There is no single best way to price an AI agent.
Some AI products charge for access. Others charge for tokens, requests, workflows, resolutions or business outcomes. Enterprise deals can combine several models in the same contract.
The right pricing model should answer four questions:
What does the customer value? What can you measure? What does delivery cost? Does the price leave enough margin?
This guide compares 20 ways to price AI agents, from traditional subscriptions to outcome-based and hybrid pricing.
20 AI Agent Pricing Models at a Glance
| # | Pricing Model | Customer Pays For | Best When | Main Risk |
|---|---|---|---|---|
| 1 | Subscription | Access over time | Usage is predictable | Heavy users reduce margin |
| 2 | Per-Agent | Each deployed AI agent | Agent represents a defined role | Workload varies by agent |
| 3 | Per-Seat | Human users | AI assists employees | AI activity grows without seat growth |
| 4 | Per-Token | Tokens consumed | Buyers understand model usage | Tokens may not reflect value |
| 5 | Per-Request | Requests | Interactions have similar economics | Request costs vary |
| 6 | Per-Tool-Call | Tool actions | Tools create measurable cost or value | Retries create leakage |
| 7 | Per-Workflow | Completed workflows | Work has a clear unit | Workflow complexity varies |
| 8 | Per-Resolution | Resolved cases | Resolution is easy to verify | Reopens and disputes |
| 9 | Per-Output | Accepted deliverables | Output is measurable | Regenerations add cost |
| 10 | Per-Outcome | Business results | Results can be attributed | Vendor absorbs failure cost |
| 11 | Credit-Based | Credits consumed | Many capabilities share one currency | Credit value becomes unclear |
| 12 | Wallet / Prepaid | Prepaid balance | Buyers want budget control | Expiry and rollover complexity |
| 13 | Minimum Commitment | Contracted minimum spend | Enterprise deals | Poor discounting hurts margin |
| 14 | Tiered Usage | Usage across tiers | Rates should decline gradually | Complex calculations |
| 15 | Volume | Total usage band | Large customers earn lower rates | Margin cliffs |
| 16 | Hybrid | Multiple pricing units | One metric cannot capture the economics | Operational complexity |
| 17 | Cost-Plus | Cost plus commercial return | Delivery cost varies materially | May underprice value |
| 18 | Token-to-Value | Value mapped from AI cost | Tokens matter internally, not to buyers | Cost-to-value mapping |
| 19 | Enterprise Custom | Negotiated terms | Large or complex deals | Contract complexity |
| 20 | Platform + Consumption | Access plus variable usage | Product has fixed and variable value | Included usage can be mispriced |
Why AI Agents Need Different Pricing Models
Traditional SaaS often connects revenue to access:
User → Seat → Subscription
AI agents can create value differently:
Token → Request → Tool Call → Workflow → Output → Outcome
The market already reflects this shift.
McKinsey's analysis of AI-native software companies found the following mix of pricing metrics:
| Pricing Meter | Share |
|---|---|
| Activity / Consumption | 40% |
| Flat Fee | 35% |
| Capacity | 15% |
| Successful Outcome | 10% |
There is no single dominant pricing model.
Buyer preferences are moving too. Simon-Kucher's 2026 AI pricing research found that 86% of surveyed buyers preferred usage- or outcome-based pricing for AI solutions over traditional seat-based structures.
The important distinction is:
The unit that creates cost does not have to be the unit the customer buys.
An agent may consume thousands of tokens internally but be sold as one completed workflow.
It may make several tool calls but be priced as one successful resolution.
It may perform dozens of actions before producing one business outcome.
AI pricing therefore needs to connect:
Customer Value → Billable Unit → Delivery Cost → Margin
The 6 Families of AI Agent Pricing
The 20 models become easier to understand when grouped by what the customer is actually buying.
1. Access & Capacity Pricing
Subscription
Customer pays for: access over time
Best for: predictable usage and recurring product value
Main risk: fixed revenue while AI consumption increases
Subscription pricing is still useful for AI. It becomes risky when one customer can consume significantly more model and tool resources than another while paying the same amount.
Per-Agent
Customer pays for: each AI agent deployed
Best for: agents representing defined business roles
Main risk: workload can vary significantly per agent
A support agent, research agent or accounts-payable agent can be sold as a defined unit even if the work performed underneath it changes.
Per-Seat
Customer pays for: human users
Best for: AI that primarily assists employees
Main risk: agent activity grows without seat growth
Per-seat pricing is not obsolete. It works when human access remains closely connected to the value delivered.
2. Consumption Pricing
Per-Token
Customer pays for: tokens consumed
Best for: developer products and model infrastructure
Main risk: tokens measure computation better than customer value
A common calculation is:
Input Tokens × Input Rate + Output Tokens × Output Rate
Per-Request
Customer pays for: eligible requests
Best for: interactions with reasonably similar economics
Main risk: one request can cost much more than another
This matters especially for autonomous agents.
McKinsey's 2026 analysis of agentic economics cites programming research in which the same task showed as much as a 30× difference in cost between separate completions.
An agent may choose a different path, call different tools, generate more tokens or retry before succeeding.
One request therefore does not always equal one predictable cost.
Per-Tool-Call
Customer pays for: eligible tool actions
Best for: tool-heavy agents, APIs and MCP-style products
Main risk: retries and internal calls become accidental billable usage
Revinci Bill explicitly describes native metering for tokens, actions and tool calls, alongside other AI usage signals.
3. Work-Based Pricing
Per-Workflow
Customer pays for: completed work
Best for: repeatable processes with clear completion criteria
Main risk: workflow complexity changes while price stays fixed
Examples include:
- lead researched,
- invoice processed,
- document reviewed,
- claim analysed.
Per-Resolution
Customer pays for: successfully resolved cases
Best for: AI customer-service agents
Main risk: defining what counts as resolved
A live example comes from Intercom. Fin's current outcome pricing includes $0.99 per resolution, alongside other qualifying outcome types.
The contract still needs clear rules for reopened cases, human escalations and disputed resolutions.
Per-Output
Customer pays for: accepted deliverables
Best for: products where the deliverable itself represents value
Main risk: regeneration cost
Outputs could include reports, images, analyses, documents or code artifacts.
The customer may pay for one accepted output even if several generations were required to create it.
4. Value & Economics Pricing
Per-Outcome
Customer pays for: verified business results
Best for: agents whose business value can be attributed clearly
Main risk: vendor absorbs the cost of failed attempts
Examples include qualified meetings, recovered payments, successful resolutions and completed transactions.
Simon-Kucher's AI pricing framework notes that output- and outcome-based models become more viable when business impact can be measured consistently.
The stronger the value alignment, the more important the attribution rules become.
Cost-Plus
Customer pays for: cost plus an agreed commercial return
Best for: highly variable AI delivery costs
Main risk: protecting margin without capturing enough customer value
For a target gross margin:
Required Price = Cost ÷ (1 − Target Gross Margin)
Illustrative example: cost per workflow = $40, target gross margin = 60%.
$40 ÷ 0.40 = $100
We at Revinci currently list our cost-plus margin product among the AI-native pricing approaches within Revinci Sell.
Token-to-Value
Customer pays for: a value-oriented unit while tokens remain an internal cost signal
Best for: products where customers should not have to understand raw LLM economics
Main risk: weak mapping between technical cost and customer value
For example:
Tokens → Cost → Completed Workflow → Customer Price
Instead of telling a customer:
You consumed 4 million tokens.
the commercial model might say:
You completed 500 research workflows.
Revinci explicitly describes token-to-value mapping within its AI-native pricing strategy.
5. Credits & Commitments
Credit-Based
Customer pays for: a common credit currency
Best for: products with many different AI capabilities
Main risk: customers cannot understand what a credit represents
For example:
| Activity | Credits |
|---|---|
| Basic action | 1 |
| Research workflow | 10 |
| Premium model task | 20 |
Credits let different activities use one commercial unit while the underlying cost can still vary.
Wallet / Prepaid
Customer pays for: usage from a prepaid balance
Best for: customers wanting stronger budget control
Main risk: expiry, rollover and refund complexity
A simple balance calculation is:
Opening Balance + Top-Ups − Usage = Remaining Balance
Revinci Bill describes multi-wallet support including grants, burns, expirations, rollovers and automatic top-ups.
Minimum Commitment
Customer pays for: at least an agreed level of spend
Best for: enterprise contracts
Main risk: discounts are agreed without understanding the cost of the committed workload
A simplified structure is:
Bill = Maximum of Minimum Commitment or Eligible Usage
depending on the contract.
Revinci Sell describes commitment management alongside prepaid drawdowns, overages and minimum-commit reconciliation.
6. Scale & Hybrid Pricing
Tiered Usage
Customer pays for: usage across graduated pricing tiers
Best for: gradually reducing unit rates at higher consumption
Main risk: complicated rate calculations
| Usage | Rate |
|---|---|
| First 10,000 units | $0.10 |
| Next 40,000 | $0.08 |
| Above 50,000 | $0.06 |
Volume Pricing
Customer pays for: usage according to the achieved volume band
Best for: high-volume customers
Main risk: sudden price changes at thresholds
Volume pricing can look similar to tiered pricing, but the invoice economics can be very different.
Hybrid Pricing
Customer pays for: multiple pricing components
Best for: AI products where one metric cannot capture the full economics
Main risk: unnecessary pricing and billing complexity
Examples include:
- Subscription + Usage
- Minimum Commitment + Overage
- Per-Agent + Workflow
- Base Fee + Outcome
Simon-Kucher's Global Software Study reports that 45% of companies plan to use two or more pricing metrics for their AI offerings.
Its examples include platform fee + usage credits, user fee + output fee and base agent fee + performance fee.
Enterprise Custom
Customer pays for: negotiated commercial terms
Best for: large enterprise contracts
Main risk: bespoke terms become difficult to operate
A single enterprise deal might combine:
- custom rate cards,
- minimum commitments,
- volume discounts,
- credits,
- entitlements,
- usage rules,
- outcome terms.
The challenge is not only negotiating these terms. They also need to move correctly from quote → usage → invoice.
Platform + Consumption
Customer pays for: platform access plus variable usage
Best for: products with both fixed and variable value
Main risk: too much included usage destroys the economics
A simple structure is:
Total Price = Platform Fee + Variable Consumption
For example: platform fee $3,000/month, included usage 50,000 units, additional usage $0.06/unit.
Google Cloud Marketplace's AI-agent pricing documentation explicitly supports this structure as combined pricing, where a base subscription is supplemented by usage charges.
How Do You Choose the Right AI Agent Pricing Model?
Start with five questions.
1. What Does the Customer Value?
| Customer Primarily Values | Models to Consider |
|---|---|
| Access | Subscription, Per-Seat, Per-Agent |
| Consumption | Token, Request, Tool-Call |
| Completed Work | Workflow, Resolution, Output |
| Business Result | Outcome |
| Budget Predictability | Credits, Prepaid, Commitment |
| Several of the Above | Hybrid |
Do not automatically price the easiest technical metric to meter.
2. What Drives Your Cost?
Identify what changes the actual cost of serving the customer:
Tokens + Model Choice + Tool Calls + APIs + Compute + Retries
This is especially important for autonomous agents.
McKinsey's agentic economics research notes that agents can take different execution paths for the same task, creating significant cost variability. In the programming research it cites, that difference reached as much as 30×.
If the customer-facing unit and the underlying cost driver are different, you need visibility into both.
3. Can You Reliably Measure the Billable Event?
Clear workflow completion? Workflow pricing may work.
Verified business result? Outcome pricing becomes possible.
Neither is reliable? A simpler consumption metric may be safer.
4. How Predictable Does the Customer's Bill Need to Be?
Higher predictability:
- Subscription
- Credits
- Prepaid
- Minimum Commitment
Closer consumption alignment:
- Token
- Request
- Tool-Call
- Workflow
Need both? Hybrid.
5. Does the Model Protect Margin?
Before finalising the pricing model, test what happens when:
- usage increases,
- expensive models are selected,
- workflows require more tools,
- agents retry,
- outcomes fail,
- enterprise discounts are added.
That leads to the most important check.
The Revinci Cost-to-Margin Test
When evaluating an AI pricing model, follow the commercial unit through four stages:
Price → Usage → Cost → Margin
| Question | Example |
|---|---|
| What do we charge? | $50 per resolution |
| What do we meter? | Verified resolutions |
| What creates cost? | Tokens + tools + failed attempts |
| What margin remains? | Revenue − attributable cost |
Illustrative example: 100 successful resolutions × $50.
Revenue = $5,000. Model and tool cost = $1,400. Failed-attempt cost = $450. Total attributable cost = $1,850.
Gross margin = $3,150. Gross margin percentage = 63%.
The customer may be buying resolutions while the infrastructure is consuming tokens, models and tools. Both sides of that equation matter.
That becomes particularly important when agent execution cost is variable rather than fixed. McKinsey describes agentic cost as behaving more like a distribution than a predictable per-task number.
Real AI Pricing Already Uses Multiple Models
This shift is already visible in the market.
Google Cloud Marketplace currently allows AI agents to use:
- subscription pricing,
- usage-based pricing,
- combined subscription + usage pricing.
Custom enterprise arrangements can also be handled through private offers.
Intercom Fin provides a live example of outcome-oriented pricing, including $0.99 resolution outcomes.
And McKinsey's AI-native software analysis shows the market split across consumption, flat-fee, capacity and outcome metrics rather than converging on one model.
AI pricing is becoming a portfolio of models, not a replacement of one model with another.
From Pricing Model to Revenue: How Revinci Operationalizes It
Choosing the pricing model is only the first decision.
It still needs to flow through:
Product → Price → Quote → Usage → Charge → Invoice → Cost → Margin
At Revinci we boast 20+ pricing models across our Agentic Revenue Platform.
The platform connects four parts of that process.
Sell — Configure the Commercial Model
Revinci Sell supports commercial configuration across products, pricing, commitments, discounts and entitlements, including AI-native approaches such as:
- outcome pricing,
- cost-plus margin,
- token-to-value mapping.
Bill — Meter and Rate the Billable Event
Revinci Bill handles usage signals including:
- subscriptions,
- tokens,
- actions,
- tool calls,
- workflows,
- outcomes,
- credits.
The goal is to turn the commercial rule into an auditable charge and invoice.
SmartCost — Attribute Delivery Cost
Revinci SmartCost is designed to attribute COGS across:
Customer → Agent → Workflow
including costs associated with tokens, compute, storage, APIs and infrastructure.
SmartMargin — Measure What Remains
Revinci SmartMargin connects attributable cost back to revenue so gross margin can be evaluated by customer, agent and deal.
Revinci's role is not to decide that every AI company should use the same pricing model.
It is to help operationalize the model that fits the commercial relationship while keeping the connection between:
Price → Usage → Cost → Margin
Which AI Agent Pricing Model Should You Start With?
| If You Need... | Start By Considering... |
|---|---|
| Simplicity | Subscription |
| An AI worker / digital role | Per-Agent |
| Technical consumption pricing | Token, Request or Tool-Call |
| Completed work | Workflow, Output or Resolution |
| Measurable business results | Outcome |
| Budget control | Credits, Prepaid or Minimum Commitment |
| Predictability + variable growth | Hybrid or Platform + Consumption |
There is no universally best model.
The strongest pricing model is one where the customer understands what they are buying, the business can measure it reliably, and the economics remain healthy as usage grows.
Frequently Asked Questions
What are AI agent pricing models?
AI agent pricing models determine what customers pay for when using an AI agent. The billable unit can be access, users, deployed agents, tokens, requests, tool calls, workflows, outputs, outcomes, credits or a combination of several units.
What is the best pricing model for AI agents?
There is no single best model. The right choice depends on customer value, the measurable billable unit, underlying AI costs, required spending predictability and target margin.
Should AI agents be priced per token or per outcome?
Token pricing aligns closely with technical consumption. Outcome pricing aligns more closely with customer value. Simon-Kucher's AI-agent pricing framework similarly places raw usage metrics toward the resource end of the pricing spectrum and outcomes closer to measurable customer value.
What is the difference between workflow and outcome pricing?
Workflow pricing charges for completing a defined unit of work. Outcome pricing charges for the business result produced by that work. A workflow can therefore be successfully completed without necessarily producing the desired outcome.
Can an AI product use more than one pricing model?
Yes. Hybrid models can combine structures such as platform fee + usage, minimum commitment + overage, subscription + credits, or base fee + outcome. Simon-Kucher reports that 45% of companies in its Global Software Study plan to use two or more pricing metrics for AI offerings.
Your first pricing model does not have to be your last. AI companies can begin with a subscription, introduce usage as consumption grows, add credits for flexibility and negotiate commitments or outcome pricing for enterprise customers.
The important thing is being able to move between those models without losing the connection between:
Price → Usage → Cost → Margin