Published research

Is enterprise AI moving toward pay-per-outcome pricing?

Generated Sep 24, 2026 · EXPERT Research · 17 sources · 2918 words

Summary

Enterprise AI has credible pay-per-outcome offerings, especially in support and verified payment recovery, but seats, tokens, actions and conversations remain prominent. OpenAI’s reported 2026 experiment is limited and its public API remains consumption-priced. Under pure outcome contracts, vendors generally bear the compute cost of failed attempts; profitability therefore depends on success, retries, unit cost and an auditable definition of completion. Hybrid pricing is more likely than a near-term, wholesale replacement of SaaS subscriptions. (theinformation.com)

Is enterprise AI moving toward pay-per-outcome pricing?

Review period: June 17, 2024–September 24, 2026 Scope: Enterprise AI pricing models, their billing definitions, unit economics, failure costs, suitable workloads, and implications for SaaS. Source policy: Prioritize current vendor terms and original statements, official pricing documentation, original research, and primary reporting; distinguish advertised prices from contract terms and illustrative models from observed results.

Executive summary

Yes, but selectively—not as a wholesale replacement for seats or tokens. Intercom, Zendesk, Gorgias, Sierra and, since June 2026, a specific Salesforce Agentforce self-service offering charge for defined results. Yet Salesforce also sells actions, conversations and seats; Ada says most of its customers choose conversation pricing; and OpenAI’s generally published API prices remain token- and tool-based. OpenAI’s reported pay-on-success arrangements concern selected large customers, with no public outcome rate or contract definition. (intercom.com)

The decisive economic difference is who pays for unsuccessful work. Under a pure per-success contract, the application vendor receives no outcome fee for a failed task but still incurs its model, search, tool, hosting and verification costs; it ordinarily continues to owe a token-billing model supplier. That makes success rate, retries, task selection and the contractual meaning of “success” determinants of margin. The scenario analysis below shows why a $0.99 resolution can be attractive for cheap, reliable work but uneconomic for expensive, failure-prone work. These are Revoqra-derived illustrations, not vendor margin estimates. (intercom.com)

What has—and has not—changed at OpenAI

In January 2026, CFO Sarah Friar identified outcome-based pricing, alongside licensing and intellectual-property agreements, as a possible future way to share in value created. In July 2026, she advocated measuring cost per successful task and useful work per dollar. Neither statement says that OpenAI’s standard API is billed per successful task: one addresses prospective business models; the other addresses how customers should assess economics. (openai.com)

The Information subsequently reported, rather than OpenAI publicly documented, that some major customers had been offered payment conditional on AI completing tasks such as support interactions. The accessible original report does not disclose customers, prices, success tests, failed-run treatment beyond the reported payment condition, or how broadly the option is available. Treat this as evidence of a limited experiment, not an announced migration of enterprise contracts. (theinformation.com)

OpenAI’s public pricing continues to demonstrate the distinction between optimizing for outcomes and billing for outcomes. Its API rate card charges by input/output tokens and separately for tools such as web search; its September 2026 GPT-6 Sol and Luna additions are likewise listed by token. Separately, OpenAI introduced pay-as-you-go Codex seats in 2026: removing a fixed seat charge still does not make a successful coding task the billing unit. (developers.openai.com)

Comparing the actual billing units

Offering, as documented by September 24, 2026	Charge trigger and published indication	Classification and qualification
Intercom Fin	$0.99 for a qualifying resolution, configured procedure handoff or disqualification; $9.99 for a defined lead qualification. At most one outcome is charged per conversation, regardless of actions. An unanswered attempt or failed procedure is not an outcome charge. (intercom.com)	Conditional-outcome billing, but not every named outcome is an end-to-end solved problem: a configured handoff is billable, and a qualified lead means matched criteria and routed—not a won sale. A resolution can be assumed when a customer leaves without asking for more help; a later return to the same conversation reverses that resolution. Fin for Sales also requires an Intercom plan. (intercom.com)
Zendesk AI agents	Advertises automated resolutions as low as $1.50 each, alongside paid per-agent Suite/Support plans and included resolution allowances. Zendesk says the agent must resolve the request without human escalation; it describes an independent AI evaluation model as a second check on charged resolutions. (zendesk.com)	Resolution-based AI charge in a hybrid subscription. “As low as” is not every customer’s contracted rate. Zendesk documents differing legacy and evolved resolution arrangements; its evaluator is a Zendesk-operated check, not proof that every end customer agrees the underlying problem was fixed. (zendesk.com)
Salesforce Agentforce, general options	$500 per 100,000 Flex Credits and generally 20 credits per standard action—an implied $0.10/action at that list rate—or $2/conversation; employee access also has user-based options. (salesforce.com)	Action, conversation or seat pricing—not pay-on-success. An action such as updating a record or answering a question is an executed function, not necessarily a completed customer request. (salesforce.com)
Salesforce Agentforce Help Agent, separate self-service option	$2 per resolution, with Agentforce actions and Data 360 queries advertised as unmetered within that resolution. The resolution-based add-ons became available June 19, 2026. Salesforce says qualifying sessions need at least two user–agent turns, no explicitly negative final feedback and either no escalation or explicitly positive final feedback; unresolved sessions are not billed. (salesforce.com)	Genuine conditional-resolution billing, subject to a proxy test. No negative feedback and no escalation can qualify even without explicit confirmation; a continuing interaction past the specified two-hour window can count as another resolution. This option must not be confused with Agentforce’s $2 per conversation offering. (help.salesforce.com)
Gorgias AI Agent	Advertises about $0.90 per AI-resolved interaction on most plans, with plan allotments and overages. Its billing documentation says a separate helpdesk ticket fee can apply to any ticket receiving a helpdesk message, including one the AI later hands to a person. (gorgias.com)	Hybrid: no automation fee on an unresolved handoff, but “no charge when AI fails” would be inaccurate for the total bill. (docs.gorgias.com)
Sierra; Ada	Sierra says it is paid when it completes contracted tasks and usually not for unresolved conversations; public examples include resolutions and other agreed outcomes, without a universal public unit price. Ada says its usual model charges for every conversation, while it also offers resolution-based pricing for enterprises with specific needs. (sierra.ai)	Sierra is an outcome-pricing example with negotiated definitions; Ada is a useful counterexample to claims that the whole agent market has moved to resolutions. (sierra.ai)
Stripe Smart Disputes; Chargeflow Automation	Stripe’s AI-assisted Smart Disputes charges 30% of the disputed amount only on wins, but the merchant’s separate dispute-received fee still applies. Chargeflow advertises 25% of recovered chargeback amounts for its automation product, while separately selling per-deflection and per-scanned-transaction products. (stripe.com)	A financially verified recovery is a stronger outcome than a chat ending. Neither product makes all merchant or vendor costs success-contingent; distinguish the success-fee SKU from adjacent activity fees. (stripe.com)

Interpretive rule: Per conversation bills contact, whether solved or not; per token bills compute; per action bills a step; per completed task approaches outcome pricing only if completion includes independently checkable acceptance criteria and unsuccessful attempts do not trigger the task fee. A posted, policy-compliant refund is closer to an outcome than a generated refund recommendation. Stripe’s pricing framework likewise separates workflow/task units from successful business outcomes. (stripe.com)

Economics of failure and retries

Cost bearer. For an application vendor buying metered models, every attempted inference and reasoning continuation can consume billable tokens; web search, file search and containers can add charges. If its customer owes only a successful-outcome fee, the application vendor funds failed attempts and retries, unless its supplier separately agrees to assume that risk. A vendor running its own models instead bears the associated infrastructure costs directly. Customers can still bear implementation charges, base subscriptions, third-party transaction fees, human escalation costs and the harm of a wrong answer. These are contractual distinctions, not a claim that any named vendor has a particular unpublished supplier agreement. (developers.openai.com)

Why token price understates cost. At OpenAI’s published standard, short-context GPT-6 rates, an illustrative attempt using 8,000 input and 2,000 output tokens costs approximately $0.0018 on Luna, $0.036 on Sol, or $0.18 on Astra for those tokens. Three separately charged standard web-search calls add $0.03, before search-content tokens, hosting or verification. These are Revoqra-derived price-card calculations, not measured costs of a deployed agent or OpenAI’s own inference cost. (developers.openai.com)

A reproducible margin model

For one eligible submitted task, define s as the fraction that meet the contract’s billable success test; r as the average additional attempts per submitted task (so attempts average 1 + r); c as the all-in variable delivery cost per attempt; h as other delivery cost per submitted task; v as verification or other delivery cost per billable success; and P as the fee per success. All scenarios below are hypothetical, with no observed sample (N: not applicable). The Revoqra-derived delivery-margin proxy is:

Revenue/task = sP; cost/task = h + (1 + r)c + sv; margin = 1 − [h + (1 + r)c + sv]/sP.

To attain a 70% illustrative target margin, s must be at least [h + (1 + r)c]/[0.30P − v], provided the denominator is positive. At a specified s, the maximum affordable c is [s(0.30P − v) − h]/(1 + r).

This is a delivery-cost proxy, not a reported or fully allocated GAAP gross margin: engineering, support, depreciation and other costs omitted from h would lower actual gross margin. Seventy percent is an analytical hurdle, not a measured industry requirement. Prices near published resolution fees provide context; the inputs do not purport to describe those vendors’ costs or success rates. (intercom.com)

Illustrative workload	P/success	Success s	Extra attempts r	c/attempt	Revenue/task	Cost/task	Modeled margin	Minimum s for 70% margin
Efficient support	$0.99	90%	0.10	$0.04	$0.891	$0.091	89.8%	24.0%
Mid-cost support	$0.99	65%	0.35	$0.12	$0.644	$0.202	68.7%	68.2%
Difficult low-price support	$1.50	40%	0.80	$0.25	$0.600	$0.482	19.7%	111.9%—impossible
Higher-price, reliable support	$2.00	80%	0.50	$0.25	$1.600	$0.419	73.8%	69.3%
Same $2 price, lower success	$2.00	40%	0.80	$0.25	$0.800	$0.482	39.8%	82.5%
Expensive, higher-value task	$10.00	60%	0.75	$1.25	$6.000	$2.408	59.9%	81.7%

Method: Revoqra-derived arithmetic from the stated equations, rounded for display. The first five rows assume h = $0.02/task and v = $0.03/success; the last assumes h = $0.10/task and v = $0.20/success. r is a mean count, not the probability of retrying. At $0.99, 65% success and 0.35 extra attempts, achieving 70% margin requires c ≤ approximately $0.114/attempt, rather than the illustrated $0.12. At $1.50 and 40% success with 0.80 extra attempts, it requires c ≤ approximately $0.082/attempt. These are thresholds under stated assumptions, not empirical forecasts. Public AI and tool rate cards motivate varying c, but do not establish any vendor’s internal costs. (developers.openai.com)

The model also exposes an incentive problem: a vendor can improve measured s by selecting easier tickets, calling silence a resolution, narrowing eligibility or changing when it escalates. Some selection is prudent cost control; it becomes misleading if contracts imply coverage of all incoming work. Buyers should report eligible-volume share, independently checked success, reopen/reversal rates, cost per truly resolved task and human rework, not just the vendor’s billable-resolution rate. Intercom’s assumed-resolution and Salesforce’s no-negative-feedback provisions show why billing tests and actual customer value must be assessed separately. (intercom.com)

Customer ROI and revenue predictability

Revoqra-derived customer illustration: Suppose 10,000 eligible support requests, a hypothetical $4 avoidable human handling cost per request, 65% billed AI resolutions, and $0.99 per billed resolution. If all billed resolutions really avoid human work and every failure goes to a human, the AI fee is $6,435, avoided human cost $26,000, and net savings $19,565 before integration or supervision. If only 90% of billed resolutions truly avoid human handling, net savings fall to $16,965; the invoice is unchanged. The $4 is an assumption, not a sourced labor benchmark. Thus an outcome-denominated bill improves auditability but does not itself establish incremental ROI or quality. (intercom.com)

For the seller, per-success revenue varies with eligible volume and acceptance rate; costs can rise as retries increase even while billings fall. Annual prepaid outcome buckets, minimum commitments, base fees and caps can steady cash flow or limit exposure, but make the deal less purely pay-on-success. Intercom has introduced annual resolution buckets; Deloitte describes prepaid-success, pay-as-you-go and hybrid accounting examples, which should not be mistaken for evidence that its hypothetical contract prices are market quotes. Recognition and unused-credit treatment depend on the actual agreement. (stripe.com)

Where an outcome can be contracted—and where it cannot yet be trusted easily

Workload	Stronger, auditable payment event	Principal attribution or quality trap
Support and IT service	Specified issue completed without human takeover, confirmation or a defined no-reopen interval; verify underlying account or ticket state for transactional requests.	Silence, ticket closure or a short no-escalation window may not mean the problem is fixed; exclude spam and duplicate contacts. Intercom and Salesforce explicitly use conversational signals in their billing definitions. (intercom.com)
Payments recovery and disputes	Cleared recovery, issuer-confirmed dispute win, or funds returned within an agreed window.	Ordinary payments would sometimes occur without the agent; refunds, reversals and separate processor fees change net value. Stripe and Chargeflow’s win-linked products provide real commercial examples. (stripe.com)
Lead qualification	Deduplicated prospect satisfies buyer-defined criteria and is accepted by sales; later commercial results require their own attribution rule.	Routing or booking is measurable but is not revenue won. Intercom charges on matching criteria and routing, demonstrating both the opportunity and the proxy risk. (intercom.com)
Document processing and transaction workflows	Correct invoice posted to the ERP, approved under policy and not reversed within an agreed period; or verified refund/payment execution.	Extracting fields or starting a workflow is not straight-through completion. Deloitte uses these stronger acceptance tests in illustrative accounting scenarios, not as reports of named vendors’ deployments. (dart.deloitte.com)
Coding	Accepted change plus agreed automated tests, security checks and human review, potentially followed by a defect-free period.	Passing tests or merging a pull request cannot prove maintainability or absence of later defects; substantial work can also produce a useful rejected patch. This is an analytical assessment of possible contract tests, not evidence of an established per-merged-PR market. (stripe.com)

Least suitable for simple per-task fees: open-ended strategy, research quality, creative work, managerial advice, diagnosis, and claimed incremental sales lift. A deliverable can be counted, but the valuable result may be subjective, delayed, jointly produced or dependent on an unknowable counterfactual. A scoped milestone, buyer acceptance, independently audited lift or a base fee plus incentive may be more defensible than claiming each generated answer is an outcome. This is Revoqra’s assessment, applying the measurement and attribution requirements described by Stripe and Zendesk. (stripe.com)

Could this reshape SaaS?

Model	What the buyer purchases	Dominant incentive and risk allocation
Seat subscription	Access over time, often with predictable committed spend.	Vendor benefits from retained seats; buyer carries adoption and much execution risk. Automation that removes paid seats can cannibalize the incumbent’s revenue. (salesforce.com)
Token/consumption or action/conversation	Compute, executed steps or interactions.	Vendor revenue tracks metered activity—including many unsuccessful attempts—while the buyer manages variable bills and business effectiveness. Metered actions can be easier to price than verified results. (developers.openai.com)
Conditional outcome	A contractually accepted result.	Vendor takes more delivery and retry risk and has stronger reason to improve cost per accepted result; buyer retains definition, oversight, integration and downstream business risks. Measurement disputes and volatile revenue become more consequential. (dart.deloitte.com)

Cannibalization illustration, not a company forecast: 100 seats at a hypothetical $100/month produce $120,000/year. If automation leaves 60 seats, seat revenue is $72,000; an added $2 per accepted outcome would require 24,000 annual outcomes merely to restore the original revenue, before outcome-delivery costs. At 50,000 accepted outcomes, combined revenue would be $172,000, but higher revenue alone says nothing about gross profit. The coexistence of Salesforce seat, action, conversation and resolution offerings shows why hybrid portfolios, rather than an immediate elimination of seats, are plausible. These figures are Revoqra-derived assumptions, not Salesforce customer data. (salesforce.com)

Bottom line—Revoqra synthesis: Outcome pricing could become a major incremental enterprise-software model if agents reliably finish economically valuable work; success can be independently checked and attributed; retries and expensive cases can be bounded or priced by complexity; contracts allocate reopens, errors and third-party fees; and vendors can finance variable revenue without hiding failure costs in mandatory subscriptions. Support resolutions, verified recoveries and tightly specified transaction workflows already offer working examples. Current coexistence with seat and consumption pricing—and Ada’s stated preference among its customers for conversations—does not establish that the broader SaaS market has crossed that threshold. (intercom.com)

Sources and evidence limits

Original positions and prices: OpenAI CFO statement, scorecard and API rate card; Intercom’s billing definitions; Salesforce’s Agentforce and Help Agent pricing and billing rules; Zendesk’s pricing and product announcement. (openai.com)
Additional first-party comparisons: Gorgias billing documentation, Sierra and Ada pricing descriptions, and Stripe and Chargeflow dispute-product terms. Vendor descriptions document offers and rules, not independently audited resolution rates or gross margins. (docs.gorgias.com)
Interpretation and corroboration: The Information’s original reporting on OpenAI’s selective experiment; Stripe’s pricing framework; Deloitte’s illustrative revenue-accounting analysis; and the original τ-bench research on the difficulty of reliable tool-using agents. Public sources reviewed do not establish the prevalence, negotiated price or realized vendor margins of outcome contracts across enterprise AI. (theinformation.com)