How SaaS Valuations Are Being Rewritten for the Agent Era

Table of Contents

• The Old Valuation Model and Why It Is Cracking
• Four Valuation Inputs Being Rewritten
• Net Revenue Retention Loses Its Halo
• Gross Margin Meets Inference Cost
• TAM Reframed From Software to Labor
• The Discount Rate of Disruption Risk
• Why Some SaaS Is Re-Rating Up, Not Down
• How GaaS Pricing Changes the Comparable Set
• What the Numbers Look Like in Practice
• Insights Most People Overlook
• References

The Old Valuation Model and Why It Is Cracking

For roughly fifteen years, software was the best business model anyone had ever underwritten. You sold seats, the seats renewed, customers added more seats as they grew, and the marginal cost of serving the next user rounded to zero. That combination -- recurring revenue, negative churn, near-zero cost of goods -- is why investors were willing to pay 10x, 15x, sometimes 30x forward revenue for category leaders. The multiple wasn't a number plucked from the air. It was a bet that this year's revenue would compound for a decade with very little erosion.

The agent era attacks that bet at its foundation, and it does so in a specific way that a lot of commentary misses. The threat isn't that "AI will make software obsolete." Software isn't going anywhere. The threat is that agentic AI changes who and what you're paying for. When the unit of value shifts from "a person using a tool" to "a task that gets completed," the seat -- the thing the whole revenue model was denominated in -- stops being the natural unit of consumption.

Consider what a sales team actually buys today. They buy CRM seats, sales-engagement seats, dialer seats, enrichment seats. Now imagine an agent that handles outbound research, sequencing, and CRM hygiene end to end. The customer doesn't need twelve reps each holding five tool licenses. They might need three reps supervising agents, plus a pool of agent capacity priced per qualified meeting. Seat count goes down. The buyer reappears as someone purchasing an outcome. The vendor's revenue per customer can actually rise -- but only if the vendor captured the agent layer instead of selling the seats the agent just made redundant.

That fork is the whole story. And it's why the same macro trend that craters one company's multiple can expand another's. Valuation in the agent era is no longer a sector-wide multiple; it's a question of which side of the seat-to-outcome transition each company lands on. This is a recurring theme across the GaaS cluster, and it sits right next to the broader debate over whether SaaS pricing fundamentally breaks when one agent replaces a ten-seat team.

Four Valuation Inputs Being Rewritten

Rather than talk about "the multiple" as one thing, it helps to decompose what a revenue multiple actually prices: how durable the revenue is, how profitable it is, how large the opportunity is, and how risky the path is. Agents touch all four.

Net Revenue Retention Loses Its Halo

Net revenue retention (NRR), sometimes called net dollar retention, was the single metric public investors learned to fetishize. Above 120% meant a company could grow even if it never landed another logo, because existing customers expanded on their own. The dirty secret is that most of that expansion was seat expansion -- more employees, more licenses, mechanical growth tied to headcount.

Now run the agent scenario forward. If your customers are quietly automating workflows and their own headcount stops growing -- or shrinks -- the seat-expansion engine reverses. A vendor reporting 125% NRR in 2023 on the back of seat growth could be looking at 100% or below as customers consolidate licenses behind agents. Investors have figured this out, which is why seat-count shrinkage has become a metric SaaS investors now openly fear, and why "expansion revenue" footnotes get far more scrutiny than they did three years ago. The question every analyst now asks is uncomfortable but correct: how much of your NRR is just your customers hiring people?

Gross Margin Meets Inference Cost

The second rewrite is brutal in its simplicity. Classic SaaS carried 75-85% gross margins because serving another user cost almost nothing. Agentic products carry a real, variable cost of goods sold: model inference. Every autonomous task burns tokens, and tokens cost money to a foundation-model provider.

A vendor that bolts an agent onto its product and prices it the old way -- flat per-seat -- can watch margins compress as usage scales, the opposite of the SaaS dream. This is why the move to per-task and per-outcome pricing isn't just a go-to-market preference; it's a margin-defense mechanism. McKinsey's work on the economic potential of generative AI frames the value at the workflow level, but the corollary for vendors is that you have to price at the workflow level too, or inference eats you. Investors now model a gross-margin range, not a gross-margin constant -- and a company that can't articulate how its margins behave as agent usage grows gets penalized for the uncertainty alone. The deeper version of this debate lives in the cluster's discussion of platform risk when you build agents on a foundation-model provider.

TAM Reframed From Software to Labor

Here's the rewrite that cuts the other way, and it's why the doom framing is incomplete. Software's total addressable market was always capped by IT budgets. Labor budgets are an order of magnitude larger. When an agent does work a person used to do, the relevant budget line isn't "software spend," it's "the cost of getting that work done."

a16z and others have argued that agentic services let software vendors reach into the much larger services and labor budget rather than fighting over the software line item. If you accept that reframe, a vendor that successfully sells outcomes isn't capped by a shrinking seat market -- it's playing for a slice of payroll. That's a fundamentally bigger TAM, and a bigger TAM, properly underwritten, supports a higher multiple. The catch is the word "successfully." Most companies asserting this expansion haven't earned it yet, and investors have learned to discount the claim until the outcome-based revenue actually shows up on the income statement.

The Discount Rate of Disruption Risk

The fourth input is the least quantified and arguably the most powerful: the risk premium. A multiple is a discounted stream of future cash flows, and the discount rate embeds risk. The agent era has injected a new, hard-to-model risk into every SaaS name -- the chance that an agent-native startup, or a foundation model itself, disintermediates the product entirely.

You can see this in how differently the market now treats a "system of record" versus a "system of action." Record-keepers with deep, proprietary, hard-to-replicate data get a lower risk premium because their data is a moat. Thin workflow layers get a higher one because an agent can plausibly rebuild that workflow. This is the same tension explored in the system-of-record versus system-of-action battle -- and it shows up directly in valuation as a spread in discount rates between companies that, on a 2020 spreadsheet, would have looked identical.

Why Some SaaS Is Re-Rating Up, Not Down

It's tempting to read all of this as uniformly bearish for software. It isn't. The clearest counter-signal is that incumbents with distribution and proprietary data are, in some cases, being re-rated upward as the market concludes they're more likely to capture agent value than to lose to it.

The logic runs like this. An agent is only as good as the data and the actions it can reach. If a vendor already sits on the system of record -- the CRM with a decade of customer history, the ERP wired into every transaction -- then the agent layer is something they can build on top of an asset nobody else has. Distribution compounds the advantage: shipping an agent to fifty thousand existing enterprise customers is a different proposition than acquiring those customers cold. This is why the smart-money framing has shifted toward distribution beating capability as incumbents' real advantage. The model wars commoditize raw capability; distribution and data don't commoditize.

So you get a bifurcation. Vendors perceived as agent-exposed seat-sellers compress. Vendors perceived as agent-enabled data owners hold or expand. The market is, in effect, repricing the moat, not the revenue -- and the agent era made moats legible in a way the SaaS-multiple-for-everyone era never bothered to.

How GaaS Pricing Changes the Comparable Set

There's a subtler valuation problem nobody fully solved yet: what do you even compare a GaaS company to? Per-seat SaaS had a clean comp set and a clean metric stack -- ARR, NRR, magic number, Rule of 40. Agentic-AI-as-a-service businesses, priced per task or per outcome, don't slot cleanly into that frame.

Outcome-based revenue behaves more like a usage business (think cloud infrastructure or payments) than a subscription business. Usage businesses are valued on different signals: consumption growth, cohort consumption curves, gross margin trajectory, and the durability of the outcome being sold. When procurement starts buying outcomes instead of software, the buyer's own accounting changes too -- the spend migrates from a software line to something closer to cost-of-services or labor, which is exactly the reframe CFOs are now wrestling with.

The practical consequence for valuation is that the comp set is being rebuilt mid-flight. A pure outcome-priced agent company arguably deserves to be benchmarked against a blend: the gross-margin profile of usage software, the retention profile of an embedded services provider, and the TAM of a labor-replacement business. Nobody has a tidy formula for that blend yet, and the absence of a settled comp set is itself a source of valuation volatility. Markets discount what they can't categorize.

What the Numbers Look Like in Practice

Strip away the narrative and the rewrite shows up as a few concrete shifts in how a deal or a public name gets modeled.

First, revenue quality gets unbundled from revenue growth. Two companies growing 30% are no longer worth the same multiple if one grows via seat expansion that agents threaten and the other via consumption that agents drive. Analysts now tag revenue by its exposure to the seat-to-outcome transition, and that tag moves the multiple more than the headline growth rate does.

Second, gross margin becomes a forward variable, not a static fact. Models now carry a margin curve as agent usage scales, and a wide or unknown curve commands a discount. The Bessemer Cloud Index commentary on efficient growth and durable margins captures the broader mood: the market stopped paying for growth-at-any-cost in 2022, and the agent era sharpens that into growth-at-known-margin.

Third, the terminal-value question got existential. A DCF's terminal value assumes the business still exists and earns economic rents at the end of the forecast. For a seat-selling workflow layer with a thin data moat, that assumption is no longer safe, and a lower terminal value drops the whole valuation regardless of near-term numbers. The companies winning the re-rating are the ones that can credibly argue their terminal value is protected -- by data, by distribution, or by owning the agent-of-record relationship with the customer.

Put it together and the "SaaS multiple compression" headline is really three separate repricings stacked on top of each other: retention quality, margin durability, and terminal-value risk. Each maps to a real mechanism in the agent transition, and each can cut either direction depending on the company. That's why the same six months can produce both a brutal de-rating of exposed names and quiet record valuations for the data-rich incumbents on the other side of the trade.

Insights Most People Overlook

• The compression is a transfer, not a destruction. Most coverage treats agent disruption as value evaporating out of the software sector. It's closer to value moving -- from seat-sellers to data-owners, from workflow layers to systems of record, and from software budgets into labor budgets that software vendors can now reach. The sector-level "SaaS is being de-rated" story hides a violent reallocation happening underneath it. The winners' gains are roughly the size of the losers' losses.

• High NRR was a lagging risk indicator, not a strength. The companies with the most spectacular seat-expansion-driven retention are, perversely, the most exposed -- because their growth was most tightly coupled to customer headcount, which is exactly what agents compress. The metric everyone treated as the proof of a moat was actually measuring the size of the cliff.

• Outcome pricing can mask margin risk as easily as it solves it. Everyone frames per-outcome pricing as the elegant fix for inference costs. But if a vendor prices a fixed outcome and inference costs spike, or the agent needs more retries than modeled, the vendor is now short volatility on token costs. Outcome pricing transfers usage risk to the vendor. That's a real liability that the bullish "services-as-software TAM" framing tends to gloss over.

• The most dangerous spot is the "good enough" middle. Deep-moat systems of record are defended; nimble agent-native startups are growing into the gap. The companies in genuine trouble are the mid-market workflow tools that are too commoditized to defend on data and too slow to out-build the startups -- and they're often the ones still reporting healthy-looking SaaS metrics, which is why the market hasn't fully repriced them yet. The re-rating there is still ahead, not behind.

• Valuation volatility is partly an accounting artifact. Part of the multiple chaos is simply that the comp set hasn't settled. Is an agent company SaaS, usage software, or a services firm? Until the market agrees on the right benchmark, identical businesses will get wildly different multiples depending on which comp the analyst reached for -- and a chunk of the "agent era repricing" is really just the market arguing with itself about taxonomy.

References

• The economic potential of generative AI: The next productivity frontier -- McKinsey & Company
• The Next Big Thing: Services-as-Software -- Andreessen Horowitz (a16z)
• State of the Cloud 2024 -- Bessemer Venture Partners