Business Model, Pricing & Governance

Why AI breaks the per-seat model

The per-seat subscription was the right pricing for a tool a human operates. A seat is a clean proxy for value when value rises with the number of people doing work inside the software. AI severs that link. Andreessen Horowitz puts the proposition flatly: “Per-seat is no longer the atomic unit of software,” because AI handles work previously done by humans, and value attaches to the work completed rather than the headcount logged in.³ Bain & Company, analysing more than thirty SaaS vendors that added generative-AI capabilities, reaches a more measured version of the same conclusion: “Seats may not be dead, but they are no longer the only game in town.”¹

Bain identifies two mechanisms doing the breaking. First, AI delivers value through automated background processes that need minimal human engagement, which decouples value from headcount. Second, AI introduces substantial new costs — model inference, fine-tuning, AI-specific R&D — that vendors must recover somehow.¹ That second mechanism is the one most strategy discussions underweight. Bessemer Venture Partners states the economics bluntly in its 2026 pricing playbook: “AI economics ≠ SaaS economics. Every AI query costs money (compute, inference, human-in-the-loop). You’ll see 50–60% gross margins vs. 80–90% for traditional SaaS.”² Microsoft, Bessemer notes, lost roughly $20 per user per month on GitHub Copilot at launch.² A model in which marginal cost is real changes pricing from a packaging decision into a margin-survival decision.

There is a measurement problem layered on top of the margin problem. McKinsey finds that only about 30% of software companies have published quantifiable ROI in dollar terms from real customer deployments, and warns that AI-enabling a full customer-service stack could imply a 60–80% increase in list prices.^44b Charging materially more for a benefit most vendors cannot yet quantify is the central commercial tension of the moment. McKinsey also flags a cost the deployment economics often miss: for every $1 spent on model development, firms should expect roughly $3 in change-management expenditure — training and performance monitoring.^44b One Fortune 100 HR executive captured the resulting paradox in the McKinsey work: “All of these copilots are supposed to make work more efficient with fewer people, but my business leaders are also saying they can’t reduce head count yet.”⁴

The pricing ladder: seats → consumption → outcomes

The cleanest framework in the corpus is Bessemer’s mapping of business model to charge metric. It splits AI products into three types and matches each to a pricing approach: copilots that assist humans price per seat or consumption, much like SaaS; agents that execute entire workflows autonomously price on outcome or workflow tied to tangible ROI; and AI-enabled services that blend automation with human oversight price from consumption toward outcome, benchmarked against the cost of a full-time employee.² The deeper point is that the charge metric is not a billing detail. In Bessemer’s words, “Your charge metric isn’t a billing decision — it’s a statement about what you believe your AI is worth and what you’re willing to stake your margins on to prove it.”²

The three rungs of the ladder trade off the same two variables in opposite directions. Consumption pricing — tokens, API calls — gives clean margins and predictable costs, but customers do not think in tokens, so it works mainly for technical buyers. Workflow pricing — per task completed — sits closer to how work actually happens. Outcome pricing — per result delivered — offers maximum value alignment but maximum cost risk to the vendor.² Bessemer’s summary of the trade is the through-line of this whole section: “As you move from consumption → workflow → outcome-based pricing, you accept more cost risk for tighter value alignment.”²

The same logic appears, sharpened, in the venture and consulting literature. a16z lists three pricing shifts AI forces: software becoming labour (service work rendered as scalable software), outcome-based pricing (charging for results, not users), and variable costs (foundation-model API calls scale with usage, breaking flat-fee logic).³ Its canonical worked example is Zendesk: customers historically paid roughly $115 per support-agent seat per month, but as AI handles more resolution, the natural pricing metric shifts toward “successful outcomes.”³ That single number — $115 per seat — is the thing the whole outcome-pricing movement is trying to replace.

Outcome pricing in the wild

The strongest evidence that the ladder is real and not theoretical is that named vendors are charging on its top rung. Bessemer calls Intercom’s $0.99 per resolved ticket “the gold standard” of outcome pricing.² Intercom’s official pricing for its Fin agent confirms it: $0.99 per outcome, where a resolution is defined as “no further help is requested after Fin’s last answer,” charged at most once per conversation regardless of how many actions the agent takes.⁶ The page now frames resolution as one of four billable outcome types, with prospect qualification priced higher at $9.99 — a tell that “outcome” is not one thing but a menu of differently-valued results.⁶

Salesforce is the most instructive case because it shows the model evolving under commercial pressure. Agentforce launched with “an initial $2 fee per agent conversation”⁵ — a conversation, as Salesforce defined the metric, running from the first agent response until the issue is resolved, closed, or inactive for 24 hours. Within months Salesforce added flexible models: Flex Credits at $0.10 per action (20 credits per action; 100,000 credits for $500), a Flex Agreement letting organisations convert user licences into credits and back, and per-user licensing — alongside pay-as-you-go and pre-commit payment options.⁷ The pivot was demand-driven; Salesforce’s own CIO research found that “90% of CIOs report that managing AI costs limits their ability to drive value.”⁷ Reading the two announcements together, the trajectory is not “seat to outcome” but “outcome back toward a consumption meter the buyer can predict.”

Zendesk supplies the most sophisticated mechanism. It announced in August 2024 that it was “first in CX industry to offer outcome-based pricing for AI agents,” with customers incurring costs “only for issues that are resolved autonomously by AI” and a starter usage level free.⁸ Its 2026 Relate launch added a verification layer that anticipates the central objection to outcome pricing: every charged resolution is “verified — both by the AI agent resolving the interaction end-to-end and independently confirmed by a dedicated AI evaluation model,” with spam and routine exchanges excluded.⁹ The per-resolution dollar figure — clustering around $1.50 — does not appear on Zendesk’s own pages and is documented only in third-party contract teardowns.^9b Harvey, the flagship legal vertical-AI vendor, is the instructive counter-case: it remains seat-anchored, using opaque enterprise per-seat pricing estimated at roughly $12,000–$16,800 per seat per year and positioning itself as “a labor cost substitute” priced at about 5–7% of associate labour cost — outcome logic in the pitch, seat logic in the invoice.²⁵

Vendor / product	Unit	Price	Source character
Intercom Fin	per resolution (outcome)	$0.99	Official (fin.ai)
Intercom Fin	per qualification outcome	$9.99	Official (fin.ai)
Salesforce Agentforce (original)	per conversation	$2.00	Press (Salesforce Ben)
Salesforce Agentforce Flex	per action	$0.10	Official via MarTech
Zendesk AI agent (committed)	per verified resolution	≈ $1.50	Third-party teardown
Zendesk legacy seat	per agent / month	$115	a16z (the model being disrupted)
Harvey (legal, base)	per seat / year	$12k–$14.4k	Third-party estimate

The measurement problem, and why outcomes get gamed

Outcome pricing has a structural flaw that gets worse the better the AI gets. Zendesk’s billing turns on a “Verified Resolution,” confirmed by a second language model that asks whether the AI’s reply actually solved the problem. As one teardown observes, this produces a counterintuitive dynamic: “a better-tuned AI agent costs more, not less,” because improved performance increases the verified resolutions that hit the bill.^9b The customer’s incentive to improve the product and the vendor’s incentive to bill for it point in opposite directions — a misalignment that pure seat pricing never created.

Attribution is the deeper problem. Lago, a billing-infrastructure vendor, argues outcome pricing is operationally fragile because outcomes are deceptively hard to measure even with full data access. Its illustration is precise: “If a user rolls their eyes, thinks ‘what a useless AI chatbot’ and slams their laptop shut in anger, they didn’t submit a ticket. Does the system count that as a resolution?“²⁴ Lago adds a boundedness problem — “a proactive, outcome-based AI SDR might get 500 meetings in a week, but if you only have 2 salespeople, that’s counter-productive” — and concludes that support tickets are a rare exception of clear, countable, uniform outcomes, while most professional work resists outcome billing because the outcomes vary qualitatively or cannot be metered at all.²⁴

Jason Lemkin of SaaStr supplies the durability counterpoint. His thesis is that proven, low-friction models win — payment processors charge per transaction, CRM vendors charge per seat — and that outcome pricing may be “the cart driving the horse.” He cites a SaaStr Fund portfolio company whose outcome-based deal crossing $1m a year reverted: “the customer quickly moved to a fixed contract.” And he notes that as AI costs in B2B SaaS fall toward zero, the economic rationale for outcome pricing weakens; major vendors charge “$1–$3 per outcome-based resolution,” yet HubSpot and Box have largely abandoned the approach.¹⁸ His closing line is the discipline this section needs: “A pricing model is not a product. And a pricing model doesn’t make a mediocre product great.”¹⁸

Convergence: hybrid platform + consumption

The market is not resolving to pure outcome pricing. It is resolving to a hybrid. Bain’s analysis of 30-plus AI-enabled SaaS vendors is the hard data point: roughly 35% raised per-seat prices while bundling AI into existing tiers (e.g. Zoom), and roughly 65% adopted hybrid models layering an AI usage or outcome meter on top of seat-based pricing (e.g. Adobe, Salesforce). Critically, 0% monetised AI purely as a standalone add-on, and 0% had fully transitioned to usage- or outcome-only pricing.¹ The pure-outcome endpoint that the venture commentary celebrates had, at the time of Bain’s study, zero incumbent adopters.

McKinsey reaches the same destination from the advisory side, recommending hybrid pricing that blends per-user subscription fees with consumption-based metrics — similar to Microsoft’s Copilot structure — and expecting many vendors to start with hybrid models in which consumption beyond a capacity cap is metered (e.g. limiting tokens processed per day, week, or month).⁴ Deloitte quantifies the trajectory: citing Gartner, it projects that “by 2030, at least 40% of enterprise SaaS spend will shift toward usage-, agent-, or outcome-based pricing,” and reports that 83% of AI-native SaaS companies already offer usage-based pricing.²¹ Note the framing: a 40% shift by 2030 implies a majority still anchored on seats or subscriptions — convergence on hybrid, not abandonment of the base. Bessemer’s prescription closes the loop: hybrid models — “base subscription + usage/outcome tiers” — “provide predictability while capturing upside,” with an illustrative structure of a platform fee set at roughly twice delivery cost plus outcome credits.²

There is a forward-looking economic argument that outcome and consumption pricing become more feasible, not less, as agents proliferate. An NBER chapter on the “Coasean Singularity” argues that AI agents “dramatically reduce transaction costs” by lowering “the costs of preference elicitation, contract enforcement, and identity verification,” which “expand[s] the feasible set of market designs.”¹⁷ Pricing tied to verifiable outcomes is exactly the kind of market mechanism that becomes cheaper to operate when verification is cheap — which is what Zendesk’s AI-verifier-on-AI-resolution architecture is, in microcosm. But the same paper warns the transition “also introduce[s] frictions such as congestion and price obfuscation” and “novel regulatory challenges,“¹⁷ which is the natural bridge to governance.

The governance stack: voluntary frameworks and binding law

As agents take on consequential decisions, three reference frameworks have become the de facto governance stack, and they differ on the one axis that matters most — whether they are binding. The U.S. NIST AI Risk Management Framework (AI RMF 1.0, January 2023) is voluntary and organised around four functions: Govern (organisational structures, policies, accountability), Map (identify systems, context, and risks), Measure (test and monitor performance and impact), and Manage (intervene and allocate resources to mitigate risk). Govern is cross-cutting and informs the other three.¹⁰ It names seven characteristics of trustworthy AI, including “Accountable and Transparent,” and was extended by a Generative AI Profile (NIST AI 600-1) in July 2024.¹⁰

ISO/IEC 42001:2023 is the world’s first international, certifiable AI management-system standard, built on a Plan-Do-Check-Act cycle and requiring AI impact assessment, lifecycle controls, and — notably — third-party supplier oversight.²⁶ Where NIST gives a risk vocabulary and ISO gives a certifiable management system, the EU AI Act (Regulation (EU) 2024/1689, in force 1 August 2024) supplies the binding law. It sorts systems into four risk tiers — unacceptable (banned), high, limited, and minimal — with high-risk systems in Annex III (which explicitly includes recruitment AI) subject to risk-management, documentation, data-governance, and human-oversight obligations, phased in over 6 to 36 months.¹¹ Penalties run up to €35 million or 7% of global annual turnover for prohibited-practice violations.¹¹ Underneath all three sit the OECD AI Principles (2019, updated 2024), the first intergovernmental AI standard, which establish accountability and traceability as named obligations: “AI actors should be accountable for the proper functioning of AI systems… based on their roles, the context, and consistent with the state of the art.”²³

Framework	Publisher / year	Core structure	Binding?	Key figure
AI RMF 1.0 (AI 100-1)	NIST · 2023	Govern · Map · Measure · Manage	Voluntary	7 trustworthiness traits; GenAI Profile 600-1 (2024)
EU AI Act (2024/1689)	European Union · 2024	4 risk tiers: unacceptable / high / limited / minimal	Binding law	Up to €35m or 7% of global turnover
ISO/IEC 42001:2023	ISO/IEC · 2023	Plan-Do-Check-Act AI management system	Voluntary (certifiable)	First international certifiable AIMS
OECD AI Principles	OECD · 2019 (am. 2024)	5 values-based principles + 5 recommendations	Soft law	First intergovernmental AI standard

The frameworks are converging on substance even as they differ on enforceability. Each centres accountability, traceability, risk management, and human oversight; ISO/IEC 42001 explicitly complements rather than replaces the NIST RMF and the EU AI Act.²⁶ And the activity is intensifying: Stanford’s 2025 AI Index records that in 2024 the OECD, EU, UN, and African Union all published RAI frameworks, even as the number of reported AI incidents hit a record 233 — a 56.4% jump over 2023.²⁷ The same Index flags the gap that frameworks alone do not close: a McKinsey survey found leaders cite inaccuracy, regulatory compliance, and cybersecurity as top RAI risks (64%, 63%, and 60% of respondents), yet “not all are taking active steps to address them,” and standardised RAI benchmarks remain sparse among major model developers.²⁷

You can’t outsource accountability

The governance principle that matters most for buyers is the one that survives any choice of framework: accountability does not transfer with the workload. IBM states it directly — asked who should be accountable for responsible AI outcomes, the three most common answers it hears are “no one,” “we don’t use AI,” and “everyone,” and “none of which are correct and all are concerning.” Its conclusion: “The principle that accountability cannot be outsourced remains critical; while AI can enhance efficiency and enable innovations, human oversight, judgment, and accountability must remain central.”¹⁶ IBM adds the procurement edge most organisations miss: “most software and cloud vendor contracts lack explicit commitments that make them accountable for providing responsible AI,” and some include disclaimers removing liability outright.¹⁶ The OECD encodes the same idea as the duty of each AI actor to be accountable “based on their roles,” supported by mandatory traceability of datasets, processes, and decisions across the lifecycle.²³

The hard legal form of this principle is being decided in Mobley v. Workday (N.D. Cal., No. 3:23-cv-00770-RFL). Lead plaintiff Derek Mobley submitted over 100 applications through employers using Workday’s screening system and was rejected each time, alleging the AI tools “unfairly penalize older candidates.”¹⁴ The pivotal move is the theory of liability. The EEOC’s April 2024 amicus brief argues that Workday — the vendor, not the employer — can be directly liable under Title VII, the ADA, and the ADEA under three theories: as an employment agency, as an indirect employer controlling access to opportunity, and as an agent of employers to whom hiring authority was delegated.²² The EEOC rejects the size defence outright: “there is no authority for the novel proposition that an entity can be too big to qualify as an employment agency… whether Workday performs those tasks for one employer or thousands of employers.”²²

The court has let the case advance and grow. In July 2024 it allowed the agent theory to proceed; in May 2025 it granted conditional nationwide certification of the ADEA collective, covering applicants aged 40 and older denied recommendations through Workday since 24 September 2020.¹³¹⁴ The scale is extraordinary: per Workday’s own filings, roughly 1.1 billion applications were rejected during the period, with the collective potentially numbering in the “hundreds of millions” — one of the largest collectives ever certified in employment litigation.¹³ A later order (filed 6 March 2026) preserved the ADEA disparate-impact claim while dismissing the related state-law FEHA claims as pled, noting that “attending a historically Black college can be used by AI screening tools as a proxy for race.”¹² Workday maintains the suit “lacks merit” and stresses the rulings are preliminary.¹⁴ The strategic lesson holds regardless of final outcome: deploying a vendor’s AI to perform a regulated function does not move the legal exposure off the deployer — and may pull the vendor in too.

Architecture and autonomy: the lever that ties it together

Pricing, margin, and liability all scale with the same hidden variable — how much autonomy the buyer cedes to the agent. The academic literature is converging on treating autonomy as a deliberate, governable design choice rather than a byproduct of capability. Feng, McDonald, and Zhang define a five-level spectrum by the user’s role — operator, collaborator, consultant, approver, observer — ranging from direct control to light supervision, and argue that “an agent’s level of autonomy can be treated as a deliberate design decision, separate from its capability and operational environment.”¹⁹ They propose “AI autonomy certificates” to govern agent behaviour, framing autonomy as “a double-edged sword.”¹⁹ This maps onto the familiar human-in-the-loop / human-on-the-loop / human-out-of-the-loop taxonomy: the operator and approver roles keep a human in or on the loop; the observer role approaches out-of-the-loop.

The risk gradient along that spectrum is the crux. Mitchell, Ghosh, Luccioni, and Pistilli of Hugging Face argue, from the title down, that fully autonomous agents should not be developed, on the principle that “risks to people increase with the autonomy of a system: The more control a user cedes to an AI agent, the more risks to people arise.”¹⁵ Their position favours constrained, semi-autonomous systems with human oversight over fully independent ones.¹⁵ The practical design implication is a trade-off between two viable postures: high autonomy on a narrow scope, where a tightly-bounded task can be safely automated end-to-end, versus broad scope at low autonomy, where wide-ranging action is permitted only behind human verification gates. The danger zone is the combination the frameworks all warn against — broad scope and high autonomy with no gate.

The autonomy lever is not abstract risk theory; it shows up in the pricing data. The cell where outcome pricing actually works — Intercom resolutions, Zendesk verified resolutions — is high-agency, narrow-scope: discrete, contained, measurable tasks, exactly Lago’s “rare exception” of clear, countable outcomes.²⁴ The cell that frightens regulators is high-agency, broad-scope. NBER work on AI-powered trading shows reinforcement-learning agents “autonomously sustain collusive supra-competitive profits without agreement, communication, or intent” — price-fixing cartels emerging with no instruction to collude, degrading market efficiency.²⁰ That is the empirical case for verification gates on any agent with both broad scope and pricing or transactional authority. The design recommendation that falls out of the whole corpus is consistent: pair high agency with narrow scope, or pair broad scope with low agency behind human-on-the-loop gates — and price, govern, and accept liability accordingly.

References

1. Maltiel, E., & Sandberg, J. (2025). Per-Seat Software Pricing Isn’t Dead, but New Models Are Gaining Steam. Bain & Company. Accessed 2026-06-16.
2. Bessemer Venture Partners (2026). The AI Pricing Playbook for Founders. Bessemer Venture Partners. Accessed 2026-06-16.
3. a16z Enterprise team (2024). AI Is Driving a Shift Towards Outcome-Based Pricing (December 2024 Enterprise Newsletter). Andreessen Horowitz. Accessed 2026-06-16.
4. McKinsey & Company (2025). Upgrading Software Business Models to Thrive in the AI Era. McKinsey & Company. Accessed 2026-06-16 (figures verified by corroboration).
5. The Register (2025). McKinsey wonders how to sell AI apps with no measurable benefits. The Register. Accessed 2026-06-16 (independent corroboration of the McKinsey figures).
6. Salesforce Ben editorial (2024). Salesforce’s Bold New Pricing Strategy: What You Need to Know. Salesforce Ben. Accessed 2026-06-16.
7. Intercom / Fin (2026). Fin AI Agent Pricing. Intercom (fin.ai). Accessed 2026-06-16.
8. MarTech (2025), reporting Salesforce’s May 2025 announcement. Salesforce Introduces New Agentforce Pricing Models (Flex Credits). MarTech.org. Accessed 2026-06-16.
9. Zendesk (2024). Zendesk First in CX to Offer Outcome-Based Pricing for AI Agents. Zendesk newsroom. Accessed 2026-06-16.
10. Zendesk (2026). Zendesk Introduces the Autonomous Service Workforce (Relate 2026). Zendesk newsroom. Accessed 2026-06-16 (verification mechanism; no dollar figure on page).
11. eesel AI (2026). Understanding Zendesk AI Pricing: A Complete Pay-per-Resolution Guide. eesel AI. Accessed 2026-06-16 (third-party teardown; sources the ~$1.50 figure).
12. NIST — Tabassi, E., et al. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. U.S. Dept. of Commerce / NIST. Accessed 2026-06-16.
13. European Union (2024). EU AI Act — Regulation (EU) 2024/1689 (high-level summary). artificialintelligenceact.eu (Future of Life Institute). Accessed 2026-06-16.
14. U.S. District Court, N.D. Cal. (2026). Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL — Court Order (Doc. 267, filed 2026-03-06). govinfo.gov. Accessed 2026-06-16.
15. Proskauer Rose LLP (2025). AI Bias Lawsuit Against Workday Reaches Next Stage (Conditional ADEA Certification). Proskauer Rose LLP. Accessed 2026-06-16.
16. Holland & Knight LLP (2025). Federal Court Allows Collective Action Over Alleged AI Hiring Bias. Holland & Knight. Accessed 2026-06-16.
17. Mitchell, M., Ghosh, A., Luccioni, A. S., & Pistilli, G. (2025). Fully Autonomous AI Agents Should Not Be Developed. arXiv:2502.02649 (Hugging Face). Accessed 2026-06-16.
18. IBM (2025). Who Is Accountable for Responsible AI? IBM Think Insights. Accessed 2026-06-16.
19. Shahidi, P., Rusak, G., Manning, B. S., Fradkin, A., & Horton, J. J. (2025). The Coasean Singularity? Demand, Supply, and Market Design with AI Agents. NBER. Accessed 2026-06-16.
20. Lemkin, J. (2025). The Real Question for Outcome-Based Pricing Is If It Will Stick. SaaStr. Accessed 2026-06-16.
21. Feng, K. J. K., McDonald, D. W., & Zhang, A. X. (2025). Levels of Autonomy for AI Agents. arXiv:2506.12469. Accessed 2026-06-16.
22. Dou, W. W., Goldstein, I., & Ji, Y. (2025). AI-Powered Trading, Algorithmic Collusion, and Price Efficiency. NBER Working Paper 34054. Accessed 2026-06-16.
23. Deloitte (2026). SaaS Meets AI Agents (TMT Predictions 2026). Deloitte Insights. Accessed 2026-06-16.
24. U.S. Equal Employment Opportunity Commission (2024). Brief of EEOC as Amicus Curiae in Support of Plaintiff, Mobley v. Workday, Inc. EEOC. Accessed 2026-06-16.
25. OECD (2019, amended 2024). OECD AI Principles (incl. Accountability). OECD. Accessed 2026-06-16.
26. Lago team (2025). Outcome-Based Pricing Is NOT the Future. Lago (Substack). Accessed 2026-06-16.
27. Metronome (2025). Harvey AI Pricing Index. Metronome. Accessed 2026-06-16 (pricing = third-party estimate; Harvey publishes none).
28. ISO/IEC JTC 1/SC 42 (2023). ISO/IEC 42001:2023 — Artificial Intelligence Management System. ISO. Accessed 2026-06-16.
29. Stanford HAI (2025). 2025 AI Index Report, Chapter 3: Responsible AI. Stanford Institute for Human-Centered AI. Accessed 2026-06-16.