Agentic AI SaaS Pricing: How to Manage Token Costs Without Losing Margin
12-minute read
There's a question that comes up whenever someone builds a digital product powered by AI agents: How should you price it?

Contents
- Per-seat pricing no longer tells the whole story
- The variable cost of inference
- Three main approaches to agentic AI pricing
- Why a hybrid pricing architecture is often the safest choice
- The Salesforce Agentforce example
- Compare prices without mixing units
- Enterprise contracts aren't universal benchmarks
- Practical ways to protect your gross margin
- Define success in writing
- A forecast for adoption
- Move beyond the pricing dilemma: design the system
- Sources
Behind it is a deeper concern: what if your product grows its customer base while quietly destroying its own profitability?
Pricing isn't just a table of plans. It's a business architecture decision. It defines the value you promise, how you measure usage, which risks you assume, and who absorbs fluctuations in AI inference costs.
Per-seat pricing no longer tells the whole story
For years, software-as-a-service (SaaS) products were mainly priced per user. That made sense: more people accessing the software generally meant more collaboration, more usage, and more value.
But AI agents change that relationship. An agent can complete tasks, query systems, generate documents, or handle customer conversations without requiring another human seat.
Per-seat pricing still makes sense for access, permissions, collaboration, and platform capacity. It just doesn't always reflect the work performed by agents.
That's why vendors increasingly combine a base platform or user fee with measured consumption: actions, credits, conversations, or outcomes.
The shift matters. Customers aren't only paying to access an application; they're also paying for work the system performs.
Pricing an AI agent solely by seats is a little like charging a fleet of self-driving cars according to how many parking spaces it uses.
The questions become:
- How many workflows are completed?
- How many actions does the agent perform?
- What outcome does it produce?
- What variable cost does each operation generate?
- How much economic value does the customer capture?
The variable cost of inference
Traditional SaaS benefits from relatively favorable economics: the product is developed once, and serving an additional customer often has a low marginal cost.
In an agentic system, every task may trigger:
- Model calls.
- Reasoning steps.
- Tool calls.
- Searches and context retrieval.
- Retries.
- Long responses.
- Validation steps.
- Escalations to human staff.
This introduces a significant cost driver: inference.
In agentic SaaS, gross margin is often influenced not just by the fixed cost of operating software but by variable inference, tooling, and oversight costs.
And token prices are only one piece of that equation. You also need to account for context length, steps per task, model selection, cache hit rates, external tools, errors, retries, and handoffs to people.
Token prices are falling, but consumption can rise
Inference has become dramatically cheaper for some capability levels. The AI Index documented a decline of more than 280-fold in inference cost for GPT-3.5-level performance between November 2022 and October 2024 (source).
That doesn't mean every model, provider, or performance tier got cheaper at the same rate.
Nor does it mean your application's total bill will decline. A multistep agent workflow may use many more tokens than a traditional chat interaction because of longer contexts, tool calls, retries, and validation.
The economics depend on how quickly each factor changes:
- Price per token.
- Number of tasks.
- Complexity of tasks.
- Steps per task.
- Average context length.
Pricing on the assumption that model costs will keep falling is risky. Higher consumption can easily outweigh a lower unit price.
Three main approaches to agentic AI pricing
The most useful distinction is between three billing units.
Token-based pricing
Customers pay for model usage, usually measured in input and output tokens.
Advantages
- Better aligns variable revenue with variable cost.
- Helps protect the provider's gross margin.
- Is straightforward to calculate internally.
- Allows usage increases to flow through to customer bills.
Challenges
- Many business buyers don't know what a token is.
- Monthly spending may be hard to forecast.
- Technical decisions outside the buyer's control can affect their bill.
- The same task can cost different amounts depending on the model and context length.
This can work for developers and technical teams, but tokens are often a poor primary unit for nontechnical business buyers.
Task- or action-based pricing
The provider charges for a defined operation, such as classifying a document, generating a proposal, updating a record, or executing a workflow.
This is easier for customers to understand because it resembles the work they're purchasing. But an action isn't the same as an outcome.
An action is an operation the agent performs. A resolution is a result defined by the vendor. One successful resolution may require several actions.
That's why $0.10 per action cannot be compared directly with $0.99 per resolution or $2 per conversation.
Outcome-based pricing
Customers pay when a defined measure of success is achieved, such as:
- Meetings booked.
- Invoices collected.
- Support tickets resolved.
- Requests completed.
- Documents approved.
- Incidents closed.
This approach can align vendor and customer incentives, but only if success is defined precisely.
Your terms need to specify:
- Which event triggers billing.
- What evidence counts as success.
- How abandoned cases are treated.
- What happens when a human takes over.
- How disputes are handled.
- What constitutes partial success.
- The time window used to attribute outcomes.
These definitions are part of the product itself. A “resolution” might mean a customer explicitly confirmed that their issue was solved—or it could rely on an indirect signal, like the customer no longer replying.
So charging per outcome doesn't automatically create better incentives. The definition and measurement of that outcome are what matter.
Why a hybrid pricing architecture is often the safest choice
The market isn't uniformly replacing per-seat SaaS with pure outcome-based pricing.
A survey cited by Kyle Poyar of more than 230 B2B software and AI companies identified hybrid pricing as a common structure: seat-based or platform subscriptions combined with usage, credits, or AI consumption (source).
The reasoning is straightforward:
- A base fee covers access, platform capacity, and fixed costs.
- Included usage creates predictability.
- Overage charges reflect intensive consumption.
- Caps and alerts help avoid surprise bills.
- The provider keeps revenue more closely aligned with variable costs.
A typical offer might include a platform fee, a specified number of users or permissions, included tasks or credits, published overage rates, usage or rate limits, an annual commitment option, and surcharges for premium models or complex workflows.
The included allowance should cover ordinary use. If it's so small that every month becomes a dispute about overages, the model isn't serving anyone well.
The Salesforce Agentforce example
Salesforce illustrates how the market is experimenting with different billing units.
Agentforce has published several pricing approaches:
- Per conversation.
- Flex Credits for actions.
- User licenses.
- Platform and complementary product options.
Salesforce lists $500 per 100,000 Flex Credits. Its reference schedule assigns 20 credits to a standard action, equivalent to $0.10 per action. It also lists $2 per conversation for certain use cases (Salesforce Help; Agentforce pricing).
In its fiscal 2026 fourth-quarter results, released February 25, 2026, Salesforce reported:
- $800 million in Agentforce annual recurring revenue.
- 169% year-over-year growth.
- More than 29,000 deals closed.
- 2.4 billion Agentic Work Units delivered.
- Nearly 20 trillion tokens processed cumulatively (investor release).
Salesforce also said that more than 60% of fourth-quarter Agentforce and Data 360 bookings came from existing-customer expansion. Because that statement combines both products, it shouldn't be attributed to Agentforce alone.
These results show commercial traction. They don't prove that Agentforce is cannibalizing per-seat revenue. That's a possible strategic hypothesis, not a conclusion established by Salesforce.
An AI agent might replace some licenses, expand a customer's contract, or do both. The outcome depends on the use case and the way the product is packaged.
Compare prices without mixing units
Published market prices are useful orientation points, but they aren't necessarily comparable.
Provider or model | Billing unit | Published reference |
|---|---|---|
Salesforce Agentforce | Flex Credits | $500 per 100,000 credits; a standard action may consume 20 credits (Salesforce Help) |
Salesforce Agentforce | Conversation | $2 per conversation for the published option (Salesforce) |
Intercom Fin | Outcome or resolution | $0.99 per outcome, according to Intercom's published pricing (Intercom) |
Zendesk | Resolution or AI Agent plan | Roughly $1.20–$2 depending on plan, contract, and structure; not a universal rate (Intercom; Agent Marketplace) |
Freshdesk | Session or interaction | Depends on the product, plan, and billing unit; $0.10 should not be treated as a universal benchmark |
A conversation can include several messages. A resolution can involve many actions. An action may produce no useful outcome at all.
A more meaningful comparison looks at the entire chain:
Cost per actual resolution
= actions per conversation
× conversations per resolution
× cost per actionA conversation-based plan may charge for unsuccessful conversations. With resolution-based pricing, the provider assumes more performance risk but has an incentive to define “resolved” carefully.
Enterprise contracts aren't universal benchmarks
Platforms such as Sierra, Decagon, and Ada may negotiate custom enterprise agreements that include:
- Implementation.
- Integration with internal systems.
- Security and compliance.
- Support.
- Committed volume.
- Customization.
- Monitoring.
- Human review.
- Service-level agreements.
It would therefore be misleading to present $50,000 to more than $600,000 per year as a public, directly comparable rate across those vendors.
The same caution applies to price ranges quoted in Spain. A range of €15 to €5,000 per month may combine self-service products, lightweight automations, usage-limited solutions, and bespoke enterprise deployments. It may offer commercial context but isn't a representative market benchmark.
Practical ways to protect your gross margin
Model routing
Not every request needs your most capable—and usually most expensive—model.
Smaller models may work well for classifications, structured extraction, routine answers, intent detection, and simple validations.
Reserve more powerful models for complex tasks, ambiguous decisions, and workflows requiring advanced reasoning.
The goal isn't simply to use the cheapest model. It's to send each task to the model that meets the required quality level at the lowest total cost.
Context caching
Agents frequently reuse stable context: system instructions, documentation, policies, and schemas.
Caching can lower the cost of repeated input tokens, but discounts depend on the model and provider. OpenAI initially documented a 50% discount on cached inputs for compatible models; some providers and models offer larger discounts, potentially reaching 90% under specific conditions (OpenAI).
Don't promise a blanket 90% reduction. Examine which prompt segments can be cached, cache duration, minimum sizes, prefix invalidation, any write charges, and actual cache-hit rates.
Batch and asynchronous processing
Tasks that don't require immediate responses can be handled in batches:
- Overnight data enrichment.
- Bulk classification.
- Offline scoring.
- Report generation.
- Document processing.
- Knowledge-base reindexing.
Some providers offer discounts around 50% for batch processing, subject to API conditions and delivery windows. Anthropic, for example, documents 50% discounts for its Batch API (Claude documentation).
Limits, alerts, and included allowances
Customers need predictability. A practical plan can provide:
- A generous included allowance.
- Alerts before limits are reached.
- Configurable usage caps.
- Approval for exceptional consumption.
- Published overage prices.
- The option to purchase additional capacity.
The goal isn't to conceal consumption. It's to prevent an unusual workflow from generating an unpredictable bill.
Workflow-level monitoring
Measuring the total API bill isn't enough. Track:
- Cost per task, customer, and resolution.
- Tokens per workflow.
- Average steps and retry rate.
- Escalations to humans.
- Gross margin by plan and segment.
Your dashboard should explain how much each workflow costs—not just what the model provider charged your account last month.
Define success in writing
In outcome-based pricing, the definition of success matters more than the advertised rate.
Contracts should state what counts as a resolution, what happens if a user abandons a case, how escalations work, whether reopened issues trigger a second charge, how AI-human contributions are attributed, what happens when the system fails, and what each party may audit.
The more ambiguous the outcome, the more likely billing disputes become.
A good success metric should be observable, auditable, relevant to the customer, difficult to manipulate, and stable enough to budget for.
A forecast for adoption
Gartner projected that 40% of enterprise applications would include task-specific AI agents by the end of 2026, compared with less than 5% in 2025 (Yahoo Finance; IT Brief).
That prediction concerns applications, not the share of companies that have deployed agents. Nor does it guarantee that every agent will have broad autonomy, create positive economic value, or use the same pricing model.
It's important to distinguish among an application with an agentic feature, an agent executing a limited workflow, a multiagent system working across tools, and a product priced directly on outcomes.
These involve different product capabilities, costs, and risks.
Move beyond the pricing dilemma: design the system
You won't solve agentic SaaS pricing by copying someone else's pricing page.
You solve it by designing the complete system:
- What does the product promise?
- Which unit of value does the customer understand?
- Which unit of cost can the provider control?
- How much usage is included?
- What happens when the customer exceeds it?
- How is success defined and verified?
- Which risks does each party absorb?
- How will you monitor margins?
- How will prices adapt as underlying models change?
One resilient approach combines a predictable base fee with a variable component tied to usage or value. The base fee supports the platform. An included allowance makes costs more predictable. Overage charges capture intensive use. Outcome-based billing can be reserved for workflows where success is genuinely measurable.
AI that serves people should also support sustainable business economics: charge for the value you actually deliver, using a model that sustains the company instead of undermining it.
If you're building an AI-powered digital business and already have technology, referrals, and plenty of ideas—but no clear monetization structure—that's the place to begin strategic work.
In a Structure and Clarity Session, we'll organize the model and define a concrete plan. The Digital Expansion Assessment helps evaluate where the business stands and choose a growth path. And when you're ready to build everything—brand, website, products, sales, content, and automations—the Complete Ecosystem turns your expertise into a business designed to scale.
Sources
- Why your AI product has SaaS pricing and utility-grade costs
- AI Is Eating Software Margins
- Your AI Feature Is Quietly Destroying Your Gross Margin
- Agentic AI Pricing Models: Token, Task and Outcome-Based
- Accounting for Outcome-Based Pricing in Agentic AI Software — Deloitte
- Pricing Models for Agentic AI Services: A Multi-Sided Market Analysis
- The 2026 State of B2B SaaS and AI Monetization
- Salesforce introduces outcome-based pricing for Agentforce AI
- Salesforce Agentforce Pricing: Full Cost Breakdown
- AI Agent Pricing Comparison 2026
- From seats to consumption — Flexera
- 2025 SaaS Pricing Report — Maxio
- How Much Does an AI Agent Cost in Spain?
- Why Per-Seat SaaS Pricing Doesn't Survive the AI Agent Era
Plan the next step for your business
In a free, 45-minute one-on-one video call, we’ll review where you are and identify where to start.


