By clicking “Accept”, you agree that cookies may be stored on your device in order to improve site navigation, analyze site usage, and help us with our marketing efforts. Check out our for more information.

Guides

AI billing: how to charge for AI and rebill tokens to customers

#QUOTE-TO-CASH Insights
Opencell #QUOTE-TO-CASH Insights cover: AI billing, the guide

AI billing: how to charge for AI and rebill tokens to customers

OpencellNovember 21, 202618 min read

Adding AI to software has become easy. Billing it without losing margin has not.

Software vendors have added artificial intelligence at high speed. According to the Numeum and EY TOP250 Panorama, published on November 4, 2026 from a survey of 274 companies, nearly four French software vendors in five have now built AI features into their solutions, 19 points more than a year earlier. The same panorama notes that subscription remains dominant, but that usage-based billing is expected to grow.

That is where the difficulty begins. Every question asked to the assistant, every document summarized, every task handed to an agent consumes tokens from an AI model provider. That cost varies from one customer to another, from one month to the next, sometimes from one request to the next. Yet most vendors still sell a fixed subscription. The result: the more customers adopt AI, the more the margin melts.

AI billing is about solving that problem: choosing how customers pay for AI, rebilling all or part of token consumption, protecting margin and producing invoices customers understand. This guide is for software and SaaS vendors that put AI at the core of their product.

For the bigger picture, see our guides to AI monetization, SaaS billing and usage-based billing.

The problem: a variable cost inside a fixed price

When a vendor adds AI to software sold by subscription, it introduces a variable cost, token consumption, into a fixed price. While usage stays low, the margin holds. When customers adopt AI, cost rises with usage while the price does not move: the margin shrinks, customer by customer.

Classic software has an almost zero marginal cost: serving one more user costs very little. AI changes that equation. Every use has a direct cost, billed by the model provider, and that cost is very uneven:

  • from one customer to another: a customer who uses AI ten times a day costs far more than one who tries it once a month, for the same subscription price,
  • from one feature to another: summarizing an email consumes almost nothing, analyzing a hundred-page contract or launching an agent that chains dozens of calls costs much more,
  • from one month to the next: the provider changes its prices, releases new models, and usage evolves.

Example: a vendor sells a subscription at €40 per user per month. It adds an AI assistant included in the price. If an average user consumes €2 of tokens a month, the margin remains comfortable. If the most active users consume €30, each of them becomes almost unprofitable. The problem is not the average cost, but its dispersion.

Where AI costs come from

AI cost is mainly measured in tokens, fragments of text processed by the model. Providers bill input tokens, the question and context sent, and output tokens, the answer produced, separately, at prices that vary by model. Agents, which chain several calls to complete a task, multiply consumption.

The main cost drivers:

Driver What makes the cost vary
Input tokens Length of the question, documents and history sent to the model
Output tokens Length of the answer produced, usually more expensive per token
Model choice A powerful model costs much more than a light one for the same task
Context Documents, customer data or history added to each request
Agents One task can trigger several successive calls, searches and checks
Images, audio, video Specific rates, often higher than text
Provider prices Frequent changes, new models, differences between providers
Currency Major providers often bill in US dollars

One trend complicates things further: the unit price of tokens keeps falling, but usage becomes more demanding, with longer contexts and agents that reason in several steps. The cost per task therefore does not necessarily fall. A pricing policy based on one month's cost can be outdated a few months later.

6 ways to charge customers for AI

There are six main ways to charge for AI: include it in the subscription with fair use, sell it as a per-user add-on, sell credits, rebill token consumption with a margin, charge per business action or charge per outcome. Vendors often combine several in a hybrid offer.

Model Principle Advantages Risks
AI included with fair use AI is part of the subscription, with a cap or fair-use policy Simple, drives adoption Margin exposed if the cap is set wrong
Per-user add-on A monthly surcharge per user who accesses AI Predictable for customer and vendor Does not follow actual consumption
Credits The customer buys or receives a stock of credits, consumed with each use Links price to usage, cash collected upfront Requires managing balances, expiry and top-ups
Token rebilling The customer pays actual consumption, at purchase cost plus a margin Margin protected, transparent Unit means little to the customer, variable invoice
Per action A price per business task: document analyzed, ticket handled, report generated Understandable and value-linked Cost per action varies and must be monitored
Per outcome The customer pays only when AI achieves a result: ticket resolved, meeting booked Aligned with value Outcome hard to measure and easy to dispute

In practice, the strongest offers are hybrid: a subscription that includes a volume of AI, then credits or consumption billed beyond it. The customer keeps a predictable budget, the vendor protects its margin on heavy users.

The choice depends on three questions:

  • can the customer link usage to value? If so, a price per action or per outcome is easier to sell than a price per token,
  • is cost widely dispersed between customers? If so, a purely flat model is risky,
  • does the customer need a fixed budget? Enterprise accounts and public buyers often prefer an annual commitment with a pool of credits.

Rebilling tokens: the method

Rebilling tokens means measuring each customer's consumption, rating it at the provider's purchase cost plus a margin, then presenting it on the invoice in a unit the customer understands. The difficulty lies in differences between providers, price changes and readability.

The method in five points:

  1. Measure per customer. Every call to the model must be attached to a customer, and ideally to a user and a software feature. Without that attribution, no rebilling is possible.
  2. Rate at real cost. Apply the provider's purchase price for each model, distinguishing input and output tokens, and converting currency if needed.
  3. Add a margin. The margin covers infrastructure, support, currency risk and price changes. It can be a coefficient on cost or a sale price set per unit.
  4. Translate into an understandable unit. Few customers know what a million tokens represents. It is often better to convert tokens into credits, requests or business actions.
  5. Separate sale price from purchase price. Set a sale price per unit, reviewed periodically, rather than passing every provider change through to the invoice. The customer gains predictability, the vendor keeps control of its margin.

Example: a vendor buys tokens from a provider and finds that a document analysis costs it €0.04 on average, with wide variations. Rather than billing tokens, it sells credits: one analysis is worth one credit, sold at €0.10. Every month it tracks the real cost per credit and adjusts the scale if the margin falls below its threshold.

Designing AI credits

AI credits turn technical consumption into a simple commercial unit. To work, a credit model must define the value of a credit, the consumption scale per feature, expiry and rollover rules, top-ups, alerts and the accounting treatment of credits sold in advance.

The decisions to make:

Question Common options
How much is a credit worth? A fixed price per credit, decreasing with the volume bought
How many credits per use? A scale per feature: a simple question, an analysis, an agent task
How are they obtained? Included in the subscription, bought in packs, or both
Do they expire? Monthly expiry for included credits, longer for purchased credits
What happens at zero? Blocking, automatic top-up, or consumption billed as overage
How does the customer track the balance? Counter in the software, alerts at 50, 80 and 100%
How are they accounted for? Prepaid credits only become revenue as they are consumed

Three mistakes come up often: a scale so detailed nobody understands it, credits given away without limit that create costly habits, and no clear rule for unused credits at the end of the contract. The accounting and tax treatment of credits sold in advance should be validated with your accountant. See our guide to IFRS 15 revenue recognition.

Protecting your margin

Protecting AI margin requires knowing real cost per customer and per feature, setting caps and alerts, reserving the most expensive models for uses that justify them and reviewing prices regularly. Steering must be continuous, not at year-end close.

The levers:

  • track cost per customer and per feature, not just total AI cost, to spot unprofitable customers and uses,
  • set caps in included offers, with a clear path to a higher tier or to credits,
  • warn the customer before they reach their limit, rather than sending a surprise invoice,
  • choose the right model for each task: a light model for simple tasks, a powerful model only when it makes a real difference,
  • decouple sale price from purchase price, with a scale reviewed at regular intervals,
  • track AI gross margin as a separate KPI, alongside classic SaaS KPIs.

What your billing system must be able to do

To bill AI, the billing system must collect consumption in real time per customer, rate it on several metrics, manage prepaid credit balances, apply caps and overages, alert customers, produce readable invoices and spread revenue over time. A classic subscription module is not enough.

Capability Why it is needed
Usage collection Millions of usage events, attached to the right customer, without loss
Several metrics Tokens, requests, actions, credits, users, combined in one offer
Flexible rating Scales per feature, tiers, volume discounts, negotiated prices
Credit balances Purchase, consumption, expiry, top-up and rollover, in real time
Caps and overages Automatic switch from an included plan to billed consumption
Alerts Consumption thresholds visible to the customer and the sales team
Fast offer launches Test a new price or AI option without development
Readable invoices Consumption detail the customer can understand
Revenue Revenue recognized as credits are consumed
Integration Links with the software, CRM, ERP and payment tools

The last point is often underestimated: AI billing evolves fast. A vendor that has to wait for development to change a credit scale will not keep up with its providers' prices or its competitors'. Our billing guide details how to choose a platform.

Enterprise accounts: commitments and credit pools

Enterprise accounts are reluctant to accept an unpredictable AI invoice. They prefer an annual commitment that gives access to a credit pool, usable by all their teams, with clear rules for overage and tracking by entity.

The most common mechanisms:

  • the annual commitment: the customer commits to an AI amount for the year, in exchange for a lower unit price,
  • the shared pool: a stock of credits common to several subsidiaries, teams or projects, with tracking by entity,
  • controlled overage: beyond the pool, consumption billed at the contractual rate, or a top-up approved by the customer,
  • internal chargeback: consumption detail by entity lets the customer allocate the cost across its own teams.

These mechanisms require managing account hierarchies and tailored contracts in the billing system, as for any enterprise account. Our guide to quote to cash explains how to link these negotiated contracts to the invoice, and our guide to multi-entity billing covers hierarchies.

Setting up AI billing in 6 steps

To set up AI billing, measure real cost per customer, choose a unit of value, build a hybrid offer, equip usage collection and billing, test on a customer segment, then steer margin every month.

  1. Measure real cost. Attach every model call to a customer and a feature. Calculate cost per customer and spot the dispersion.
  2. Choose the unit of value. Tokens, credits, actions or outcomes: keep the unit your customers understand and that you can measure reliably.
  3. Build the offer. Usually a volume of AI included in the subscription, then credits or consumption billed beyond, and commitments for enterprise accounts.
  4. Equip billing. Usage collection, credit balances, caps, alerts and detailed invoices.
  5. Test. Launch the new offer on a customer segment or on new customers, and measure adoption and margin.
  6. Steer. Track cost per customer, AI margin and changes in provider prices every month, and adjust the scale.

Where Opencell fits

Opencell is an open source monetization platform designed for usage-based models. It collects and rates large volumes of usage events, combines subscriptions, prepaid credits, tiers, commitments and overages in a single offer, and lets you launch a new price without development. It integrates with the vendor's software, CRM and ERP through more than 200 APIs and connectors. See the platform.

Born in telecoms in 2007, where billing billions of usage events is the norm, Opencell now manages €14 billion in revenue for around fifty customers, including cloud providers such as OVHcloud and Outscale. Opencell also develops its own AI agents, presented in our guide to AI agents for billing.

Frequently asked questions about AI billing

How do you charge customers for AI?

There are six main ways: include AI in the subscription with fair use, sell it as a per-user add-on, sell credits, rebill tokens with a margin, charge per action or per outcome. The strongest offers combine an included volume with consumption billed beyond it.

How do you rebill AI tokens to customers?

Measure each customer's consumption, rate it at the provider's purchase cost plus a margin, then present it in an understandable unit, such as credits or actions. It is better to set a stable sale price, reviewed periodically, than to pass every provider change through.

What is a token in AI?

A token is a fragment of text, often part of a word, processed by an AI model. Providers bill input tokens, sent to the model, and output tokens, produced by the model, at prices that vary by model.

Should AI be included in the subscription or billed separately?

Including a volume of AI drives adoption, but exposes the margin if usage exceeds forecasts. The most common solution is to include a capped volume and bill consumption beyond it, through credits or usage.

What are AI credits?

AI credits are a commercial unit representing AI consumption. The customer receives them in the subscription or buys them in packs, then each use consumes a defined number. They make AI cost predictable for the customer and usage-linked for the vendor.

How do you protect your margin on AI?

Track cost per customer and per feature, set caps and alerts, choose the right model for each task, decouple sale price from purchase price and review the scale regularly.

How do you bill AI agents?

Agents can be billed per completed task, per outcome achieved or through credits consumed. Because an agent often chains several model calls, tracking the real cost per task is essential to set a profitable price.

Which software should you use to bill AI?

You need a billing platform able to collect consumption in real time, handle several metrics and credit balances, apply caps and overages and launch new prices without development. A classic subscription module is usually not enough.

AI is not billed like software. It is billed like a consumed service.

Putting AI at the heart of software takes vendors into a world that telecom operators and energy suppliers have known for a long time: a cost that varies with usage, customers who want predictability, and a margin that is won or lost in billing.

Key takeaways:

  • AI puts a variable cost inside a fixed price: dispersion between customers is what threatens the margin,
  • six models exist, and hybrid offers, included volume then consumption, are the strongest,
  • rebilling tokens requires measuring per customer, adding a margin and using a unit the customer understands,
  • AI credits need clear rules for expiry, top-ups and accounting,
  • billing must follow usage in real time and let you change a price without development.

AI margin is not won in the model. It is won on the invoice.

Go further with Opencell

Opencell is an agile, open source monetization platform that automates quote-to-cash, from quote to cash collection, for subscription and usage-based services. Born in telecoms in 2007, where billing billions of usage events is the norm, it manages €14 billion in revenue for around fifty customers, including cloud providers such as OVHcloud and Outscale.

Adding AI to your software and working out how to bill it?

Optimize your recurring revenues with Opencell

Get a demo