THALES Article

Subscription vs. Consumption vs. Outcome-Based AI Pricing

Summary: Using AI has a direct cost. Inference, model calls, and automated tasks consume compute and often incur third-party model fees. AI ISVs need pricing that reflects those costs as usage grows. Subscription, consumption, and outcome-based pricing approach that problem differently, while hybrid models combine a predictable base price with charges for heavier use.

Historically, SaaS was priced around more predictable costs. Once the software was developed and hosted, adding another user was relatively inexpensive. Per-user subscriptions provided recurring revenue, and customers found predictable bills comforting.

AI changes the relationship between users, use, and cost. A product often needs fewer human users as it becomes more capable, while the amount of machine activity behind those users increases. Revenue tied to seats can fall at the same time as the cost of providing the service rises.

Software companies are responding with different combinations of subscription, consumption, and outcome-based pricing. The right choice depends on how customers use the product, what drives the provider's costs, and what customers consider valuable enough to pay for.

Why Software Pricing Models Are Evolving

For years, the move from perpetual licenses to subscriptions dominated software monetization. Customers moved from buying software outright to paying monthly or annually for continued access.

Software is now consumed through cloud applications and APIs, as well as embedded systems and connected devices. AI adds machine activity that may bear little relationship to the number of people using the product. Each inference, model call, or automated task has a cost attached, be it compute and infrastructure, energy, data processing, or third-party model fees. As customers use more AI, those costs grow.

McKinsey's July 2026 Enterprise AI FinOps Survey found that 93% of companies have exceeded their AI budgets, while only 26% have real-time visibility into what running AI is costing them. Sixty percent say the cost trajectory of their AI initiatives is unpredictable, right as CFOs, boards, and investors are demanding proof of ROI. For AI software companies, growing usage carries a cost that's hard to see coming.

Per-seat pricing presents a separate challenge. As AI takes on work previously done by people, customers may need fewer users even as they get more value from the software.

Consider an AI product that once supported ten people doing a particular job. If automation allows two people and several AI agents to produce the same output, a per-user model now captures revenue from two seats. The agents may be generating thousands of model calls and other chargeable activity in the background.

More product use can carry more cost without generating corresponding revenue.

There are two decisions involved when a software company reviews its pricing. The price model covers how the customer pays, such as through a subscription or pay-as-you-go arrangement. The price metric is the unit being charged for, such as a user, API call, transaction, or completed task.

Those decisions can be mixed in different ways. A subscription can include a fixed amount of usage rather than being priced per user, for example. This gives software companies more options than choosing between subscription and consumption as two entirely separate models.

The Usage Monetization Era

The Business Model for the AI Revolution

Consumption-based pricing is on every software vendor’s radar today.

Find out how to implement the right infrastructure without disrupting your tech stack. Read this new white paper, The Usage Monetization Era: The Business Model for the AI Revolution.

The Usage Monetization Era The Business Model for the AI Revolution

What Is Subscription Pricing?

Subscription pricing gives customers access to software for a recurring fee, usually paid monthly or annually. The price can vary according to the package purchased and the entitlements included.

How Subscription Pricing Works

A customer chooses a package and pays a fixed amount for access during the subscription period. Different packages can provide access to different features, usage allowances, or service levels.

AI features can work within the same arrangement. An AI capability might be included in a particular subscription tier, or the subscription could include a defined allowance for AI use.

The important point is that subscription pricing does not necessarily mean charging per seat. The recurring payment is the price model. The vendor still has to decide what metric determines the price and what the customer receives for it.

Advantages and Limitations of Subscription Pricing

Subscriptions give customers predictable bills and provide vendors with recurring revenue. They are also familiar to procurement teams and easy to explain. They work particularly well when AI usage stays within a reasonably consistent range. The provider can price the subscription with enough room to absorb the expected cost of serving each customer.

Large variations in usage make that harder. Two organizations paying the same monthly fee could generate very different inference costs.

Agentic AI adds another variable to the equation. For example, a user request sets off a workflow involving multiple model calls, tools, and actions. McKinsey's research into AI costs claims that the same task can generate very different token volumes and agent chains.

Heavy users can then become progressively more expensive to serve while their subscription remains unchanged. If those costs rise faster than subscription revenue, higher usage starts to eat into margin.

A per-user subscription has a further limitation for products where AI replaces some of the work users previously performed. Fewer seats may reduce revenue even while the software handles more work.

When to Use Subscription Pricing

Subscriptions make sense when customers use the product at broadly predictable levels and the cost of providing it does not vary dramatically between them. They can also suit products where customers want to know their software costs in advance. A company may prefer a fixed annual commitment to a monthly bill that changes according to activity.

AI Independent Software Vendors (ISVs) need to question how much variation lies behind that fixed price. If one customer's AI consumption costs several times as much to support as another customer's, charging both the same amount is harder to justify or sustain.

Usage data can show whether that gap is developing. The vendor can then adjust packages or introduce an allowance and overage charge rather than replacing subscriptions altogether.

What Is Usage-Based Pricing (Consumption Pricing)?

Usage-based pricing, also called consumption pricing, charges customers according to how much of a product or service they use.

For AI software, the chargeable unit might be an API call, token, transaction, amount of data processed, or another measurable activity.

How Usage-Based Pricing Works

The provider first needs to decide what to meter. The most obvious technical unit is not always the best commercial one. A vendor may be able to count model calls precisely, for example, while its customers care more about the number of documents processed or tasks completed.

Once the metric is chosen, the provider measures consumption and uses it to calculate the customer's charge. The rate can be fixed per unit or vary according to volume and other commercial terms.

Tokens or credits can make consumption pricing easier to manage. A simple model call might use only a small number of credits, while a longer agent workflow uses substantially more. Customers still pay according to their use without having to try and make sense of every technical unit behind it.

The vendor decides how many credits each capability consumes and can include an allocation within a subscription. New AI capabilities can also be assigned their own credit values as the product develops.

For more on static, usage, and outcome metrics, including metering and tokens for consumption pricing, see our guidance on AI monetization.

Advantages and Limitations of Usage-Based Pricing

Consumption pricing charges for those activities that can be measured, like API calls, tokens, compute time, transactions, data processed, or finished AI tasks. As usage increases, revenue rises with it, giving the provider a way to recover the higher costs of serving heavier users. Customers can start with relatively little commitment and spend more as they use more.

Consumption pricing can create a ‘taxi meter’ effect. If customers are not sure what their AI use will cost, they may start keeping a closer eye on every call or token and hold back on using the product as much as they otherwise would have. McKinsey has named this lack of price predictability as a worry for enterprise AI buyers.

The vendor also has to measure consumption in a reliable way. Cloud-native products can report usage centrally, but AI software does not always run in a continuously connected cloud environment. Models may be deployed on-premises, at the edge, or on devices that spend periods offline.

A consumption model in those environments depends on reliable metering and entitlement enforcement. The provider must still know what has been used before it can charge for it.

The unit being measured also has to make commercial sense. Customers may understand a charge per transaction or completed task more readily than one based on the underlying compute needed to produce it. The metric needs to make sense to the customer as well as the billing system.

When to Use Usage-Based Pricing

Usage-based pricing is useful when consumption varies enough that a fixed price no longer reflects how customers use the product. AI products with large differences in model usage are a clear candidate, as are products where customers can start small and increase consumption substantially without adding more users.

There needs to be something dependable to meter. Ideally, that metric should also have a relationship with the value the customer receives. Charging for an internal technical event that customers neither understand nor control can make the bill difficult to explain.

The provider also needs to know how the proposed model would behave before rolling it out. Historical usage data can be used to test what existing customers would have paid under a new metric and identify unusually large changes before they appear on an invoice.

What Is Outcome-Based Pricing?

Outcome-based pricing charges for the work the software produces. A customer might pay for a support case resolved, a fraud event prevented, or an automated task completed.

AI makes this increasingly plausible because its outputs can often be counted. An agent completing a defined task gives the vendor something more tangible to price than access to the software.

How Outcome-Based Pricing Works

The charge is linked to an agreed result. For a customer-service application, that could be a successfully resolved case. An automation product could charge for completed tasks.

The work required to reach that result does not necessarily determine the price. One task might require several model calls while another requires many more. If both meet the agreed definition of a successful outcome, the customer pays for that result.

That approach requires both parties to agree on what counts as success. The provider also needs evidence that the result occurred and that its software was responsible for it.

Advantages and Limitations of Outcome-Based Pricing

Outcome pricing can make the unit being bought easier for customers to understand. They are paying for a result instead of the technical activity it takes to produce it. It can also suit AI products that perform work once carried out by people. Charging for the completed work may reflect the product's value more closely than counting the number of people overseeing it.

Defining the outcome is where the model gets tricky. Both sides must agree on how success is measured, and the software's contribution must be identifiable. A fraud detection product, for instance, might flag a suspicious transaction, but establishing an appropriate financial value for that action is not as easy.

Delivery also has to be consistent enough to support a commercial agreement. These requirements leave outcome pricing better suited to products for which results can already be measured and attributed reliably.

When to Use Outcome-Based Pricing

Outcome-based pricing works best when the customer and provider can agree on what counts as a successful result. Defined tasks make that easier. If an AI agent processes a claim or completes another verifiable task, there is a specific outcome to charge for.

The model becomes more difficult when judgment enters the definition of success. It can also break down when several systems or human decisions contribute to the final result and nobody can isolate the software's contribution.

Companies do not need to work their way through subscription and consumption pricing before considering outcomes. If a product already produces results that can be measured and attributed reliably, outcome pricing can be considered from the start.

How Hybrid Pricing Combines Subscriptions and Consumption

For many AI ISVs, hybrid pricing is a good combination of a predictable subscription with charges that increase as AI use grows. A base subscription can include a set amount of usage, with the customer paying for anything above that allowance.

Credits can also be used within this model. A subscription might include a credit allocation, with different AI capabilities using credits at different rates. Customers know the base cost, while heavier use generates additional charges.

A hybrid model also gives an established subscription business a way to introduce consumption gradually. The provider can keep the recurring base fee and decide how much usage to include before overage charges begin.

The same model does not have to apply across an entire product portfolio. One product may have predictable consumption that fits a subscription, while another may have wide discrepancies in AI use that make an allowance and overage model more appropriate.

What Systems Do You Need to Support Evolving Pricing Models?

A pricing team can decide to charge by API call relatively quickly. Putting that decision into production takes more work.

First, the company has to know what customers are using. If it has historically sold fixed subscriptions without tracking consumption, it may have little evidence for deciding how much usage to include or what an overage should cost.

That makes product telemetry useful before the pricing change as well as after it. Historical usage gives the company a baseline for testing proposed metrics against existing customer behavior. Once consumption affects the bill, the same data has to support accurate charging.

Entitlements then connect the commercial offer with what the customer can use. If a package includes particular features or a monthly allowance, the software needs to enforce those terms. Upgrades, additional capacity, and changes to the package also have to reach the customer's entitlement without creating a manual licensing job each time.

Deployment complicates the picture. A cloud service can usually report consumption to a central system as it happens. Software running on-premises, at the edge, or on intermittently connected devices may need to record usage locally and report it later. The metering approach has to work wherever the product is deployed.

Billing needs the resulting usage data in a form it can use. Under a hybrid model, for example, the system has to know how much consumption the subscription includes and when the customer has moved into overage. A credit model needs to keep track of the customer's balance and how quickly different capabilities consume it.

Customers need access to some of that information, too. A company paying according to consumption should be able to see how much it has used and how much of an allowance remains. Without that visibility, the first sign of unexpectedly heavy consumption may be the invoice.

Packaging also becomes harder to change when every commercial variation has to be coded into the product. Software companies experimenting with new pricing need a way to change feature access, usage limits, trial conditions, and other entitlements without rebuilding the underlying application for every offer.

The Usage Monetization Era

The Thales Usage Monetization Era whitepaper sets out five stages for managing usage-based pricing - ingestion, mediation, metering, invoicing, and monetization.

Software companies need to capture usage wherever their products run, turn raw telemetry into reliable records, and measure consumption against each customer’s entitlements. When customers reach their limits, vendors can issue a warning, restrict further use, or allow consumption to continue with overage charges. The resulting data can go into existing billing systems without replacing them.

Sales teams can pinpoint customers approaching their limits, while product teams can see which features customers use, and assess whether their pricing still makes sense. It can also help vendors assess new pricing metrics against actual customer behavior.

Keeping metering and entitlement management separate from billing, where the requirements justify it, allows new vendors to introduce new pricing models without having to rebuild their existing systems.