AI Usage Billing: How to Meter, Price and Invoice It

TL;DR: Charge for what customers use, not a flat fee alone. Most AI products do best with a hybrid: a base subscription plus one simple usage metric. Build metering in from the first sprint, stop bill shock with limits and alerts, and make sure usage invoices meet ATO GST rules.

AI usage billing means you charge customers for what they use, not just a flat fee. That could be tokens, model calls, compute time or finished tasks. For most AI products the best answer is a hybrid: a base subscription, plus a metered charge for heavy use. It only works if you build the metering in from the first sprint. This guide covers what to meter, which model to pick, how to stop surprise invoices and what GST means for your usage invoices.

The short version

• AI costs rise with use. A flat fee lets heavy users run up your bill while light users pay the same.
• Hybrid is the usual answer. A base subscription plus one simple usage metric.
• Meter every billable event. Give it a customer ID, a unique event ID, the model used and a timestamp. Never edit raw events.
• Stop bill shock in the product. Use limits, alerts and credit balances before an invoice is ever sent.
• GST still applies. If you are GST-registered, usage invoices must meet ATO tax invoice rules.

Why AI products need usage-based billing

A normal SaaS product costs about the same to serve whether a customer logs in once or fifty times. An AI product does not. Every prompt, lookup and agent step uses model inference. That cost lands on your bill from your model provider.

A token is the small chunk of text a model reads or writes. It is the most common unit providers charge by. Two customers on the same flat plan can use very different amounts of it.

Margins make this urgent. The 2026 State of B2B Monetization report, as summarised by Schematic, puts AI product margins near 50%. Traditional SaaS sits at 70 to 80%. Charging in line with use keeps your margin steadier when usage, models or provider rates change.

Agents add another wrinkle. Stripe's guide points out that agent loops can fan out into many tool calls and that token counts vary a lot. One user request can trigger dozens of billable steps. If you do not measure them, you cannot price them.

Start with your own cost per unit. Our guides to what the OpenAI API really costs and LLM API pricing compared show how to work it out. Your cost per unit sets the floor for your price.

What you can meter

Metering is the measuring layer. It records what each customer did so you can charge for it. Common units for AI products:

• Tokens, split into input and output. Zuora's guide rates them as separate lines because their costs differ.
• Model or API calls, when each call costs about the same.
• Embedding requests, for RAG systems that index customer documents.
• Compute seconds, for image, audio, video or fine-tuning work.
• Agent actions or finished tasks, like a resolved support ticket or a processed document.

The unit you charge does not have to be the unit you pay for. Ganivra's guide on infro.io says documents or completed workflows are easier for customers to predict than raw tokens. Meter tokens inside your product for cost control. Bill in documents, minutes or conversations.

Australian providers already do this. VerticalAI, an Australian voice AI service, meters use past plan allowances at A$0.40 a minute, A$0.12 a text and A$0.05 a chat conversation, billed monthly in arrears. A customer can picture a minute or a chat. Few can picture a token.

How to record usage so invoices hold up

Most billing disputes start with weak event data. Each usage event should carry:

1. A stable customer ID, plus a project or tenant ID where it matters.
2. A unique event ID, so duplicates can be dropped.
3. A timestamp.
4. The unit and the quantity.
5. The provider and model version, because rates change by model.
6. A correlation ID that ties the invoice line back to your app logs.
7. A billable flag, a reason code and a schema version.

Model version matters more than most teams expect. When you swap a model, your cost per call changes. Our guide to AI model versioning in SaaS shows how to keep that traceable.

A few rules keep the data honest:

• Use the usage numbers your provider reports instead of guessing token counts.
• Keep test traffic out of production data. Internal experiments must never reach a customer.
• Keep raw events fixed. Ganivra suggests fixing mistakes with adjustment events, not edits, so you keep an audit trail.
• Never treat an unknown cost as zero. Mark it unpriced until it is sorted.
• Decide attribution up front. If one agent session spans several users or tenants, decide who pays before the first invoice.

Then keep the layers apart. Event capture, ingestion, metering and invoicing should each be correct on their own. A bug in invoice formatting should never corrupt the usage record. This data usually sits in a warehouse next to your product analytics. Our piece on API usage analytics shows how the same data feeds revenue decisions.

Choosing a pricing model

There are three broad models. Each suits a different product.

Stripe's pricing guide lists hybrid as one of the core AI pricing models and says to start with the simplest model that fits. The m3ter team looked at the top 20 AI pricing pages and found 17 of 20 had a usage-based part.

Here is a local tiered example. Elfworks, an Australian AI software provider, set three tiers in September 2025. Light is $19 a month. Professional is $49 with a 15,000,000 token limit. Premium is $89 with 30,000,000. All prices are GST exclusive, and the token limits are soft. Customers are not cut off mid-task when they pass one.

To pick a model, ask three questions:

• Does cost per customer vary by more than a few times? If yes, you need a usage part.
• Can the customer predict their own use? If not, give a base fee with an included allowance.
• Is the buyer a finance team or a developer? Finance teams want a steady monthly total. Developers accept pay-as-you-go.

Setting prices and protecting margin

Stripe's advice is to pick a value metric that tracks both customer value and your costs. Keep the usage part to one clear metric, not a grid of rates.

In practice:

1. Work out cost per unit. Count model calls, embeddings, vector storage and any third-party services. Then add your overhead.
2. Add a margin that survives rate changes. Providers change prices and release new models. Review your price list on a set rhythm.
3. Pick a unit customers understand. Documents, chats or minutes sell better than tokens.
4. Set allowances by plan. The allowance should cover a normal customer's month, so extra use feels like growth, not a penalty.
5. Publish your rate-change terms. VerticalAI commits to 30 days' notice before rates change.

Clarity is a legal issue in Australia too. In October 2025 the ACCC took Microsoft to the Federal Court. It alleged Microsoft misled about 2.7 million Australian customers about their plan options after adding Copilot to Microsoft 365. The Personal plan rose from $109 to $159 a year, The Guardian reported. Whatever you build, say plainly what customers pay for and what their options are.

Preventing surprise customer bills

Bill shock is the quickest way to lose a customer. Agents make it easy to cause. A loop that keeps retrying a failed tool call can burn a month's allowance in minutes. Put the controls in the product, before the invoice.

• Spend limits and alerts. Tell customers at set points, like 50%, 80% and 100% of their allowance.
• Credit balances and reservations. Stripe describes credit ledgers and reservations, so you never do work that cannot be paid for.
• Circuit breakers. Stop runaway agent loops automatically.
• Anomaly alerts. Flag a customer whose use jumps tenfold overnight. Contact them before the invoice goes out.
• Rate limits and quotas. These protect your costs and your platform. See our guide to enterprise API rate limits.
• A live usage dashboard. Customers who can see their use dispute less.

Test billing before launch

Customers find billing bugs, and they cost a lot. The Revenium docs walk through setting up a plan with usage included and overage charged after. Check that your usage data flows, your plan charges the right amount and the invoice total matches, all before you go live.

The LangWatch docs add two habits worth copying:

• Verify webhook signatures and drop duplicates using a stable event ID.
• Replace an early usage figure with the settled one. Do not add both.

Then run a reconciliation every month. Compare your provider invoices to your metered totals. A gap means events are being lost, duplicated or priced wrong.

GST and invoicing in Australia

Usage billing does not change your GST duties, but it makes invoices more detailed. Check the points below with your accountant against current ATO guidance. Treatment depends on your registration and your customer.

• Registration. GST registration is generally required once your 12-month GST turnover hits A$75,000, and GST is generally 10% on most services, per ExpenseFlow's guide.
• Tax invoice contents. The ATO's GSTR 2013/1 covers what a tax invoice needs. That includes who supplied it, their ABN, the date and what was supplied. The buyer's identity or ABN is needed when the total is $1,000 or more. A supplier must give a tax invoice within 28 days of a request.
• Usage lines. Show the unit, the quantity and the rate on each line.
• Published prices. Say whether advertised prices include or exclude GST.

Your invoices come from metered data. So your billing system has to produce the ABN, line details and GST amount on its own. Build that into the invoice layer from the start.

Build usage billing into your AI product

Usage billing is part of the platform, not a side project. The same events that drive an invoice also show which features customers value and what each customer costs you. Plan it with the first production build. Retrofitting means rebuilding usage history that was never captured.

Devwiz builds AI agents, AI-first platforms, custom apps and the data and automation underneath them. We have shipped 200+ apps, with work for NSW Government, Briometrix, Vivid and Huskee. To scope metering, pricing logic and invoicing as part of your build, start with our AI program and platform development service, or see the full services list. Our AI platform architecture guide and post-launch support guide show where billing and monitoring fit.

For more on pricing in the AI era, see AI-Led's take on usage-based pricing and The AI Orchestrators on LLM cost optimisation.

Got an AI product that needs billing built in? Book a call with Devwiz and we will map it out.

Frequently asked questions

What is usage-based billing for AI?

It is pricing based on measured use, such as tokens, API calls, compute time or agent actions, instead of a flat access fee. It can run as pay-as-you-go, prepaid credits or a hybrid with a base subscription. It suits AI because your own costs rise with each request.

Should I bill customers per token?

Only if your customers are technical and comfortable with it. Many products meter tokens inside the product and charge in units customers understand, like documents, conversations or minutes. That keeps invoices predictable while you still track your real cost.

How do I stop customers getting a surprise invoice?

Build the controls into the product: usage alerts at set points, spend limits, prepaid credit balances, circuit breakers for runaway agent loops and a live usage dashboard. Contact customers when a sudden spike is flagged, before the invoice goes out.

Do usage invoices need to include GST?

If you are GST-registered, a taxable supply of AI software or services generally attracts GST at 10%, and the invoice must meet the ATO's tax invoice rules in GSTR 2013/1. Confirm the treatment for your customers with your accountant or the ATO.

When should I add metering to my AI product?

During the first production build. Retrofitting means rebuilding usage history that was never captured, which makes early invoices hard to defend. Design event capture, cost tracking and attribution with the feature itself.