The AI pricing shift and what it means for your business

AI companies are discovering their current pricing models are not scalable – the fixed subscription cost model doesn’t cover costs, and the token pricing model has an upper limit on what clients are willing to pay. And on top of that, the margins stay thin, even as revenue grows exponentially. So, they are trying something new: outcome-based AI pricing, or charging only when the AI actually delivers and weaving themselves into the software people already use. It’s a clever fix commercially. However, it also raises a question regulators are already asking: who’s responsible when things go wrong, and who’s really controlling access to your data?

The numbers behind the shift: Leading AI Labs

Foundation-model providers disclose surprisingly thin, and in some cases shrinking, margins on the business of simply selling model access:

OpenAI’s adjusted gross margin fell from 40% in 2024 to 33% in 2025, as inference costs quadrupled to over USD 8 billion – even as revenue reached USD 13 billion, ahead of its own USD 10 billion target.^[1]
Anthropic cut its 2025 gross margin guidance from 50% to 40% after inference costs running on Amazon and Google servers came in 23% above budget, even as revenue is on track to hit roughly USD 4.5 billion – four times 2024’s USD 1 billion.^[2]
Z.ai’s Open Platform and API business grew gross profit from RMB 1.6 million (USD 238,400) in 2024 to RMB 36.0 million (USD 5.364 million) in 2025, and gross margin climbed from 3.3% to 18.9%, helped by growth in its cloud deployment business, the launch of coding-focused subscription packages, and improved inference efficiency.^[3]
MiniMax’s 2025 revenue grew to USD 79 million, while the gross profit margin increased from 12.2% in 2024 to 25.4% in 2025, but only USD 26 million of that came from selling Open Platform and other AI-based enterprise services directly – most of the growth came from MiniMax’s own consumer apps.^[4]

A caveat on comparability. These four margins aren’t directly comparable, since each company classifies its business differently, and not all the statistics strictly reflect the pure per-token economics of the company. Even so, all four point the same way: margins on token-based AI products are thin and under pressure.

Even at scale, and even among the best-funded labs, charging for raw model access looks more like a services business (18-40% margins) than a software business (SaaS leaders have historically run above 60%).^[5]Customer concentration adds to that fragility: financial operations platform Ramp found that 80% of OpenAI’s and Anthropic’s enterprise revenue comes from just 1% of their customers, a concentration unseen in any other software category Ramp tracks, and it isn’t improving as more businesses adopt AI.^[6] This gap is what’s pushing the AI vendors toward different pricing and offering models.

How the AI providers are responding: underwrite the result, not the compute

The idea

The logic is simple: instead of billing for tokens processed, bill only when the AI finishes the job successfully. OpenAI has begun offering some large customers outcome-based pricing contracts.^[7]Sierra and Fin, two customer-service AI vendors (Salesforce is acquiring Fin for USD 3.6 billion), already charge only when a case is resolved without a human stepping in.^[8] Cognition, a coding-agent company, has gone further still, publicly guaranteeing up to USD 10 million in credits if the engineering value its agent Devin delivers falls short of what a client paid for.^[9]

To see the mechanics, take a simple example.

Suppose an AI voice agent places 1,000 outbound sales calls, each involving roughly 3,000 tokens of conversation (say 1,800 input tokens of context and instructions, 1,200 output tokens of generated speech). Using Anthropic’s published list price for Claude Sonnet 4.6 – USD 3 per million input tokens, USD 15 per million output tokens – the token bill for all 1,000 calls comes to roughly USD 23.4 (about USD 5.4 for 1.8 million input tokens, USD 18 for 1.2 million output tokens).^[10]Crucially, that USD 23 bill is identical whether the calls produce zero sales or fifty.

Under an outcome-based contract charging, say, USD 50 per qualified lead generated, the same underlying compute could instead generate USD 2,500 in revenue if 50 of the 1,000 calls convert – or USD 250 if only 5 do.

In this example, the vendor’s revenue can potentially swing by 100x on the same compute cost, which is exactly why outcome pricing is commercially attractive. But that same math shows outcome-based pricing only works where success rates are already high: every attempt burns compute regardless of outcome, so a low success rate leaves too few paying outcomes to cover the failures. That’s why early deals cluster around narrow tasks with a clean success signal, such as a ticket resolved, a call converted, or a coding task benchmarked against engineer-hours.

This produces two consequences for how outcome-based pricing plays out:

It splits the market by task difficulty. Agents spanning multiple internal systems fare worse on this math: Stanford’s AI Index put real-world computer-use success at 66% in 2026, still a third failing, and 89% of enterprise agents never reach production.^[11] Few vendors will price a fixed outcome fee against that failure rate, so the harder, multi-system automation problems are likely to stay on ordinary token or subscription pricing for now.
It turns the definition of “success” into a contractual, and eventually legal, question. Payments provider Stripe has already published guidance warning vendors to define attribution rules up front, because a sales increase may stem from product adjustments, marketing campaigns, or seasonal factors rather than the AI itself, leaving room for disputes over whether a result should be credited to the software at all.^[12]

How the software providers are responding: don’t be just a tool, become the interface

While AI providers are re-calibrating pricing models to escape thin margins, the mirror problem is faced by the other side: what happens to a software company when its customers stop needing to open its app at all?

The idea

If a customer’s employees stop opening software because an AI agent now does the work for them, the software company’s traditional revenue base, the per-seat subscription, erodes right along with it. A smart response is not to resist that shift, but to make the agent-to-data connection itself the thing that gets monetized.

A clear example of this strategy in action is ServiceNow. In 2026 it launched Action Fabric, opening its enterprise workflows to external AI agents so they can execute governed actions directly on the ServiceNow platform, with Anthropic’s Claude as the launch partner.^[13] Rather than blocking outside agents, ServiceNow lets them in and meters what they do: every action an agent completes is reportedly tracked as an “assist,” the unit ServiceNow uses to measure AI usage. Third-party analyses of its licensing model indicate that “assists” beyond a subscription’s allotment are billed separately, though ServiceNow does not publish that pricing itself.^[14]

Separately, ServiceNow retired its legacy per-seat tiers in favor of three tiers, Foundation, Advanced, and Prime, mapped to AI maturity rather than just features: only Prime unlocks full autonomous-agent capability.^[15] ServiceNow, through its Action Fabric partnerships, pairs ServiceNow’s control over enterprise workflow data with an outside model provider’s technology, each capturing value from a different layer of the same arrangement.

In plain terms, the login screen is disappearing, but the toll booth isn’t – it has moved one level down, to the subscription tier that decides whether an outside AI is allowed to touch the data. That is a tying arrangement (buy the higher-priced plan to let your AI agent talk to your own CRM records), and it might fall into the range of bundling that competition regulators have spent two decades scrutinizing in other digital markets, i.e. self-preferencing, restricted interoperability, and gatekeeping over access to data that a customer owns.

Where this becomes a regulatory question

The commercial fixes above create two distinct governance problems that regulators in multiple jurisdictions are already examining separately:

Liability and attribution. When an autonomous agent is paid only for results, and the agent’s actions span several vendors (a model provider, a tooling layer, a platform, the enterprise deploying it), proving what went wrong and who is responsible becomes materially harder than under a simple software license. Singapore’s IMDA, in a May 2026 discussion paper, mapped a nine-actor AI value chain and concluded that existing contract, negligence, and product-liability doctrines can likely address many disputes in principle, but claimants will face real practical difficulty proving attribution, foreseeability, and causation in practice.^[16] It is one of several parallel efforts: the EU’s revised Product Liability Directive extends strict liability to AI-enabled products from December 2026,^[17] the UK’s Competition and Markets Authority published guidance applying existing consumer-protection law to agentic AI in March 2026,^[18] and the US FTC has signaled it will treat unsubstantiated AI performance claims as an unfair or deceptive practice under its existing authority.^[19]
Competition and bundling. Arrangements with AI-native tiered pricing raise the same interoperability and self-preferencing questions that competition authorities have already applied to app stores and search – specifically, whether a dominant enterprise platform can condition third-party AI access to a customer’s own data on the customer buying a more expensive subscription.

Neither of these is a hypothetical future concern: both are visible in contracts and product launches happening now, which is why they are likely to draw regulatory attention on a similar timeline to the commercial rollout itself, rather than years behind it.

On top of these regulatory concerns, there’s the added question of trust. As verification itself gets automated, with AI systems deployed to check whether another AI’s task counts as “done”, the confirmation of success becomes one more machine-generated claim rather than a human judgment. That raises its own legal, contractual, and ethical question: whether a “completed” label produced at a scale no human team can independently review is a fact anyone can actually stand behind.

How companies can get ahead

Companies on either side of these arrangements have a practical window to get ahead of the issues above before a dispute or a regulatory inquiry forces the question. Several things are worth doing now:

Get attribution and failure definitions in writing before signing, rather than after a dispute. If a contract charges by outcome, it should already specify what counts as a “resolved” ticket, a “qualified” lead, or a “delivered” engineering result, and how responsibility is split if a multi-vendor agent chain is involved.
Map own position on the AI value chain. Whether the company is a model developer, a tooling or platform provider, a deployer, or an end user, it is important to understand what the company is exposed to (IMDA’s stakeholder responsibility framework is a useful template for doing this exercise internally).
Normalize independent checks on AI performance. Just as companies get their finances audited, they may soon need someone independent to verify that an AI actually did what it was paid for, and getting ahead of that now is a chance to help set the standard rather than wait for regulators to impose one.
Recognize that today’s contracts are tomorrow’s regulatory template. Harder tasks might slowly move to outcome-based pricing as confidence builds, so the definitions of “success” and “responsibility” being written into today’s easier contracts are likely to harden into industry norms, and eventually regulatory reference points. Whoever defines those terms first has outsized influence later, which makes this an opening for companies to shape the terms proactively, not just a risk to manage.
Model the cash-flow gap before signing up for outcome pricing. A client has to pay for compute as soon as it runs, while outcome-based revenue from clients might only realize weeks or months later upon confirmation, so the AI service providers might end up fronting their own AI costs out of pocket in the meantime.
Scrutinize bundling terms in platform partnerships before they become dependencies. If a partner’s AI agent requires a higher subscription tier to access your own data, negotiate those terms now, while you still have leverage, rather than discovering later that a pricing line item has quietly become a structural dependency you

[1] Reported by news; neither company has confirmed these figures publicly, https://www.humai.blog/openai-makes-25-billion-a-year-and-is-preparing-for-an-ipo-here-is-what-the-numbers-actually-mean/

[2] Reported by news; neither company has confirmed these figures publicly. https://www.webpronews.com/anthropics-margin-squeeze-inference-costs-bite-as-revenue-surges/; https://www.theinformation.com/articles/anthropic-lowers-profit-margin-projection-revenue-skyrockets; https://taptwicedigital.com/stats/anthropic

[3] https://stockn.xueqiu.com/02513/20260419954892.pdf

[4] https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results

[5] https://eqvista.com/gross-profit-margin-saas-tech-companies/

[6] https://ramp.com/data/ai-index

[7] https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/

[8] https://www.intercom.com/learning-center/ai-customer-service-agent-pricing-comparison; https://www.salesforce.com/news/press-releases/2026/06/15/salesforce-signs-definitive-agreement-to-acquire-fin/

[9] https://cognition.ai/blog/ai-guarantee

[10] This is an illustrative calculation using public rate cards, not a disclosed contract.

[11] https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance

[12] https://stripe.com/en-sg/resources/more/outcome-based-pricing

[13] https://newsroom.servicenow.com/press-releases/details/2026/ServiceNow-opens-its-full-system-of-action-to-every-AI-Agent-in-the-enterprise/default.aspx

[14] https://www.servicenow.com/community/upgrades-and-patching-forum/servicenow-ai-native-licensing-in-2026-a-practical-guide-to/td-p/3565858

[15] https://www.servicenow.com/products/itsm/pricing.html

[16] https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/agents-legal-responsibility.pdf

[17] https://www.europarl.europa.eu/RegData/etudes/BRIE/2023/739341/EPRS_BRI(2023)739341_EN.pdf

[18] https://www.gov.uk/government/publications/complying-with-consumer-law-when-using-ai-agents

[19] https://www.ftc.gov/news-events/news/press-releases/2026/07/ftc-seeks-public-comment-policy-statement-addressing-ai-accuracy