The Contrarian GaaS Bets VCs Are Quietly Making

Table of Contents

• The Consensus Trade Everyone Is Crowded Into
• Bet One: Boring Verticals Over Horizontal Glamour
• Bet Two: Companies That Sell Outcomes And Eat The Liability
• Bet Three: Reliability Infrastructure, Not Agents
• Bet Four: Margin Expansion As Models Commoditize
• Bet Five: The "Not An AI Company" Founder
• How These Investors Underwrite The Downside
• Insights Most People Overlook
• References

The Consensus Trade Everyone Is Crowded Into

Walk any Sand Hill Road partner meeting in 2026 and you'll hear the same pitch flavors on repeat. An agent that writes code. An agent that books meetings. An agent that does customer support. An "AI workforce" you hire by the seat. These categories absorbed the overwhelming majority of agentic AI-as-a-service capital over the past eighteen months, and the valuations show it -- premium multiples on revenue that, in many cases, is barely a year old.

There's nothing wrong with those bets individually. The problem is crowding. When forty funded companies attack the same SDR-agent wedge with the same underlying model and roughly the same demo, the differentiation collapses into distribution and burn. The winner is often just whoever raised the biggest war chest, which is a capital game, not a product game. McKinsey's own work on the economic potential of generative AI put trillions of dollars of value on the table, and the venture market read that as permission to fund the same obvious applications dozens of times over.

The contrarians I've watched aren't ignoring the size of the prize. They're avoiding the part of the board where everyone else is standing. Here's where they're standing instead.

Bet One: Boring Verticals Over Horizontal Glamour

The loudest GaaS companies are horizontal -- one agent, every industry, land-and-expand. The quiet money is going the other way: into agents that do one unglamorous thing for one unglamorous industry, and do it so completely that no horizontal player will ever bother to compete.

Think freight brokerage reconciliation. Title-search abstraction for real estate closings. Prior-authorization handling for a specific category of medical procedures. Dealer-warranty claims for heavy equipment. These are not categories that generate Twitter threads. They are categories where a competent agent replaces a genuinely painful manual workflow, where the buyer has a budget line item that already exists, and where domain data is the moat rather than the model.

The logic is straightforward once you say it out loud. A horizontal agent has to be good at everything and is therefore deeply defensible at nothing. A vertical agent that has ingested ten years of a niche industry's edge cases -- the weird state-specific form, the carrier that rejects claims for an obscure reason, the seasonal pattern nobody documents -- builds an accuracy lead that compounds. a16z has argued for years that in applied AI, the data and workflow lock-in matter more than the model, and the vertical-agent bet is that thesis taken to its logical conclusion.

The contrarian wrinkle: these companies look small at Series A. Their TAM slides are unimpressive. Their logos are regional. That's exactly why the valuation entry point is sane, and why a patient investor can own a category instead of renting a feature.

Bet Two: Companies That Sell Outcomes And Eat The Liability

Most GaaS pricing is still a thin disguise over SaaS -- a per-seat or per-month subscription with "agent" stapled to the name. A smaller, more interesting group prices per outcome and, critically, absorbs the cost when the outcome is wrong.

That second clause is the whole bet. Anyone can charge per resolved ticket. The contrarian companies guarantee the resolution and reimburse the customer when the agent fails -- a refunded claim, a covered chargeback, a paid-for human escalation. They are, in effect, selling insurance on their own reliability. It sounds reckless. It's actually a filter.

A founder who will eat the liability is telling you something a deck never can: they have measured their agent's real-world failure rate, priced it, and concluded the unit economics survive it. Most agent startups cannot make that claim honestly, which is why most won't offer the guarantee. The ones that do have effectively pre-completed the buyer's due diligence. They've also built a business that gets structurally harder to dislodge, because switching to a competitor means giving up a guarantee and re-taking the operational risk yourself.

Investors backing this model are underwriting a different question than the usual "how fast does ARR grow." They're underwriting "is the loss ratio stable and improving" -- a question that belongs more to a property-casualty underwriter than to a software VC. That mismatch with the standard playbook is precisely why the category is under-funded, and why early movers can get in cheaply.

Bet Three: Reliability Infrastructure, Not Agents

The crowd buys agents. A quieter set of funds is buying the tooling that makes agents trustworthy enough to deploy -- evaluation harnesses, simulation environments, guardrail and policy layers, observability built specifically for non-deterministic systems, and the audit trails that compliance teams will demand before any agent touches a regulated workflow.

This is the "picks and shovels" instinct applied to the agent economy, and it rhymes with how the smartest money played previous platform shifts. During the cloud build-out, monitoring and security companies often returned better than the apps they watched over. The bet here is that as agents move from pilots to production, the bottleneck stops being "can the agent do the task" and becomes "can we prove it did the task correctly, and catch it when it didn't."

The contrarian edge is timing. Reliability infrastructure feels premature when most agents are still demos. But the companies building it now are positioning for the moment -- arriving faster than most realize -- when an enterprise board asks who is liable for an autonomous decision, and the answer requires the exact observability and audit layer these startups sell. Anthropic's own published guidance on building effective agents underscores how much of real agent work is orchestration, evaluation, and control rather than raw model capability -- which is a direct tailwind for the infrastructure bet.

Bet Four: Margin Expansion As Models Commoditize

A widely held fear among GaaS skeptics is the burn-rate problem: agents are expensive to run, inference costs can devour gross margin, and a price war among model providers could crush the application layer. Several contrarian investors look at the same facts and reach the opposite conclusion.

Their reasoning: model capability per dollar has been falling fast and shows no sign of stopping. For an agent company whose customer pays a fixed per-outcome price, every drop in inference cost flows straight to gross margin. The companies that look margin-thin today, running on frontier models, become structurally profitable tomorrow simply by riding the cost curve down -- or by swapping in a cheaper open model for the eighty percent of tasks that don't need the flagship.

The contrarian doesn't fear commoditization of intelligence; they're long it. The bet is to find agent companies whose pricing is decoupled from their input costs, so that the deflation in model prices becomes their margin expansion rather than a competitive race to zero. This inverts the usual anxiety captured in debates about whether usage-based AI revenue is durable -- the contrarian view is that durability improves precisely as the underlying models get cheaper, provided the company captured value at the outcome layer rather than reselling tokens.

The risk, and these investors know it, is that customers also see the cost curve and demand price cuts. The defensible companies are the ones where the buyer cares about the outcome, not the mechanism -- a freight broker doesn't audit your token bill, they care whether the load got covered.

Bet Five: The "Not An AI Company" Founder

The final quiet bet is about people. The consensus founder profile in GaaS is the ex-frontier-lab researcher or the serial AI founder who can talk transformers fluently. Some contrarians deliberately seek the opposite: the operator who spent fifteen years inside the industry the agent serves and treats the AI as plumbing, not identity.

The pattern these investors have noticed is that domain operators sell differently. They don't pitch "agentic AI." They pitch "I will cut your claims-processing cost by sixty percent," and the AI never comes up until the contract is signed. They have the relationships, they speak the buyer's language, and they instinctively understand the edge cases that wreck a generic agent in production. When the technically brilliant team and the domain-native team collide in the same vertical, the domain team often wins the enterprise deal -- because enterprise buying is about trust and workflow fit, not benchmark scores.

This bet is contrarian because it underweights exactly the credential most VCs over-index on. A pitch that barely mentions AI reads as unsophisticated to a pattern-matching associate. To a contrarian, it reads as a founder who has internalized that the agent is a means, and the business is the end.

How These Investors Underwrite The Downside

What unites these five bets isn't optimism -- it's a different way of pricing risk. Consensus GaaS underwriting borrows the SaaS template: top-line growth, net revenue retention, magic number, sales efficiency. The contrarians have quietly rebuilt the checklist for a world where the product is non-deterministic and the cost base moves under your feet.

They ask about the loss ratio when the agent is wrong. They ask whether revenue survives the model layer getting commoditized. They ask whether the moat is data and workflow lock-in or just a clever prompt that any competitor can replicate in a weekend. They stress-test what happens in a funding crunch, because a company priced for boring economics rarely needs a heroic mega-round to stay default-alive. These are the same instincts behind the broader shift in how VCs underwrite agent bets differently from SaaS -- the contrarians simply got there first and went further.

None of this guarantees returns. Boring verticals can stay too small. Outcome-liability models can be wrecked by one bad loss year. Infrastructure can get absorbed by the model labs themselves. But the contrarian premise is that the crowded trades carry their own, larger risk -- the risk of paying a consensus price for a feature that a better-capitalized competitor commoditizes. Against that, a sane entry price on a defensible, unglamorous business looks less like a gamble and more like discipline.

Insights Most People Overlook

• The guarantee is the diligence. When a GaaS company offers to eat the cost of its own failures, it has done your reliability audit for you. Founders only make that offer when they've measured the failure rate and the math still works. The presence or absence of an outcome guarantee tells you more than any benchmark in the deck.

• Commoditization of models is bullish for the right application layer, not bearish. The reflexive worry is that cheap intelligence destroys agent margins. The inverse is true for companies that priced on outcomes rather than tokens -- every cost decline becomes margin they keep. The skeptics and the contrarians are reading the same cost curve and betting opposite directions; the difference is whether pricing is coupled to input cost.

• "Unimpressive TAM" at Series A is sometimes the feature, not the bug. A vertical agent with a regional logo list and a niche market looks like a pass to growth-obsessed funds. But a small, defensible category bought at a sane price can compound into category ownership, while the horizontal land-grab dilutes into a capital war nobody wins on product.

• The least AI-fluent founder can be the strongest enterprise seller. Buyers in regulated, workflow-heavy industries don't buy "agentic AI" -- they buy cost reduction from someone they trust. The operator who treats AI as plumbing often out-sells the researcher who treats it as identity, because the enterprise sale rewards domain trust over technical pedigree.

• Reliability infrastructure is a bet on liability, not capability. The infrastructure thesis only pays off when an enterprise board demands to know who is accountable for an autonomous decision. That moment is a governance event, not a technical one -- which is why the timing feels early right up until it's suddenly late.

References

• The economic potential of generative AI: The next productivity frontier (McKinsey)
• AI Will Change Everything About the Software Business (a16z)
• Building Effective Agents (Anthropic)
• Bessemer Venture Partners -- State of the Cloud / Atlas research