MetricaOS

Glossary

AI Cost Attribution

· · Leave a Comment

What is AI cost attribution?

AI cost attribution is the process of assigning AI-related costs to the customers, users, features, models, or workflows responsible for generating them.

An AI company may know that it spent $10,000 on model providers during a month. Cost attribution explains where that $10,000 came from.

For example:

Customer A: $1,800

Customer B: $420

Customer C: $95

Document analysis: $4,200

AI assistant: $3,100

Report generation: $2,700

Instead of seeing only total provider spend, the company can understand which parts of the product are creating cost.

AI usage metering provides the underlying usage data needed to make this attribution possible.

Why AI cost attribution matters

AI products often have variable costs. Every model request, document analysis, generated response, agent action, or AI workflow can increase the cost of serving a customer.

Two customers paying the same subscription price may therefore have completely different economics.

For example:

Customer A

Revenue: $199/month

AI cost: $24/month

Customer B

Revenue: $199/month

AI cost: $260/month

Without customer-level attribution, both accounts appear to generate the same revenue.

With attribution, it becomes clear that Customer B may be unprofitable.

Cost per customer helps teams measure this relationship between customer usage, cost, and revenue more directly.

What can AI costs be attributed to?

Most AI products should attribute costs across several dimensions.

Customer

Customer-level attribution shows how much each account costs to serve.

This can reveal heavy users, unprofitable accounts, and customers that may need higher plans, overages, or enterprise pricing.

Feature

Feature-level attribution shows which parts of the product create the most AI spend.

For example:

Document summaries: $2,100

AI search: $1,400

Agent workflows: $5,800

Email generation: $700

This helps product teams understand whether expensive features are delivering enough value to justify their cost.

Model and provider

Different models can have very different cost structures.

Model-level attribution helps teams identify whether expensive models are being used where cheaper alternatives could perform adequately.

User or workspace

For B2B products, costs may also need to be attributed to individual users, teams, projects, or workspaces.

This is particularly useful for enterprise usage reporting and departmental usage controls.

Workflow

Some AI features make several model or tool calls behind the scenes.

Workflow-level attribution combines these related costs so teams can understand what a complete product action actually costs.

How does AI cost attribution work?

AI cost attribution starts when the product records a usage event.

A simplified event might contain:

Customer: cus_123

Feature: document_analysis

Model: premium_model

Input tokens: 4,200

Output tokens: 650

Estimated cost: $0.036

The usage data explains what happened.

The customer and feature identifiers explain where the cost belongs.

The estimated cost explains the financial impact.

Token metering is commonly used to capture the input and output token consumption behind LLM requests.

LLM cost tracking then connects that usage to the corresponding model cost.

Cost attribution adds another layer by connecting the cost to the relevant customer, feature, workspace, or workflow.

AI cost attribution vs LLM cost tracking

The terms are closely related but describe different parts of the process.

LLM cost tracking answers:

How much did our model usage cost?

AI cost attribution answers:

Who or what generated that cost?

For example:

LLM cost tracking:

Total model cost = $8,000

AI cost attribution:

Customer A = $1,200

Customer B = $600

Document analysis = $3,400

AI assistant = $2,100

Other features = $2,500

Cost tracking gives visibility into spend. Attribution gives that spend business context.

For a deeper explanation, see AI Cost Attribution: How to Know Which Customers and Features Drive Your LLM Spend.

AI cost attribution and pricing

Cost attribution becomes especially valuable when designing AI pricing.

If one feature consistently creates a large share of model spend, a company may decide to:

  • Increase its credit cost
  • Restrict it to higher plans
  • Add a usage quota
  • Charge overages
  • Use a cheaper model
  • Optimize the workflow

Usage quotas can prevent unusually heavy consumption from creating uncontrolled cost.

Usage-based billing can also allow revenue to increase when customer consumption increases.

The important distinction is that cost attribution measures what usage costs the company, while billing determines what the customer should actually pay.

Not every cost-generating event needs to be billable.

Common AI cost attribution mistakes

One common mistake is tracking only the monthly provider invoice. It tells the company total spend but not which customers or features generated it.

Another is recording tokens without customer or feature identifiers. Once that context is missing, reliably reconstructing it later can be difficult.

Teams may also rely too heavily on averages. An average customer cost of $20 can hide a small group of customers costing hundreds of dollars each.

Another mistake is mixing internal activity with customer usage. Development, testing, demos, and administrative usage should be identified separately.

Finally, cost and billable usage should not be treated as the same thing. Free trials, failed requests, promotional credits, and internal operations can create real cost without being charged to customers.

How MetricaOS helps

MetricaOS helps AI product teams connect usage and cost to customers, users, features, models, providers, and workflows.

This gives teams visibility beyond the provider invoice and helps them understand where AI spend actually originates.

With customer and feature-level cost attribution, AI companies can identify expensive usage, evaluate margins, improve pricing, set better limits, and make more informed product decisions.

For AI products, knowing total model spend is useful. Knowing exactly what created that spend is far more actionable.

Cost per Customer

· · Leave a Comment

What is cost per customer?

Cost per customer is the amount a company spends to serve an individual customer over a specific period.

In AI products, this usually includes the cost of model usage, tokens, AI requests, file processing, infrastructure, storage, and other product resources consumed by that customer.

For example, an AI SaaS company may have three customers on the same $99/month plan:

Customer A costs $12/month to serve
Customer B costs $38/month to serve
Customer C costs $145/month to serve

Even though all three customers pay the same price, they do not create the same cost.

This is why cost per customer is especially important for AI products. AI usage can vary heavily between customers, even when they are on the same plan.

Why cost per customer matters for AI products

AI products often have variable costs. Every prompt, response, summary, document analysis, workflow, or agent action may create real provider cost.

If a company only looks at total monthly AI spend, it may miss which customers are profitable and which customers are creating margin pressure.

For example:

Monthly subscription revenue from a customer: $99

Monthly LLM cost from that customer: $140

Gross margin before other costs: negative

That customer may look healthy from a revenue perspective, but they are actually unprofitable.

Tracking cost per customer helps AI teams answer questions like:

  • Which customers are the most expensive to serve?
  • Which customers are profitable?
  • Which plans are underpriced?
  • Which features create the most cost?
  • Should heavy users be moved to a higher plan?
  • Should usage quotas or overages be added?

This is closely related to LLM cost tracking, because teams need to understand model spend before they can calculate customer-level cost.

What should be included in cost per customer?

For AI products, cost per customer should include the product costs that increase with customer usage.

Common cost components include:

LLM provider cost

Input token cost

Output token cost

Embedding cost

Image generation cost

Speech-to-text or transcription cost

AI workflow execution cost

Vector database usage

Storage

Compute

API usage

The most important cost component is usually LLM or model provider cost, especially for products built on OpenAI, Anthropic, Google, Azure OpenAI, or other AI providers.

But cost per customer should not only track the provider bill. It should connect that cost to product context, such as the customer, user, workspace, feature, model, and billing period.

That usually requires AI usage metering and token metering at the product level.

Cost per customer and token usage

In many AI products, cost per customer is calculated from token usage.

A simple usage event may show:

Customer: cus_123

Feature: document_summary

Input tokens: 4,200

Output tokens: 650

Model: gpt-4.1

Estimated cost: $0.036

One event may look small. But across thousands of requests, the total cost can become meaningful.

For example:

Customer: cus_123

Monthly requests: 18,000

Total tokens: 52 million

Estimated LLM cost: $410

Plan revenue: $299

In this case, the customer may be using the product heavily enough to create negative margin.

This is why token metering is important. It gives the raw usage data needed to calculate customer-level cost.

Cost per customer vs average cost per customer

Cost per customer looks at the cost of serving each individual customer.

Average cost per customer looks at total cost divided across all customers.

For example:

Total monthly AI cost: $10,000

Total customers: 500

Average cost per customer: $20

That average may look healthy.

But the individual breakdown may show something very different:

100 customers cost less than $5 each

350 customers cost around $20 each

50 customers cost more than $100 each

The average hides the problem.

For AI products, individual customer-level cost is often more useful than average cost because usage patterns are uneven. A few heavy customers can create a large portion of total AI spend.

How cost per customer affects pricing

Cost per customer helps teams design better pricing.

If customers on a plan consistently cost more than expected, the company may need to adjust the plan structure.

Possible changes include:

  • Lower included usage
  • Add usage quotas
  • Introduce overage billing
  • Move heavy users to higher plans
  • Change credit consumption rates
  • Use cheaper models for some actions
  • Limit expensive workflows

For products using credit-based pricing, cost per customer helps validate whether credit allowances are sustainable.

For example, if a Pro plan includes 20,000 credits, the team needs to know whether those credits usually cost $5, $50, or $500 to deliver.

Cost per customer also supports usage-based billing, because billing rules should reflect the real cost of customer consumption.

Example of cost per customer

Imagine an AI legal document product with three pricing plans.

Starter: $49/month

Pro: $199/month

Business: $599/month

A customer on the Pro plan uploads long legal documents and runs high-volume analysis every day.

At the end of the month, the product team sees:

Subscription revenue: $199

LLM cost: $260

Storage and processing cost: $35

Total cost to serve: $295

That customer costs more to serve than they pay.

The company may need to add a usage quota, move the customer to a higher plan, charge for additional usage, or optimize the feature to use a lower-cost model.

Without customer-level cost tracking, this issue may remain hidden inside the total provider bill.

Common mistakes with cost per customer

A common mistake is only tracking total AI spend. Total spend shows how much the company paid, but not which customers created the cost.

Another mistake is using averages too early. Average cost per customer may look fine while a small group of heavy users is unprofitable.

Some teams also track usage but do not convert it into cost. Usage data is useful, but pricing and margin decisions need cost data too.

Another mistake is not separating billable usage from internal usage. Testing, demos, admin actions, and development traffic should not be mixed with customer cost.

Finally, some teams ignore cost per customer until pricing breaks. By then, customers may already be used to generous limits, unlimited plans, or underpriced AI features.

How MetricaOS helps

MetricaOS helps AI product teams track usage and cost at the customer level.

With MetricaOS, teams can connect usage events to customers, users, features, models, tokens, credits, quotas, and estimated cost.

This helps AI companies understand which customers are profitable, which features create cost, and where pricing needs to change.

For AI products, cost per customer is not just a finance metric. It is a product, pricing, and margin metric.

LLM Cost Tracking

· · Leave a Comment

What is LLM cost tracking?

LLM cost tracking is the process of measuring how much an AI product spends on large language model usage.

In AI products, every model request can create cost. A user may send a prompt, generate a response, summarize a document, analyze a file, run an AI workflow, or trigger an agent. Each of these actions may consume input tokens, output tokens, or other model-related resources.

LLM cost tracking helps product and engineering teams understand where that cost is coming from.

For example, instead of only seeing a monthly provider bill, a team can answer:

  • Which customers are generating the most LLM cost?
  • Which features are the most expensive to run?
  • Which model is driving the highest spend?
  • How much does each AI workflow cost?
  • Is usage growing faster than revenue?

This is closely connected to AI usage metering, because teams need accurate usage data before they can understand cost clearly.

Why LLM cost tracking matters

Traditional SaaS products usually have relatively stable software costs. AI products are different.

When an AI product uses OpenAI, Anthropic, Google, Azure OpenAI, or other model providers, product usage can directly increase the company’s cost. A single customer using long prompts, large documents, frequent generations, or expensive models can create much higher cost than another customer on the same plan.

Without LLM cost tracking, teams may only notice the problem when the monthly provider invoice arrives.

That is too late.

LLM cost tracking helps teams protect margin, design better pricing, set usage limits, and understand whether specific customers or features are profitable. It also helps teams avoid offering “unlimited AI” without knowing the real cost behind that promise.

For AI products using usage-based billing, LLM cost tracking is especially important because pricing needs to be connected to actual consumption.

What should LLM cost tracking measure?

LLM cost tracking should not stop at total monthly spend.

A useful cost tracking setup should break cost down by the dimensions that matter to the business.

Common dimensions include:

Customer
User
Workspace or team
Feature
Model
Provider
Token usage
Request type
Workflow
Billing period

For example, a product team may want to know that Customer A spent $42 in model cost this month, mostly through the document analysis feature, using a higher-cost model.

That is much more useful than simply knowing the company spent $3,000 on LLM usage this month.

Good cost tracking connects model usage to product context.

LLM cost tracking and token metering

Most LLM cost tracking depends on token metering.

Tokens are the basic unit many language models use to calculate usage. Input tokens come from the prompt, context, uploaded text, or conversation history. Output tokens come from the model’s response.

A simple model request may include:

Input tokens: 3,000

Output tokens: 600

Model: gpt-4.1

Customer: cus_123

Feature: contract_summary

Estimated cost: $0.027

Token metering tells the team how much usage happened. LLM cost tracking converts that usage into cost.

Even if customers do not see tokens directly, the company still needs to understand token-level cost internally. This is especially important when using credit-based pricing, because credits need to map back to real cost in a sustainable way.

LLM cost tracking vs AI usage metering

LLM cost tracking and AI usage metering are related, but they are not exactly the same.

AI usage metering tracks what customers consume.

LLM cost tracking focuses on what that consumption costs the company.

For example, AI usage metering may record:

Customer A generated 10,000 AI summaries this month.

LLM cost tracking adds:

Those summaries created $380 in model cost.

Both are needed.

Usage data helps teams understand product activity. Cost data helps teams understand margin. Together, they help AI companies decide how to price plans, set usage quotas, control abuse, and decide when customers should upgrade.

Example of LLM cost tracking

Imagine an AI support product with three AI features:

Reply generation

Ticket summarization

Customer sentiment analysis

At the end of the month, the provider bill shows:

Total LLM spend: $5,000

That number alone is not enough.

With LLM cost tracking, the team may discover:

Reply generation: $1,200

Ticket summarization: $2,900

Customer sentiment analysis: $900

Then they may go deeper:

Customer A: $800

Customer B: $120

Customer C: $40

This changes how the team thinks about pricing.

If Customer A is paying $99/month but creating $800 in model cost, that customer is unprofitable. The company may need better limits, higher pricing, overages, or a different plan structure.

This is why LLM cost tracking supports both product decisions and billing decisions.

Common mistakes with LLM cost tracking

A common mistake is only looking at the provider invoice.

Provider invoices show total spend, but they usually do not explain which customer, feature, or workflow created the cost. This makes it hard to protect margin.

Another mistake is tracking tokens without converting them into cost. Token usage is useful, but business teams also need to understand money.

Some teams also forget to track model-level cost. If one feature uses a cheaper model and another uses a premium model, total token usage alone may not tell the full story.

Another mistake is not separating internal usage from customer usage. Testing, demos, admin activity, and development traffic should be separated from billable customer usage.

Finally, some teams start cost tracking too late. By the time AI usage grows, it becomes harder to reconstruct historical cost accurately.

How MetricaOS helps

MetricaOS helps AI product teams track usage and cost across customers, users, features, models, and providers.

For LLM cost tracking, MetricaOS gives teams a structured way to connect AI usage events to estimated cost, customer attribution, credits, quotas, and billing logic.

This helps teams move beyond provider invoices and understand the real economics of their AI product.

For AI companies, LLM cost tracking is not just an accounting task. It is the foundation for sustainable pricing, margin protection, and smarter product decisions.

Usage Quota

· · Leave a Comment

What is a usage quota?

A usage quota is a limit on how much of a product, feature, or resource a customer can use within a specific period.

In AI products, usage quotas are often used to control consumption of tokens, AI credits, model requests, API calls, documents processed, messages generated, or workflows completed.

For example, an AI SaaS product may offer:

Starter plan: 2,000 AI credits per month

Pro plan: 20,000 AI credits per month

Business plan: 100,000 AI credits per month

Each plan includes a specific amount of usage. Once the customer reaches that quota, the product may stop usage, show an upgrade prompt, charge overages, or allow continued usage under a fair usage policy.

Usage quotas are closely connected to AI usage metering, because teams need accurate usage data before they can enforce limits or track included usage.

Why usage quotas matter for AI products

Usage quotas matter because AI products often have variable costs.

Every prompt, response, model call, document summary, AI workflow, or generated output can create real infrastructure or provider cost. If customers can use AI features without limits, a small number of heavy users can create unexpected cost.

This is especially important when a product uses third-party model providers. One customer may use a few short AI responses per month. Another customer may process long documents, generate reports, and run high-volume workflows every day.

If both customers are on the same fixed-price plan, but one consumes much more AI usage, the company’s gross margin can become unpredictable.

  • Usage quotas help AI teams answer questions like:
  • How much AI usage is included in each plan?
  • When should a customer be asked to upgrade?
  • Should free trial users have lower limits?
  • Should heavy usage be blocked, throttled, or billed?
  • Are credits or tokens being consumed too quickly?

Without usage quotas, AI pricing can become difficult to control. This is why quotas often sit between credit-based pricing, token metering, and usage-based billing.

Common types of usage quotas

AI products can define quotas in different ways depending on the product and pricing model.

Common quota types include:

Token quotas
Credit quotas
Request quotas
Document quotas
Message quotas
Image generation quotas
Transcription minute quotas
Workflow execution quotas
API call quotas

A developer-focused AI platform may use token quotas or API request quotas. A business-facing AI SaaS product may use credit quotas because customers often understand credits more easily than raw tokens.

For example:

Token quota:
1 million tokens per month

Credit quota:
20,000 AI credits per month

Request quota:
10,000 AI requests per month

Document quota:
500 documents processed per month

The best quota depends on what customers understand and what the company needs to control internally.

Usage quota vs usage limit

A usage quota and a usage limit are closely related, but they are not always the same.

A usage quota usually defines the amount of usage included in a plan.

A usage limit defines what happens when the quota is reached.

For example:

Quota:
20,000 AI credits per month

Limit:
Stop usage after 20,000 credits unless the customer upgrades

Or:

Quota:
20,000 AI credits per month

Limit:
Allow continued usage and charge overages

This distinction matters because two products can have the same quota but very different user experiences after the quota is reached.

One product may block usage immediately. Another may send alerts. Another may allow overages. Another may throttle usage. Another may ask the customer to upgrade.

The quota defines the allowance. The limit defines the enforcement behavior.

Hard quotas vs soft quotas

AI products usually handle quotas in two main ways: hard quotas and soft quotas.

A hard quota stops usage when the customer reaches the limit.

For example:

You have used all 10,000 AI credits for this month.

Upgrade your plan or buy more credits to continue.

Hard quotas are useful when cost control is important. They are common for free plans, trials, prepaid credits, and products where usage can become expensive quickly.

A soft quota allows customers to continue using the product after reaching the quota, but may trigger alerts, upgrade prompts, admin notifications, or overage billing.

For example:

You have used 90% of your monthly AI credits.

Additional usage may be billed as overage.

Soft quotas are often better for B2B products where interrupting customer workflows can create frustration.

The right choice depends on the product, customer type, pricing model, and cost exposure.

Usage quotas and AI credits

Many AI SaaS products use quotas through credits.

Instead of saying:

Your plan includes 1 million tokens per month.

the product may say:

Your plan includes 20,000 AI credits per month.

This is easier for many customers to understand.

Different AI actions can consume different numbers of credits:

Generate a short reply: 5 credits
Summarize a document: 50 credits
Analyze a long report: 200 credits
Run an AI workflow: 500 credits

Behind the scenes, the company may calculate those credit values based on token usage, model cost, workflow complexity, and desired margin.

This is why usage quotas often work together with credit-based pricing. Credits give customers a simpler usage unit, while metering helps the company track what is actually being consumed.

Usage quotas and token metering

If an AI product uses LLMs, usage quotas often depend on token metering.

Token metering tracks how many input and output tokens are consumed by users, customers, features, models, or workflows.

This matters because quota usage needs to be based on accurate consumption.

For example, if a plan includes 1 million tokens per month, the system needs to track:

Input tokens
Output tokens
Total tokens
Customer ID
User ID
Feature
Model
Timestamp
Billable status

Even if customers see credits instead of tokens, token metering may still be used internally to calculate how many credits should be deducted.

A product may show:

This document summary used 50 credits.

But internally, the system may know:

Input tokens: 4,200
Output tokens: 650
Estimated cost: $0.036
Credits deducted: 50

This allows the company to keep customer-facing pricing simple while still protecting margins internally.

What happens when a customer reaches a usage quota?

When a customer reaches a usage quota, the product needs a clear policy.

Common options include:

  • Block further usage
  • Ask the customer to upgrade
  • Allow the customer to buy more credits
  • Charge overage fees
  • Throttle usage
  • Notify an admin
  • Allow temporary grace usage
  • Reset usage in the next billing period
  • Each option has tradeoffs.

Blocking usage protects cost, but can interrupt the customer. Overage billing protects revenue, but can create billing surprises if not communicated clearly. Upgrade prompts can work well, but only if customers understand why they hit the limit.

For AI products, transparency is important. Customers should be able to see how much usage they have consumed, how much remains, and what happens after they reach the quota.

A good usage dashboard may show:

Monthly credits included: 20,000
Credits used: 14,500
Credits remaining: 5,500
Billing period ends: July 31
Current usage: 72.5%

This helps customers manage usage before they hit a limit.

Usage quotas and usage-based billing

Usage quotas are often part of usage-based billing.

A product may include a certain amount of usage in each plan and then charge customers for additional usage.

For example:

Pro plan: $99/month

Includes: 20,000 AI credits

Overage: $10 per additional 10,000 credits

This model gives customers predictable base pricing while allowing revenue to scale with heavier usage.

It also protects the company from high-usage customers consuming unlimited AI resources without paying more.

Usage-based billing works best when quotas are clear, usage is visible, and customers understand what happens when they exceed their included usage.

For a deeper explanation of how usage connects to pricing, read Usage-Based Billing for AI Products: How to Price AI Features Without Losing Margin.

Example of a usage quota in an AI product

Imagine an AI customer support product that helps teams generate replies and summarize tickets.

The pricing might look like this:

Starter: 2,000 AI credits per month

Pro: 20,000 AI credits per month

Business: 100,000 AI credits per month

Different actions may consume credits:

Generate short support reply: 5 credits

Summarize ticket thread: 20 credits

Generate long response: 40 credits

Analyze customer history: 75 credits

If a Pro customer uses 18,000 credits in the first three weeks of the month, the product may show an alert:

You have used 90% of your monthly AI credits.

Upgrade your plan or buy more credits to avoid interruption.

This gives the customer time to act before hitting the quota.

For the company, this protects cost and creates a natural upgrade path.

Common mistakes with usage quotas

A common mistake is setting quotas before understanding real usage patterns.

If the quota is too low, customers may feel restricted. If the quota is too high, the company may lose margin. AI teams should ideally use real usage data before finalizing quotas for each plan.

Another mistake is using the wrong quota unit. Tokens may be accurate, but they may not be easy for business users to understand. Credits may be simpler, but they need a reliable metering system underneath.

Some teams also forget to separate internal usage from customer usage. Development, testing, demos, and admin activity should not usually consume a customer’s quota.

Another mistake is not warning customers before they hit the limit. If customers only discover the quota after something breaks, the experience feels bad.

A deeper mistake is offering “unlimited AI” without internal controls. Even if the pricing page says unlimited, the business still needs internal usage visibility, abuse detection, fair usage policies, and cost monitoring.

This is one of the reasons AI teams need AI usage metering before pricing and billing become difficult to manage.

How MetricaOS helps

MetricaOS helps AI product teams track usage across customers, users, features, models, and providers.

For usage quotas, MetricaOS gives teams the metering foundation needed to understand how much each customer has consumed, how usage maps to credits or tokens, and when customers are approaching plan limits.

With MetricaOS, AI teams can connect usage data to pricing, quotas, credits, and billing decisions.

For AI products, usage quotas should not be guessed. They should be based on clear usage data, customer-level attribution, and a reliable metering layer.

Usage-Based Billing

· · Leave a Comment

What is usage-based billing?

Usage-based billing is a pricing model where customers are charged based on how much of a product or service they use.

Instead of paying only a fixed monthly subscription, customers pay according to actual consumption. In AI products, this usage may be measured through tokens, AI credits, model requests, API calls, documents processed, messages generated, or workflows completed.

For example, an AI product may charge customers based on:

Number of AI credits used

Input and output tokens consumed

Documents summarized

AI replies generated

Agent workflows completed

API requests made

Usage-based billing is closely connected to AI usage metering, because a company needs reliable usage data before it can bill customers accurately.

Why usage-based billing matters for AI products

Usage-based billing matters because AI products often have variable costs.

In traditional SaaS, two customers on the same plan may cost roughly the same to serve. But in AI SaaS, one customer may use a few short AI responses while another customer may process long documents, run multi-step workflows, or generate thousands of responses.

Both customers may pay the same subscription fee, but their cost to serve can be very different.

This is why AI companies need to understand usage before pricing becomes messy. A customer who uses more AI may create more provider cost, more infrastructure cost, and more margin pressure.

Usage-based billing helps companies connect pricing to actual consumption.

It can help AI teams answer questions like:

  • Should heavy users pay more?
  • Should each plan include a usage quota?
  • Should usage be measured in tokens, credits, or requests?
  • Should customers be charged for overages?
  • Are some customers unprofitable because of high AI usage?

This is one reason usage-based billing often works together with token metering and customer-level usage tracking.

Usage-based billing vs subscription billing

Subscription billing charges a fixed recurring amount.

Example:

Pro plan: $99/month

Business plan: $299/month

Usage-based billing charges based on consumption.

Example:

$99/month including 10,000 AI credits

Additional usage billed at $10 per 10,000 credits

Many AI products use a hybrid model: a base subscription plus included usage, with extra usage billed separately or handled through upgrades.

This is often easier for customers than pure usage-based pricing and safer for companies than unlimited AI usage.

Common usage units in AI products

AI products can measure usage in different ways.

Common usage units include:

Tokens

AI credits

API calls

Model requests

Messages generated

Documents processed

Images generated

Minutes transcribed

Agent runs

Workflow executions

The right usage unit depends on the product.

A developer-focused AI platform may use token-based billing. A business-facing AI SaaS product may use credit-based pricing because credits are easier for customers to understand than raw tokens.

Internally, the company may still track tokens and model costs. Externally, it may show customers a simpler credit balance.

What makes usage-based billing difficult?

Usage-based billing sounds simple, but it requires accurate metering.

Before charging customers based on usage, a product needs to know:

  • Who used the AI feature
  • Which customer or workspace the usage belongs to
  • Which model or provider was used
  • How many tokens or credits were consumed
  • Whether the usage should be billable
  • Whether the request succeeded or failed
  • Whether the usage was internal, free, trial, or paid

This is where many teams run into problems.

A provider dashboard may show total AI spend, but it may not explain which customer, user, feature, or workflow caused that spend. That makes billing and pricing decisions harder.

For usage-based billing to work well, teams need a proper metering layer before the billing layer.

Example of usage-based billing in an AI product

Imagine an AI document analysis product.

The company may offer:

Starter: $29/month with 2,000 AI credits

Pro: $99/month with 20,000 AI credits

Business: $299/month with 100,000 AI credits

Different actions consume different credits:

Short document summary: 25 credits

Long document summary: 100 credits

Contract analysis: 250 credits

Bulk document workflow: 1,000 credits

Behind the scenes, the company may calculate those credits based on token usage, model cost, workflow complexity, and desired margin.

This allows the customer to understand usage in a simple way while the company still protects its AI margins.

Usage-based billing and AI usage metering

Usage-based billing depends on AI usage metering.

Metering answers:

  • What was used?
  • Who used it?
  • How much was used?
  • What did it cost?
  • Should it count toward billing?

Billing answers:

How much should the customer be charged?

If the metering layer is weak, the billing layer becomes unreliable.

This is why AI teams should usually solve metering before they design complex usage-based pricing or overage billing.

For a deeper explanation, read Usage-Based Billing for AI Products: How to Price AI Features Without Losing Margin.

Common mistakes

A common mistake is trying to build usage-based billing directly from raw provider invoices. Provider invoices show total spend, but they do not always provide clean customer-level attribution.

Another mistake is using a customer-facing usage unit that users do not understand. Tokens may be accurate, but credits, requests, documents, or workflows may be easier for non-technical customers.

Teams also sometimes fail to separate raw usage from billable usage. Not every model call should necessarily be billed. Internal testing, failed requests, free trial usage, and promotional credits may need different treatment.

A deeper mistake is offering unlimited AI without internal usage visibility. Unlimited pricing can work only when the company has strong monitoring, fair usage limits, and margin controls.

How MetricaOS helps

MetricaOS helps AI product teams track usage across customers, users, models, providers, and features.

For teams planning usage-based billing, MetricaOS provides the metering foundation needed to understand consumption before it reaches the invoice.

With MetricaOS, AI teams can track usage, attribute costs, monitor customer consumption, and prepare for pricing models based on credits, quotas, or usage-based billing.

For AI products, usage-based billing should not start with the invoice. It should start with reliable usage metering.