Why Outcome Pricing Quietly Hands the Advantage to Incumbents With Data

Table of Contents

• The Promise and the Catch
• What Outcome Pricing Actually Requires
• Where the Incumbent Edge Compounds
• Baselines You Can Only Build With History
• Reliability Is a Data Problem Before It Is a Model Problem
• Attribution Belongs to Whoever Owns the System of Record
• The Margin Math That Punishes Newcomers
• How Challengers Can Fight Back
• What Buyers Should Watch For
• Insights Most People Overlook
• References

The Promise and the Catch

There is a seductive story being told across the agentic AI-as-a-service market right now, and it goes like this: forget seats, forget tokens, forget the anxiety of a metered bill. Just pay us when the work gets done. A resolved support ticket. A booked meeting. A recovered invoice. A closed deal. The vendor eats the risk; the buyer pays for value.

It is a genuinely good story, and for buyers burned by shelfware SaaS contracts, it lands. The trouble is that outcome pricing is not a neutral container. It quietly encodes a set of requirements about who gets to define an outcome, who can prove one occurred, and who can deliver one cheaply enough to make money on the deal. Once you trace those requirements to their source, you keep arriving at the same place: the company with the most data and the deepest hooks into the customer's workflow wins, often before the contest starts.

This is not an argument against outcome pricing. It is an argument for understanding its gravity. If you are a founder building a vertical agent, or a buyer evaluating one, the structural bias here will shape your odds more than your demo ever will. And it connects directly to the broader debate this cluster keeps circling: the choice between per-task, per-outcome, and per-seat models is never just a pricing decision. It is a decision about who absorbs uncertainty.

What Outcome Pricing Actually Requires

Strip outcome pricing down to its load-bearing parts and you get four:

1. A definition. Someone has to say precisely what counts as the outcome. "Resolved ticket" sounds clean until a customer reopens it two days later, or the resolution was the agent telling them to email a human.
2. A baseline. To charge for an outcome, you implicitly claim the outcome would not have happened anyway. That demands a believable counterfactual: this lead converted because of the agent, not despite it.
3. Attribution. Even with a baseline, you need to trace the result back to the agent's action through whatever messy, multi-touch reality the customer lives in.
4. Repeatable delivery at a margin. The agent has to actually produce the outcome often enough, cheaply enough, that the vendor profits across a portfolio of customers, not just the easy ones.

Each of these four looks like a contract negotiation. Each is actually a data problem. And data problems compound for whoever started collecting first. The audit and definition question alone is thorny enough that it deserves its own treatment, but for our purposes here, notice that all four levers reward accumulated information, and that accumulation is precisely what an incumbent has and a challenger does not.

Where the Incumbent Edge Compounds

Baselines You Can Only Build With History

Consider a company selling an AI sales-development agent on a per-meeting-booked basis. To price that honestly, they need to know the natural booking rate of a given lead segment, otherwise they are charging for meetings that would have happened from a well-timed inbound anyway. An incumbent CRM player like Salesforce or HubSpot, or a sales-engagement incumbent like Outreach, has seen tens of millions of sequences play out. They know the baseline conversion of a Tuesday-morning email to a VP of Finance at a 200-person SaaS company. They can price the incremental outcome with confidence and still protect margin.

A six-month-old startup cannot. It either guesses the baseline (and prices wrong, eroding margin or scaring buyers) or it charges for gross outcomes including the freebies it did not cause (and gets caught the moment a sophisticated buyer runs a holdout test). McKinsey's analysis of how generative AI value gets captured keeps returning to this theme: the economic value of AI accrues disproportionately to organizations that can measure and operationalize it, and measurement is a function of data you already hold.

The baseline problem is the quiet killer. It is invisible in a sales deck and decisive in a renewal.

Reliability Is a Data Problem Before It Is a Model Problem

The agent reliability conversation usually gets framed as a model-quality issue, better reasoning, fewer hallucinations, longer context. That framing flatters model labs and misleads founders. In production, reliability is overwhelmingly about edge cases, and edge cases are learned, not reasoned. The incumbent that has run a workflow ten million times has a library of the weird ways it breaks: the malformed PDF, the customer who replies "k" to a confirmation, the regional tax rule that only applies in Quebec.

Under outcome pricing, reliability is not a nice-to-have; it is the entire business model. If the agent fails the task, the vendor eats the cost of the attempt and earns nothing. A vendor with a 92% success rate and a vendor with a 78% success rate are not 14 points apart, at typical agent unit economics, one is profitable and the other is underwater. Anthropic's own guidance on building reliable agents stresses evaluation against real failure cases over architectural cleverness, and you cannot evaluate against failure cases you have never seen. This ties straight into margin expansion through model routing: incumbents route the easy 80% to a cheap model precisely because their data tells them which 80% is easy.

So the per-outcome model converts a data lead into a margin lead, and a margin lead into a pricing-aggression lead. The incumbent can undercut the challenger and still profit, because their cost-per-successful-outcome is structurally lower.

Attribution Belongs to Whoever Owns the System of Record

Here is the part founders underestimate most. Even if a challenger builds a better agent with a fair baseline, they still have to prove the outcome to get paid, and proof lives in the system of record. The recovered invoice shows up in the customer's ERP. The closed deal closes in the CRM. The resolved ticket resolves in the helpdesk. Whoever owns or sits closest to that system controls the ledger that the invoice is written against.

An incumbent helpdesk vendor charging per resolution, the dynamic at the heart of the Intercom Fin per-resolution model, defines "resolution" inside its own product, using its own signals, with no dispute layer the customer can appeal to. A third-party agent bolted onto that same helpdesk is attributing outcomes through someone else's API, hoping the data it needs is exposed, hoping the definitions line up. The incumbent owns the scoreboard. The challenger is playing an away game in a stadium where the home team keeps the official stats.

The Margin Math That Punishes Newcomers

Outcome pricing is, underneath, an insurance product. The vendor is selling a guarantee and self-insuring against the cases where the agent fails. Insurance is brutally unforgiving to anyone who cannot price risk accurately, and pricing risk requires a loss-history dataset, exactly what a new entrant lacks.

Play it out. A challenger launches a per-outcome agent. Adverse selection bites immediately: the customers most eager to try pure pay-for-results pricing are often the ones whose workflows are hardest, because they have already failed to solve the problem cheaply. The challenger, lacking baseline data, cannot identify and price away these hard accounts. They sign a cohort skewed toward difficulty, burn inference dollars on attempts that never convert, and watch gross margin go negative while reporting impressive top-line "outcomes delivered."

This is one mechanism behind the discounting death spiral that hits early GaaS deals. It is not only that newcomers discount to win logos. It is that outcome pricing hands them a portfolio they are structurally unequipped to price, and the data they would need to fix it only arrives after they have already lost money learning it. The incumbent, meanwhile, was handed that data by a decade of prior business. Andreessen Horowitz has written about how AI startups can struggle to hold gross margins when their cost of delivery scales with usage rather than collapsing toward zero, and outcome pricing is the sharpest version of that trap, because cost scales with attempts while revenue scales only with successes.

How Challengers Can Fight Back

The situation is biased, not hopeless. The most effective counter-moves all aim at the same target: neutralizing the data and attribution advantage rather than pretending it does not exist.

• Pick outcomes incumbents can't cleanly measure. Incumbents own the obvious metrics inside their own products. Challengers should hunt for valuable outcomes that span systems, a result that only becomes visible when you connect the CRM, the billing system, and the support inbox. Cross-system outcomes are messy to attribute, which is bad, but they are messy for everyone, which levels the field. This is the wedge behind pricing agents that act across multiple systems.
• Sell the baseline study as the product's first act. Instead of guessing a counterfactual, run an explicit holdout pilot, share the methodology, and let the customer co-own the baseline. You convert your data weakness into a trust asset. Buyers who have been burned by incumbent black-box "resolution" definitions will pay for transparency.
• Use hybrid pricing as a data-acquisition subsidy. A modest base fee plus outcome upside, structured the way hybrid base-plus-usage models work, funds the inference attempts you need to generate your own loss history. Pure per-outcome pricing starves a newcomer of exactly the data it needs to eventually compete on pure per-outcome pricing. Floors and caps are not a retreat from outcome pricing; they are how you survive long enough to earn it.
• Go vertical and deep before you go broad. Reliability data is domain-specific. A challenger that owns every edge case in, say, dental-practice insurance verification can out-deliver a horizontal incumbent in that narrow lane, and an outcome that the incumbent prices conservatively (because their data is averaged across industries) the specialist can price aggressively.

None of these erase the incumbent advantage. They route around it. The founders who lose are the ones who adopt pure outcome pricing because it tested well in a sales conversation, without doing the loss-history math that determines whether they can survive their own pricing model.

What Buyers Should Watch For

If you are purchasing an agent on outcome terms, the bias above is your leverage, use it. Ask who defines the outcome and whether you can audit the definition. Demand a holdout or A/B baseline before you accept that the agent caused the result; an incumbent confident in its data will agree, and a vendor that refuses is telling you their attribution would not survive scrutiny. Watch for definitions that quietly inflate counts (a "resolution" that is really a deflection, a "qualified lead" that never had budget).

And weigh the trade you are actually making. Outcome pricing with a data-rich incumbent often gets you a fair, defensible price and lower delivery risk. Outcome pricing with a challenger may get you a sharper agent in a narrow domain at the cost of fuzzier attribution. Neither is wrong. But knowing that the pricing model itself favors the incumbent lets you read a too-good-to-be-true outcome guarantee for what it sometimes is, a newcomer underpricing risk it cannot yet measure, which is a deal that tends not to survive its own renewal.

Insights Most People Overlook

• Outcome pricing is a moat-deepening mechanism disguised as a customer-friendly one. Every outcome an incumbent prices and delivers feeds back into its baseline and reliability data, widening the gap. The model does not just reflect the data advantage, it actively compounds it with each transaction. Buyers think they are shifting risk to the vendor; market-structurally, they are subsidizing the leader's moat.

• The hardest customers self-select into pure outcome pricing. Adverse selection is the silent assassin. The accounts most attracted to "only pay when it works" are disproportionately the ones for whom nothing has worked yet. A challenger who wins on outcome terms may be winning exactly the portfolio that bankrupts them, and mistaking it for product-market fit.

• Whoever owns the system of record owns the invoice, regardless of who owns the agent. The deepest incumbent advantage is not the model or even the data, it is control of the ledger the outcome is recorded in. A superior third-party agent can be quietly starved by an incumbent that simply declines to expose the attribution signals in its API. Distribution and data control beat agent quality under this pricing model.

• Cost scales with attempts; revenue scales with successes, and the gap is the whole game. Most pricing analysis treats per-outcome as a cleaner version of per-task. It is not. The structural asymmetry between paying for compute on every attempt while billing only on wins is what makes loss-history data existential rather than merely useful. A 10-point reliability gap is often the difference between a viable business and a charity.

• A newcomer's smartest first product may be the baseline study, not the agent. Counterintuitively, the way out of the data trap is to sell measurement before you sell automation. Co-owning the baseline with the customer turns your biggest weakness, no history, into a transparency advantage incumbents, protective of their black-box definitions, are reluctant to match.

References

• The State of AI, McKinsey & Company
• Building Effective Agents, Anthropic Engineering
• The New Business of AI (and How It's Different From Traditional Software), Andreessen Horowitz