, churn, usage or quality data, while carrying a multiple of roughly 105x on Taylor's own $150M ARR figure, and the most rigorous public critique of the company is precisely that its growth "tells us nothing about whether customers retain them long-term."

Why it moves the needle: outcome pricing makes net revenue retention an unusually honest metric, because expansion only happens when customers voluntarily route more volume to an agent that is working. Sierra has the one business model where NRR is a direct measurement of product quality rather than of a sales team's upsell motion. Publishing it would be the cheapest, most defensible answer to every bear argument in this piece, and Five9, the company Sierra positions against, publishes its retention every quarter.

The lesson is that a pricing model is a distribution decision, and most companies still treat it as a finance decision. Outcome pricing winning on the merits is the smaller half of the story.

Sierra built the most defensible revenue mechanism in enterprise software and then hid the number behind a sales call, which means the mechanism cannot do any selling for it. Five9 kept a meter everyone assumes is doomed, and quietly made it much harder to kill than the market believes, and has never explained the thing that makes it durable on a single earnings call. Both companies own an asset their go-to-market refuses to show anyone.

When a buyer's first question now goes to a language model instead of a salesperson, a published price stops being a pricing decision and becomes a distribution asset. Five9 publishes a price for every tier that does not contain advanced AI. Sierra publishes none at all. In a category whose entire promise is that you'll finally know what you're paying for, that is the finding.