Per-Resolution Is the New Per-Seat: Who Wins the AI Agent Pricing War

The Pricing Pivot Nobody Is Calling What It Is

On April 14, 2026, HubSpot announced that its Breeze agents would bill at $0.50 per resolved conversation and $1.00 per qualified lead. The announcement was framed as a customer-friendly innovation: pay only when the agent delivers. Within weeks, the pattern was visible across the entire category. Zendesk had already established a $1.50–$2.00 per automated resolution tier. Intercom's Fin agent was processing roughly two million weekly resolutions at $0.99 each. Salesforce introduced something called an "Agentic Work Unit," a proprietary metric that functions as the billing atom for its agent platform. Adobe CX Enterprise moved to credits. Sierra AI, operating on a pure outcome model, had crossed $150 million in annual recurring revenue. What looked like a series of independent product decisions was, in aggregate, a coordinated industry posture.

The narrative attached to this posture is consistent and carefully constructed. Vendors describe AI agent outcome-based pricing as "value alignment" — the idea that buyers only pay when the product works, eliminating the waste embedded in flat per-seat contracts. This framing is coherent enough to survive a sales call. It does not survive scrutiny. The asymmetry at the center of every outcome-based contract is this: the vendor defines what "resolved" means, the vendor measures whether resolution occurred, and the vendor generates the invoice. The buyer receives the bill and a summary. That is not value alignment. It is a measurement monopoly dressed in the language of fairness.

The thesis here is not subtle: outcome-based pricing, as currently structured across the AI agent market, is a margin and lock-in architecture. The audit problem — who decides what counts as a billable outcome — is not a footnote to the pricing debate. It is the entire game. Everything else, the per-unit prices, the product names, the customer success narratives, is downstream of that single structural fact.

This pattern is not confined to the application layer. Anthropic's April 2026 enterprise contract shift moved to per-seat plus API overage. OpenAI's ChatGPT Workspace Agents transitioned to credit billing after the May 6 free preview ended. The stack is repricing simultaneously at every layer, which means the pricing architecture is not an accident of product management. It is a deliberate structural choice propagating from model providers through platform vendors to end buyers.

How We Got Here: The Per-Seat Model's Slow Collapse

Per-seat SaaS pricing worked because it was administratively simple and loosely accurate. When software was primarily a productivity multiplier for human workers, the unit of consumption was a person, so the unit of billing was a person. A company with two hundred salespeople bought two hundred seats of Salesforce. The pricing model was imprecise but directionally correct: more users meant more value extracted, more seats purchased. Finance teams could predict spend. Vendors could predict revenue. The model had the great virtue of mutual legibility.

AI agents break that logic at the foundation. A single agent can handle the volume of dozens of human workers, which means per-seat pricing becomes incoherent in both directions. Price the agent at human-equivalent rates and you've built a product nobody will buy. Price it at software rates and you're massively undercharging for the value delivered. Vendors faced a genuine unit economics problem that per-seat pricing could not resolve, and the first-generation response was usage-based pricing tied to API calls, tokens, or compute minutes. This solved the volume problem and immediately created a different one: buyers couldn't predict costs, couldn't connect spend to business outcomes, and couldn't explain their AI budgets to CFOs who wanted to know what they were getting for the money.

Outcome-based pricing presents itself as the synthesis. Predictable per-unit cost, tied to a business result, no waste from unused capacity. The sales pitch is coherent. It addresses real pain points that usage-based billing created. But it is incomplete in ways that matter enormously for buyers, because it resolves the buyer's prediction problem by transferring measurement control entirely to the vendor.

The mental model for this kind of pricing was already being built before the AI agent market arrived. Clay in the sales intelligence space trained buyers to think in terms of contacts enriched and workflows executed. Apollo in outbound prospecting oriented buyers around sequences sent and qualified contacts surfaced. These tools built the cognitive infrastructure for outcome-adjacent billing, the intuition that you're paying for a result rather than a license. By the time HubSpot announced $1.00 per qualified lead for Breeze agents, enterprise buyers already had a mental category for it. That familiarity is part of what makes the current pricing architecture so effective, and so worth examining carefully.

The Audit Problem: Who Decides What "Resolved" Means?

Start with the definitional question, because it is where the entire pricing model either holds or collapses. What is a "resolved" conversation? The plausible definitions are numerous and non-equivalent. A conversation might be resolved because the customer did not escalate to a human agent. Or because the customer closed the chat window. Or because the customer did not reopen a ticket within twenty-four hours. Or because the customer rated the interaction positively. Or because the underlying issue was actually fixed. These are different things. A customer who closes a chat window in frustration and calls the support phone line has not had their issue resolved. A customer who doesn't reopen a ticket because they gave up has not had their issue resolved. Vendors choose among these definitions, and the choice has direct financial consequences at scale.

Intercom's Fin model makes the stakes concrete. At $0.99 per resolution across two million weekly resolutions, the billing surface is substantial. A five percent definitional disagreement about what counts as resolved is not a rounding error. It is a material contract dispute repeating every week. Buyers have no independent instrumentation to audit this number. They receive a resolution count from Intercom's measurement system and an invoice derived from it. The only recourse is to trust the vendor's definition or negotiate a different one before signing, which requires knowing what questions to ask.

HubSpot's $1.00 per qualified lead for Breeze agents introduces a definitional problem that is, if anything, more severe. "Qualified" is even more contested than "resolved." Qualification criteria vary by sales team, by quarter, by shift in ideal customer profile, by whether the sales manager running the team this quarter has different standards than the one who ran it last quarter. The vendor's definition of qualified is baked into the billing engine. The buyer's definition lives in their CRM, in their sales manager's head, and in the informal norms of their sales floor. These definitions will diverge. The question is not whether disputes will arise but whether buyers have any structural mechanism to surface and resolve them.

Salesforce's "Agentic Work Unit" is the most abstract version of this problem. It is a proprietary metric with no external reference point, no industry standard it maps to, no independent definition that exists outside Salesforce's platform. When the unit of billing is invented by the vendor, the audit problem becomes structurally unsolvable without access to the vendor's internal measurement logic. You cannot verify an Agentic Work Unit count any more than you could verify a claim denominated in a currency only one party can print.

The most important clause in any AI agent contract is not the price per resolution. It is the definition of resolution, buried in the terms of service.

The lock-in mechanism that follows from this is worth stating explicitly. Once a company's support workflows, escalation logic, and success metrics are tuned to a vendor's definition of resolution, switching vendors means rebuilding those definitions from scratch against a new measurement system. The switching cost is not data migration in any conventional sense. It is epistemological: you have to reconstruct what "working" means in a different vendor's terms, retrain your team's intuitions, and accept a period of measurement discontinuity during which you cannot compare performance across the old and new systems. That is an enormous barrier, and it compounds over time as the vendor's definitions become more deeply embedded in operational practice.

This problem is not hypothetical, and it is not without historical precedent. The click fraud problem in digital advertising followed an identical structure. Buyers paid for outcomes — clicks, impressions, conversions — that vendors measured unilaterally. The resolution came only after years of third-party verification infrastructure was built: ad verification companies, independent measurement panels, eventually industry standards bodies. The AI agent billing market is at the pre-audit-infrastructure stage of that same curve. The infrastructure gap is real, it is consequential, and it will eventually be filled. The question is how much money flows through unaudited billing systems before it is.

The Margin Architecture Underneath the Fairness Narrative

The definitional problem is not the only structural issue with AI agent outcome-based pricing. Underneath the fairness narrative is a deliberate margin architecture that deserves to be understood on its own terms. At scale, vendors can set resolution definitions that maximize the count of billable events while minimizing the cost of producing them. A well-tuned deflection that closes a chat without human escalation costs the vendor very little in compute. It bills at $0.99 or $1.50. The incentive structure built into outcome-based pricing rewards vendors for producing large numbers of cheap deflections that meet the minimum threshold for billing, not for producing genuinely high-quality resolutions that address the underlying customer need.

The model-layer economics make this margin architecture more durable than it might first appear. Anthropic's per-seat plus API overage structure for enterprise contracts means that application-layer vendors — HubSpot, Zendesk, Intercom — are paying per-token to Anthropic and billing per-resolution to their customers. The spread between those two cost structures is the margin. As models get cheaper, which they consistently have, that spread widens without any change in the customer's bill. The customer signed a contract at $0.99 per resolution. The vendor's cost to produce that resolution declines over time as the underlying model improves and gets cheaper. The customer sees none of that efficiency gain unless they renegotiate.

Sierra AI's reported $150 million-plus in annual recurring revenue is evidence that the commercial model works. It is not evidence that enterprise buyers are satisfied with the pricing structure over the long term. High ARR at this stage of market development tells you about sales motion and early adoption. It does not tell you about customer retention rates, dispute rates over billing definitions, or whether enterprise procurement teams are beginning to push back on audit provisions. Those signals take longer to surface, and the market is not old enough to have produced them in volume.

Adobe CX Enterprise's credit-based approach adds a further layer of abstraction. Buyers purchase credits. Credits are consumed by agent actions. The conversion rate between credits and business results is opaque by design. This is outcome-based pricing with an additional indirection layer, which means the audit problem is compounded: buyers must first reconstruct which actions consumed which credits, then determine whether those actions produced the business outcomes they were paying for. The credit model is not a simplification. It is a complexity that benefits the party who controls the conversion logic.

The contrast with workflow automation tools is instructive here. Make and n8n have built operation-count pricing that gives buyers meaningful visibility into what they're consuming. An "operation" in these platforms is a discrete, auditable technical event: a module executed, an API call made, a data transformation completed. Buyers can count operations independently. They can instrument their own workflows to verify counts. The unit of billing is a technical fact, not a business judgment. The contrast with "resolution" is not incidental. It reflects a fundamentally different relationship between the vendor and the buyer's ability to verify what they're paying for.

What Buyers Should Actually Demand — and Which Tools Are Building It

The argument against outcome-based pricing is not that it is wrong in principle. It is that it requires independent observability infrastructure to be fair in practice. A pricing model tied to business outcomes is genuinely appealing, and it can be made to work, but only if buyers have the means to verify the outcomes they're paying for. Right now, most don't. The prescription follows from the diagnosis: buyers need to demand three specific things before signing any AI agent contract.

First, a contractual definition of the billable outcome with explicit exclusion criteria. Not "resolved conversation" as a phrase, but a written specification of exactly what events must occur, and what events must not occur, for a conversation to be counted as resolved. This definition should include how the vendor handles edge cases: conversations where the customer closes the window mid-session, conversations that receive no rating, conversations where the customer contacts a different channel within a defined window. Every exclusion criterion that is absent from the contract is a potential source of billing inflation.

Second, access to raw interaction logs sufficient to reconstruct billing calculations independently. This is a data access provision, and it needs to be negotiated before signing, because vendors have no incentive to offer it after the contract is in place. The logs need to include timestamps, conversation states, escalation events, and closure events in a format that can be ingested by the buyer's own analytics infrastructure.

Third, SLA language that ties disputed resolutions to a credits mechanism. When a buyer identifies a billing discrepancy, there needs to be a defined process for raising the dispute and a defined remedy if the dispute is upheld. Without this, the practical cost of pursuing a billing dispute is high enough that most buyers will absorb the discrepancy rather than fight it, which is exactly the equilibrium the vendor prefers.

Metabase, scored 8.2/10 by the TopReviewed AI panel, is relevant as a BI layer that buyers can use to build their own resolution dashboards against exported interaction data. But the observability gap is a data access problem before it is a tooling problem. Metabase can model and visualize whatever data you can get out of the vendor's system. If the vendor doesn't provide sufficient export data, the tooling doesn't matter. The negotiation for data access is the prerequisite.

The teams best positioned to audit AI agent billing are those who have already built data infrastructure. Companies using dbt for data transformation can model their own resolution definitions against raw conversation exports and compare the results to vendor invoices. This means defining resolution in SQL, running that definition against the exported conversation logs, and producing an independent resolution count that can be reconciled against the vendor's billing summary. This is not a workflow most support teams have. It requires collaboration between support operations and data engineering, and it requires the organizational will to treat AI agent billing with the same rigor applied to other significant vendor relationships.

An emerging category of AI workflow observability tools is beginning to address this gap directly. These tools sit between the agent layer and the billing layer, logging every agent action, every escalation decision, and every conversation close event independently of the vendor's measurement system. The category does not yet have dominant players. That absence is itself a signal about where the market is heading and where the next significant infrastructure opportunity lies. The vendors who build credible, buyer-side observability into their products — not as an add-on feature but as a core architectural commitment — will have a meaningful advantage as enterprise procurement teams become more sophisticated about AI contracts.

OpenAI's credit billing model for ChatGPT Workspace Agents, launched after the May 6 free preview ended, is worth examining as a partial case. Credits are at least denominated in a unit buyers can track. The consumption of credits can be logged. But the conversion from credit consumption to business outcome remains opaque: knowing that a conversation consumed a certain number of credits does not tell you whether that conversation produced the resolution you were paying for. The credit model is a step toward auditability compared to pure outcome billing. It is not a solution to the fundamental measurement problem.

The Vendor Who Solves Observability Wins the Decade

The AI agent outcome-based pricing war will not be won on price per resolution. The numbers across the major vendors are already close enough that price alone will not differentiate at enterprise scale. $0.99, $1.50, $2.00 — these are close enough that procurement decisions will turn on other factors. Trust is the factor that will matter most over time, and trust in this context means verifiability. Buyers will consolidate around vendors who can demonstrate, not just assert, that their billing reflects genuine outcomes.

The historical parallel that closes this argument is the cloud infrastructure market. For several years, buyers trusted AWS billing because they had no alternative. The bills were complex, the pricing dimensions were numerous, and the internal expertise to audit them didn't exist in most organizations. Then gradually, FinOps practices emerged. Third-party cost management tools appeared. Cloud cost optimization became a recognized discipline with dedicated tooling, dedicated headcount, and eventually industry standards. The AI agent market is at the "trust us" stage of that same curve. The FinOps equivalent for AI agents — call it AgentOps, or resolution auditing, or AI billing governance — is the infrastructure gap that represents the next significant opportunity in the enterprise software market.

The irony is that the vendors most likely to build buyer-side observability are not the large platforms with the most to lose from audit scrutiny. They are the workflow and integration tools that sit in the middle of agent workflows and can log everything independently of the outcome-measurement systems. The category that Make, n8n, and Microsoft Power Automate occupy is structurally positioned to provide independent logging of every agent action, every escalation, every conversation state transition, because these tools are already in the data path. They don't have a billing interest in the resolution count. They have an interest in being the infrastructure layer that enterprise buyers trust precisely because it is neutral.

Any enterprise evaluating an AI agent contract in 2026 should require three things in writing before signing: a definition of the billable outcome specific enough to implement as a database query, a data export provision sufficient to reconstruct that definition independently, and a dispute resolution clause with a defined credits remedy. If a vendor won't provide those three things, the pricing model is not aligned with your interests. It is aligned with theirs.