Seed-Stage GaaS: What Investors Actually Want to See Before They Write the Check

Table of Contents

• Why Seed-Stage GaaS Diligence Looks Different
• The Five Signals Investors Weight Most
• 1. Proof the Agent Finishes the Job
• 2. Margin That Survives the Inference Bill
• 3. Revenue That Doesn't Evaporate
• 4. A Wedge, Not a Feature
• 5. Founders Who've Run the Agent in Anger
• The Metrics Seed Investors Ask For (and the Numbers They Like)
• Red Flags That Sink a Seed Round
• How the Pitch Itself Should Be Built
• Insights Most People Overlook
• References

Why Seed-Stage GaaS Diligence Looks Different

For fifteen years, seed investors evaluated software companies with a familiar lens: a sticky product, a clean SaaS gross margin in the high 70s or 80s, net revenue retention above 110 percent, and a sales motion you could pour fuel on. GaaS breaks several of those assumptions at once, and the investors who are good at this know it.

The first break is the cost structure. A SaaS company's marginal cost of serving one more customer rounds to zero. An agent company's marginal cost is a live inference bill that scales with usage, sometimes brutally. When your agent runs a twelve-step research workflow on behalf of a customer, you are paying a model provider for every token in every step, plus retries, plus the failed attempts the customer never sees. That changes what "good unit economics" even means at seed, and it's why a sharp investor will ask about your cost per completed task before they ask about your logo wall.

The second break is the value proposition itself. SaaS sells software that a human operates. GaaS sells the outcome, the contract reviewed, the lead qualified, the support ticket resolved, the candidate sourced, often with no human in the loop. That shifts the buyer from a software budget to a labor budget, which is a far bigger pool, but it also raises the bar on reliability. Software that's 95 percent reliable is annoying. An autonomous agent that's 95 percent reliable means one in twenty outcomes is wrong, and at seed stage you need a believable story about how you close that gap. Sequoia's writing on the the "Act Two" of AI built on agents and reasoning frames this well: the market has moved from selling tools to selling work, and the economics of selling work are different.

The third break is timing. Many seed-stage GaaS founders are raising on a demo and a handful of design partners rather than meaningful ARR, because the category is young and the technical risk is real. That pushes diligence away from spreadsheets and toward judgment: can this team actually ship a reliable agent in a domain where reliability is hard? Seed has always been a bet on people, but in GaaS the technical bet is unusually load-bearing.

The Five Signals Investors Weight Most

Across the seed checks getting written in this category, five signals come up again and again. Hit four of them convincingly and you're in the conversation. Miss the first two and most experienced GaaS investors will pass no matter how good the deck looks.

1. Proof the Agent Finishes the Job

The single most important thing you can show is that your agent completes the task autonomously, end to end, at a rate that would make a buyer fire the manual process. Investors have learned to distinguish a demo that works on a curated happy path from a system that holds up on messy, real-world inputs. The way you prove the difference is with a completion metric on real customer data: out of 100 tasks attempted, how many finished correctly with no human intervention?

Be honest about the denominator. A 90 percent autonomous-completion rate on a narrow, well-bounded task is genuinely impressive. A 60 percent rate where humans clean up the rest can still be a business, but then you're selling a copilot, not an autonomous agent, and you should price and pitch accordingly. The founders who get marked down are the ones who blur the two, quoting a completion number that quietly includes heavy human review.

2. Margin That Survives the Inference Bill

Seed investors in this space have been burned by companies that looked great until you backed out the model costs. The number they want is gross margin after inference and tooling, what's left after you pay your model provider, your vector database, your orchestration, and any third-party API calls the agent makes to do its job.

If your blended gross margin is 35 percent today, that's not automatically disqualifying at seed, but you need a credible curve toward 60 to 70 percent. That curve usually comes from three places: routing easy steps to cheaper models, caching and reusing intermediate work, and the secular decline in model prices. The catch, and a good investor will probe this, is that falling model prices help everyone, so they don't constitute durable margin advantage on their own. (This margin-versus-model-cost dynamic gets its own deep treatment in #357, Valuation haircuts when model costs compress margins, and the run-cost side in #346, The burn-rate problem: agents are expensive to run.)

3. Revenue That Doesn't Evaporate

GaaS pricing is often usage-based or outcome-based: per task completed, per ticket resolved, per dollar of value delivered. That's attractive because it aligns price with value, but it raises a durability question that haunts seed diligence, is this usage going to recur, or did the customer run a one-time backlog through your agent and then go quiet?

Investors want to see cohort behavior: do customers who started three or six months ago still generate similar or growing usage today? Flat-to-up usage cohorts are gold. A spiky pattern, where accounts light up and then fade, signals that you sold a project rather than a process. This is the central question of revenue quality in agent companies, and it's why some firms now run explicit usage-durability checks, a topic explored further in #335, The revenue-quality question: is usage revenue durable?

4. A Wedge, Not a Feature

The hardest question in any GaaS seed pitch is: why can't the foundation-model lab, or the incumbent SaaS vendor in your category, just do this? "We have a thin wrapper around a frontier model" is not an answer. The answer that works at seed is some combination of proprietary workflow data, deep integration into a system of record, a hard-won understanding of an ugly domain, and a feedback loop that makes your agent get measurably better with use.

Vertical agents tend to pitch this more convincingly than horizontal ones, because the moat lives in the messy specifics, the compliance rules of a niche, the data schema of an industry tool, the edge cases only operators know. A horizontal "AI agent for anything" almost always struggles to articulate why it won't be commoditized. This is also why agent startups can command the premiums they do when the wedge is real, a dynamic unpacked in #332, Why agent startups command premium valuations.

5. Founders Who've Run the Agent in Anger

Seed is a people bet, and in GaaS the most reassuring founder profile is one who has personally felt the pain of making an agent reliable in production. That experience shows up in how they talk: they discuss eval suites, guardrails, failure modes, and the long tail of edge cases without prompting. Founders who only talk about TAM and "the agentic revolution", and go vague when you ask how they measure correctness, get discounted fast.

The other thing investors look for is domain credibility. An agent for medical coding, freight brokerage, or legal review is far more fundable when at least one founder has lived inside that workflow, because they know which failures are tolerable and which ones get you fired by the customer.

The Metrics Seed Investors Ask For (and the Numbers They Like)

You won't have all of these at seed, and that's fine, but you should know which ones you have, which you're choosing not to show, and why. The strongest decks lead with two or three of these and have the rest ready in the data room.

• Autonomous completion rate. Tasks finished correctly with no human in the loop, on real data. Directionally, narrow-task agents that crack 85 to 95 percent are in strong shape; broad agents below 70 percent need a copilot framing.
• Cost per completed task / outcome. Your fully loaded inference-plus-tooling cost to deliver one unit of value. Investors want to see this trending down as you optimize.
• Gross margin after inference. Today's number plus a believable path to 60 to 70 percent.
• Usage-cohort retention. Are month-three and month-six cohorts holding or growing their usage? Flat-to-up is the goal.
• Time-to-value for a new customer. How fast does the agent go from signed to producing outcomes? Fast onboarding is a real moat in a category where buyers are skeptical.
• Eval coverage. Do you have a real evaluation harness that catches regressions before customers do? This is becoming table stakes, and its absence is a tell.
• Design partners with budget. A handful of paying design partners beats a long list of free pilots every time.

It's worth saying plainly: McKinsey's research on the economic potential of generative AI and agents frames the addressable market in the trillions, which means TAM is rarely the constraint. At seed, almost no one is rejected for too small a market; they're rejected for not having a credible path to reliable, defensible, margin-positive delivery inside it.

Red Flags That Sink a Seed Round

The fastest way to understand what investors want is to know what makes them walk. A few patterns reliably end conversations.

Agentwashing. Rebranding a deterministic workflow or a simple chatbot as an "autonomous agent" is the cardinal sin of this cycle. Experienced investors test it in thirty seconds by asking what the agent decides on its own versus what's hard-coded. If the honest answer is "not much," the premium evaporates. This is endemic enough to have its own coverage in #333, The "agentwashing" problem in fundraising decks.

Margins hidden behind subsidized inference. Some founders quietly run on free or steeply discounted model credits and present the resulting margins as if they were durable. When the credits expire, the economics invert. Smart investors normalize your costs to market rates before believing any margin number.

No eval story. A founder who can't explain how they measure whether the agent is getting better or worse is telling you they're flying blind. In a domain where a single confidently wrong output can lose a customer, that's disqualifying.

Revenue that's really one-time. Big initial usage that turns out to be a backlog dump, dressed up as recurring. Cohort data exposes this every time.

A wedge that's a wish. "Our moat is our team" or "our moat is being first" rarely survives a probing question. If the moat isn't data, distribution, integration, or genuine domain depth, it probably isn't a moat.

How the Pitch Itself Should Be Built

If you're raising, structure the narrative around the buyer's labor budget, not a software-feature list. Open with the specific, expensive, repetitive work your agent does autonomously, name the role or process it replaces or augments, and quantify the outcome. Then move straight to evidence: completion rates on real data, the unit economics, and a usage cohort or two. Save the market-size slide for last, in GaaS it's confirmation, not the argument.

Two craft notes. First, show your failure handling. Walking an investor through how your agent detects when it's unsure and escalates is more persuasive than any happy-path demo, because it proves you understand the reliability problem you're actually solving. Second, be precise about what's autonomous and what isn't. Investors trust founders who draw that line clearly far more than founders who imply full autonomy and hope no one checks. The broader shift in how capital is being underwritten here, away from the SaaS playbook, is the subject of #334, How VCs are underwriting GaaS bets differently from SaaS, and it's worth reading alongside this piece if you want to model the investor's full mental checklist.

Finally, raise the right amount. GaaS companies burn on inference and on the eng talent needed to make agents reliable, so default-dead math is easy to stumble into if you raise a thin round and price too cheaply. Know your run-rate cost per outcome, know your gross margin after inference, and raise enough runway to prove durable usage, not just initial traction.

Insights Most People Overlook

• Falling model prices are a trap, not a tailwind, if that's your whole margin story. Every founder pitches "and inference keeps getting cheaper" as a margin lever. It's true, but it accrues to your competitors equally, and to your customers, who will eventually expect price cuts. Durable margin comes from your optimizations (routing, caching, smaller fine-tuned models for sub-steps), not from the industry-wide price curve. Investors who've done this twice now mentally subtract the free-rider portion of your margin story.

• A lower autonomy percentage with honest framing beats a higher one with hidden humans. Founders instinctively inflate completion rates to look more "agentic." But seasoned investors have seen the inflated version blow up in diligence when they ask to watch ten live runs. A founder who says "we're at 72 percent fully autonomous and here's exactly how we get to 90" earns more trust, and often a better valuation, than one quoting a suspiciously clean 96 percent.

• The best moat at seed is often onboarding speed, not algorithmic cleverness. In a category where buyers are nervous about handing work to an agent, the company that gets a customer to first successful outcome in days rather than weeks builds trust and switching costs faster than the one with a marginally smarter model. This is underweighted in most decks because it feels unglamorous.

• Outcome-based pricing can mask a brutal margin problem. Charging per resolved ticket sounds aligned and investor-friendly, but if your hardest tickets cost more in inference than you charge for them, your best customers, the ones sending you the gnarly work, are your least profitable. Investors increasingly ask for margin by task difficulty, not blended, and founders who haven't looked get caught flat.

• "Who's your real competitor" usually isn't another startup, it's the status quo of doing nothing. Many GaaS seed pitches over-index on competitive differentiation against other agent startups when the actual battle is inertia and trust. Investors quietly favor founders who understand that the hard sell is convincing a buyer to let an agent do the work at all, not beating a rival agent on a feature.

References

• Generative AI's Act Two, Sequoia Capital
• The economic potential of generative AI: The next productivity frontier, McKinsey & Company