How VCs Are Underwriting GaaS Bets Differently From SaaS

Table of Contents

• The SaaS Underwriting Playbook VCs Are Unlearning
• Why Per-Outcome Revenue Forces a Different Model
• Gross Margin Is Now a First-Order Diligence Question
• The New Risk Stack: Reliability, Model Dependency, and Liability
• How Funds Are Repricing the Same Metrics
• Net Revenue Retention With an Asterisk
• The Burn Multiple Gets a Compute Adjustment
• What VCs Actually Ask in the Room Now
• Insights Most People Overlook
• References

The SaaS Underwriting Playbook VCs Are Unlearning

For most of the 2010s, underwriting a software company was close to a checklist exercise. A partner could look at a B2B SaaS deck and know within a few slides whether the business cleared the bar: annual recurring revenue growing 2.5 to 3x year over year at seed-to-Series-A, gross margins of 75 to 85%, net revenue retention above 110%, and a sales-efficiency ratio that suggested every dollar of go-to-market spend returned more than a dollar of new ARR inside a year. The beauty of seat-based SaaS was that revenue was almost entirely decoupled from cost to serve. Once you'd built the product, the marginal cost of adding one more user was rounding-error small. That's why software earned the multiples it did.

Agentic AI-as-a-Service does not work that way, and the smarter funds figured this out fast. When a company sells an agent that resolves a support ticket, drafts a contract, or reconciles an invoice, the unit being sold is an outcome, not a license. And producing that outcome costs money every single time the agent runs, because tokens, tool calls, and sometimes multiple model passes get consumed on each task. The clean separation between revenue and cost that made SaaS so financeable is gone. That single fact ripples through the entire underwriting process.

The investors who are getting GaaS right aren't throwing out the SaaS framework wholesale. They're treating it as a starting point and then asking, line by line, where the assumptions no longer hold. As Bessemer's analysts have noted in their work on the shift from per-seat to usage- and outcome-based pricing models, the metrics that matter are migrating from "how many seats" to "how much value delivered, at what cost." That reframing is the whole game.

Why Per-Outcome Revenue Forces a Different Model

Per-seat pricing gave you a revenue number you could forecast with a spreadsheet and a sales headcount plan. Per-outcome and per-task pricing gives you something closer to a consumption business, like a cloud-infrastructure vendor, except the consumption is tied to how often the customer has work for the agent to do.

This creates two underwriting problems at once. First, revenue becomes harder to predict because it floats with the customer's own activity. A legal-research agent billed per brief will see revenue collapse if the client's litigation pipeline slows. Second, and more subtly, the company only captures value when the agent actually succeeds. If you're charging per resolved ticket and the agent resolves 70% of them, you're billing on 70% while paying compute on 100% of attempts, including the failures and the retries.

VCs underwriting these businesses now spend real time modeling the attempt-to-success ratio and how it trends as the product matures. A SaaS investor never had to care whether the software "worked" on a given session because the customer paid regardless. A GaaS investor cares enormously, because reliability sits directly upstream of both revenue recognition and cost of goods. The pitch decks that survive diligence increasingly show a curve: success rate climbing while cost-per-successful-outcome falls. The decks that don't survive show top-line growth with no view into what each outcome costs to produce. (This gap is exactly the kind of thing the agentwashing problem in fundraising decks is meant to flag.)

There's also a measurement headache that good investors probe. Outcome-based revenue requires both parties to agree on what counts as a delivered outcome, and that definition can be gamed or disputed. A founder who books revenue on "outcomes" that the customer later contests is sitting on revenue quality that won't survive an audit. This is why the revenue-quality and durability questions have become their own diligence workstream.

Gross Margin Is Now a First-Order Diligence Question

In SaaS, gross margin was almost a formality. You assumed 75%-plus and moved on to growth. In GaaS, gross margin is where deals live or die, and it's volatile in a way software margins never were.

The reason is that cost of goods sold for an agent company is dominated by inference spend, which the company largely does not control. When a model provider cuts prices, margins expand overnight. When a company has to route to a more capable, more expensive model to hit an accuracy bar, margins compress. A startup can report a 30% gross margin in Q1 and a 55% gross margin two quarters later purely because of model-price movements and prompt-engineering optimizations, with no change to the actual product. That kind of swing would have been unthinkable in seat-based software.

So underwriters now demand a real cost-of-goods buildup. They want to see token consumption per task, the blended cost across the model mix, how much of the workflow runs on cheaper small models versus frontier models, and what caching and retrieval optimizations are already in place. They're trying to answer a question that simply didn't exist five years ago: is this a structurally high-margin business that's temporarily margin-thin because it's young, or is it a structurally thin-margin business dressed up as software?

The market has not fully settled on how to price this. Some funds argue agent companies should be valued on a discount to SaaS multiples until margins prove out; others argue the best agents will reach SaaS-like margins as model costs fall on a Moore's-Law-style curve, and you should underwrite to that future. McKinsey's work on the economic potential of generative AI leans toward the optimistic case on cost trajectories, but even its analysis assumes meaningful workflow redesign, not a free ride. The honest answer is that margin destiny is the central unresolved bet in GaaS investing, and where a fund lands on it largely determines what it's willing to pay.

The New Risk Stack: Reliability, Model Dependency, and Liability

Beyond the financial model, GaaS introduces categories of risk that simply weren't on the SaaS diligence checklist. Three stand out.

Reliability as an existential variable. A SaaS tool that's down for an hour annoys people. An autonomous agent that takes a wrong action, such as issuing a refund it shouldn't, sending a bad email to a customer, or misclassifying a compliance document, can create real liability. Investors now ask how the company bounds agent behavior, what human-in-the-loop checkpoints exist, and whether there's an audit trail. A company selling autonomy without guardrails is selling a lawsuit with good growth metrics.

Model dependency. Most GaaS companies are, at some layer, reselling intelligence they license from a foundation-model lab. That creates a supplier-concentration risk SaaS never had. If a single provider can raise prices, deprecate a model version, change usage policies, or simply launch a competing first-party agent, the startup's entire cost structure and moat are exposed. Diligence now routinely includes the question: what happens to this business if your primary model provider becomes your competitor? The strongest answers involve model-agnostic architectures and proprietary data or workflow assets that survive a provider switch. This tension between the labs and the application layer is one of the defining capital dynamics in the agent market.

Liability and indemnification. When you sell outcomes rather than tools, you implicitly take on more responsibility for results. Enterprise buyers are pushing agent vendors for indemnification on agent errors, which changes the risk profile of the revenue. A dollar of outcome-based revenue that carries error liability is not worth the same as a dollar of seat-based SaaS revenue that carries none. Sophisticated investors are starting to risk-adjust revenue accordingly.

How Funds Are Repricing the Same Metrics

Rather than invent an entirely new vocabulary, most funds are keeping the familiar metrics and quietly changing how they read them.

Net Revenue Retention With an Asterisk

NRR is still the headline retention metric, but its meaning shifts. In SaaS, expansion came from adding seats. In GaaS, expansion comes from the customer routing more tasks to the agent, which can grow fast but is also more reversible. A customer can ramp usage 3x in a quarter and cut it just as quickly if the agent underperforms or if they bring the workflow in-house. So a 130% NRR in GaaS is treated as higher-variance than the same number in SaaS. Underwriters dig into whether expansion is driven by genuine workflow adoption or by a few power users whose usage could evaporate. They also separate "good" expansion (more workflows automated) from "fragile" expansion (one customer, one use case, scaling unsustainably).

The Burn Multiple Gets a Compute Adjustment

The burn multiple, net burn divided by net new ARR, popularized as a capital-efficiency yardstick, gets a GaaS-specific wrinkle. Because compute is a variable cost that scales with revenue, a fast-growing agent company can show a deceptively healthy burn multiple while quietly running a low-margin business that will never throw off cash at scale. Conversely, a company investing heavily in margin optimization may look worse on burn today but is building toward durable economics. Investors now read the burn multiple alongside the gross-margin trajectory rather than in isolation. The "default alive" math that David Sacks and others applied to SaaS still applies, but the inputs are noisier.

The net effect is that GaaS underwriting trades some of SaaS's predictability for a thicker layer of scenario analysis. Funds model margin-up and margin-down cases driven by model-price movements, and they stress-test revenue against a reliability-regression scenario. That's a meaningfully more involved process than the SaaS checklist, and it's reshaping which firms have an edge: those with the technical depth to actually evaluate an agent's architecture and cost structure, not just its growth curve.

What VCs Actually Ask in the Room Now

If you sat in on a GaaS pitch at a fund that has done its homework, the questions sound different from a SaaS pitch circa 2019. Instead of "what's your magic number," you hear: What does it cost you to deliver one successful outcome today, and what was it six months ago? What's your success rate, and how is it trending? What share of your COGS is model inference, and how exposed are you to a single provider's pricing? When the model gets cheaper, who captures the savings, you or your customer? What happens to your gross margin if you have to upgrade to a more expensive model to keep accuracy? What's your liability exposure on agent errors, and is it insured or indemnified?

None of these are growth questions. They're all margin, reliability, and dependency questions. That's the tell that GaaS underwriting has genuinely diverged from the SaaS template, even as the surface-level metrics look familiar. The funds that internalized this early are paying premium valuations for the companies that can answer cleanly, and walking away from the ones that can only talk about top-line. For a fuller picture of how those premiums get justified, the broader debate over agent-company valuations is worth tracking.

Insights Most People Overlook

Falling model costs can be a trap, not a tailwind. Everyone assumes cheaper inference helps agent margins. But if a startup's pricing is benchmarked against the cost of the work it replaces, and model costs fall industry-wide, competitive pressure can force the startup to pass those savings straight to customers. The margin expansion you underwrote may get competed away. The companies that keep the savings are the ones with pricing power from switching costs or proprietary data, not the ones merely riding cost curves.

Outcome-based pricing quietly transfers model risk onto the startup. Under per-seat pricing, the customer absorbs the cost of inefficiency. Under per-outcome pricing, the vendor eats every failed attempt and retry. This means a GaaS company's margins are hostage to its own accuracy in a way SaaS never was. A two-point drop in success rate isn't just a churn risk; it's a direct hit to cost of goods. Few decks model this sensitivity, and few investors push hard enough on it.

The best GaaS businesses may look more like staffing firms than software. If you charge per resolved outcome and your cost scales with volume, your financial profile rhymes with labor outsourcing or BPO more than with classic software. Some contrarian investors are deliberately underwriting GaaS to BPO-style margins (think 40-60%, not 80%) and still finding the deals attractive because the addressable market is labor budgets, not software budgets. That reframing, agents priced against payroll rather than against SaaS line items, expands the prize enough to justify thinner margins.

Reliability data is the real moat, and it's underpriced in diligence. The proprietary asset that's hardest to replicate isn't the model or the prompt; it's the accumulated record of which agent actions succeeded and failed in production, which feeds back into better routing and guardrails. Investors fixated on model choice often miss that the durable advantage is this operational data flywheel. A company two years into running agents at scale has reliability data a well-funded newcomer simply cannot buy.

"Recurring" revenue may be the wrong word entirely. Much GaaS revenue is recurring only in the sense that the customer keeps having work to do. It's closer to transaction revenue than subscription revenue. Investors who underwrite it as sticky subscription ARR are overpaying; those who underwrite it as durable transaction flow, and separately verify the durability, are pricing the risk correctly. The vocabulary lag, calling consumption revenue "ARR," is causing real mispricing across the category.

References

• Bessemer Venture Partners, Scaling to $100 Million (cloud and usage-based metrics)
• McKinsey, The Economic Potential of Generative AI: The Next Productivity Frontier