.png)
Frameworks, core principles and top case studies for SaaS pricing, learnt and refined over 28+ years of SaaS-monetization experience.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Note: I can’t see an attached document in this chat. The draft below is therefore based on broader market research and a clear point of view: agentic AI should not primarily be priced like conventional SaaS seats or raw model usage. It should be priced around verified units of work, with outcome-based upside where value is measurable and attributable. Share the document and I can revise the article to reflect its specific language and POV precisely.
For two decades, enterprise software companies have relied on an elegantly simple pricing mechanism: charge per user, per month. It worked because software was generally a tool used by a person. More employees meant more licenses, more usage, and more value.
AI agents break that relationship.
An agent does not merely help a worker complete a task. Increasingly, it can receive an objective, reason through a workflow, use tools, retrieve information, make updates in systems of record, and hand back a completed result. The commercial question is no longer, “How many employees need access to this software?” It is, “What work was completed, what did it replace or improve, and what was that worth?”
That distinction is the foundation of agentic AI pricing.
The companies that treat agents as another feature and simply add a seat surcharge will get adoption, but may leave substantial value on the table—or create a pricing structure that collapses when one agent serves thousands of people. The companies that leap immediately to pure outcome pricing, meanwhile, may take on risks they cannot control: poor customer data, broken downstream processes, attribution disputes, and unbounded inference costs.
The best answer is neither traditional SaaS pricing nor simplistic “AI worker” pricing. It is a hybrid, value-linked model:
- Charge a platform fee for secure deployment, governance, integrations, observability, and administrative control.
- Meter verified units of work—completed resolutions, claims processed, records updated, code changes reviewed, invoices reconciled, qualified leads created.
- Add outcome-based pricing only when the outcome is measurable, attributable, and partly within the vendor’s control.
- Use commitments, rate limits, and cost guardrails to give customers budget certainty while protecting gross margins.
In short: do not charge for the agent’s existence. Charge for trusted work performed—and, where possible, for the business result that work creates.
An AI copilot is still fundamentally a software product. A human remains in charge, invokes it frequently, checks its work, and uses it to become more productive. Per-seat pricing can work well in that world. A lawyer using a drafting assistant, a developer using a coding companion, or a salesperson using an email-writing tool still experiences the product as a personal workspace.
An agent is different. It can operate in the background, serve many users, and execute workflows autonomously. A customer-service agent may resolve tens of thousands of issues without an employee ever opening the interface. A coding agent may generate pull requests overnight. A finance agent may reconcile transactions continuously. In each case, the number of licensed human users becomes a poor proxy for value.
That is why the market is already fragmenting into several models:
| Pricing model | What the customer pays for | Best fit | Main weakness |
|---|---|---|---|
| Per-seat | Access for each human user | Copilots and employee-facing assistants | Disconnects price from autonomous output |
| Usage-based | Tokens, minutes, API calls, compute, or credits | Infrastructure and highly variable workloads | Technical meter is hard for buyers to translate into value |
| Task-based | A completed job or workflow step | Repeatable, auditable operational work | Can reward vendors for volume rather than impact |
| Transaction-based | A business transaction processed | Payments, commerce, claims, bookings, logistics | Requires a clear transaction boundary |
| Outcome-based | A verified business result | Support resolution, recovered revenue, qualified lead | Attribution and vendor risk can be difficult |
| Hybrid | Platform access plus work and/or outcome charges | Most enterprise agent deployments | More complex to explain and administer |
The key is not choosing the most fashionable model. It is choosing the unit of value that most closely tracks the customer’s economic gain while remaining observable, auditable, and profitable for the vendor.
This matters because the upside is enormous but still theoretical in many deployments. McKinsey estimates that generative AI could add $2.6 trillion to $4.4 trillion in annual value across 63 use cases, with roughly three-quarters of that opportunity concentrated in customer operations, marketing and sales, software engineering, and R&D. Those are precisely the functions where agents can move from assisting people to completing work.
But potential value is not the same as captured value. Gartner has predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. Its warning should shape pricing strategy: if a buyer cannot connect the invoice to an economic result, the agent will become an expensive experiment rather than a durable budget line.
Per-seat pricing: useful for copilots, weak for autonomous labor
Per-seat pricing is not dead. It is simply no longer sufficient as the default model.
It remains attractive when an agent behaves like a productivity layer for a known population of employees. Seat pricing is easy to budget, easy for procurement to understand, and familiar to finance teams. It also aligns with value when a product’s main benefit is individual capability: better writing, faster analysis, easier document search, or assisted coding.
Salesforce, for example, offers an Agentforce User License at $5 per user per month for employee access, alongside other consumption-based and outcome-based options. That is sensible for broad employee enablement, where the product is still partly a workplace interface rather than a fully autonomous worker.
The problem emerges when an agent’s work is decoupled from the number of employees using it. Imagine an ecommerce support agent that handles one million customer conversations a month. Charging per support-manager seat would dramatically underprice the value created. Conversely, charging every employee for an internal procurement agent that only runs occasionally may overcharge for unused access.
Per-seat pricing also creates the wrong incentive for a truly autonomous system. A vendor’s revenue grows when customers add humans, while the customer’s value grows when the agent reduces human intervention. That is an economic contradiction.
Companies should therefore use seats as an access and governance fee, not as the primary monetization vehicle for autonomous work. Seats can pay for identity, permissions, audit trails, analytics, collaboration, and configuration. They should not be the main meter for an agent that independently completes high volumes of work.
Usage-based pricing: necessary under the hood, insufficient on its own
Usage pricing is the natural model for AI infrastructure because the underlying cost base is variable. An agent may use a few inexpensive model calls to answer a straightforward question, or consume substantial compute while planning, searching, using tools, processing documents, or recovering from errors.
Cognition’s Devin illustrates the logic. Its historical enterprise pricing has used Agent Compute Units, or ACUs, which reflect the complexity of actions, context gathering, browser and code execution, virtual-machine time, and related resources. Cognition explicitly frames ACUs as a measure of the work Devin performs, not simply a count of prompts.
In April 2026, Cognition updated Devin’s self-serve packaging to include free, Pro, Max, Teams, and Enterprise tiers. It said that compute-intensive features such as Ask Devin, DeepWiki, and Devin Review would be charged based on the underlying model cost of each run—an acknowledgement that sophisticated agentic work cannot always be economically supported by an all-you-can-eat subscription.
This is commercially rational. But raw compute is a poor customer-facing value metric. No CFO wants to explain why a customer-data cleanup project exceeded budget because the model needed more “thinking tokens,” retries, or context windows. The customer did not buy tokens. They bought clean records.
The answer is to separate the cost meter from the price meter. Vendors should absolutely track token consumption, tool calls, inference time, human-review rates, and failed runs internally. These metrics determine margins. But they should expose a business-readable meter externally whenever possible: per reconciled invoice, per resolved ticket, per reviewed pull request, per completed onboarding workflow.
In other words, compute should govern the vendor’s cost controls; work should govern the customer’s bill.
Task-based pricing: the practical center of gravity
For most enterprise agents today, the best pricing unit is neither a seat nor a final business outcome. It is a verified unit of completed work.
Task-based pricing works when a workflow has a discrete start and end, a clear definition of completion, and reliable instrumentation. Examples include:
Salesforce’s evolution with Agentforce is a useful example of this approach. The company initially emphasized $2 per conversation for certain customer-facing agent interactions, then introduced Flex Credits based on actions. Its current pricing offers $500 per 100,000 Flex Credits, with standard Agentforce actions consuming 20 credits, or $0.10 per action. Salesforce defines an action as a function such as updating a record, summarizing a case, answering a product inquiry, or executing a flow.
That move is important because it recognizes that a conversation is not a stable unit of work. One conversation may be a simple product question; another may require authentication, database retrieval, order updates, and escalation. By metering actions, Salesforce creates more granularity and gives customers a way to relate usage to operational activity.
But action pricing has limits. A vendor can inadvertently charge for every internal step of a workflow, turning a simple customer request into an opaque credit-burning exercise. The more autonomous an agent becomes, the less defensible it is to charge separately for every micro-action the vendor chose to perform.
The better task-based model charges for the customer-recognizable job, not every machine-level operation. A support agent might perform five actions internally, but the billable unit should be “case resolved” or “account change completed.” The vendor should bear the responsibility for making the agent efficient.
Transaction pricing: powerful where the workflow already has an economic unit
Transaction-based pricing is task-based pricing with a stronger commercial anchor. It works best where the business already counts activity in discrete units: payment processed, order placed, booking completed, loan application reviewed, insurance claim handled, shipment exception resolved, or compliance filing prepared.
This model is familiar because many industries already pay intermediaries this way. Payment processors charge per transaction. Payroll providers charge per employee paid. BPO firms charge per claim, call, document, or case. Agentic AI can tap into those established buying habits.
The important principle is that the transaction should be material, not merely technical. “One API request” is not a transaction. “One successfully completed return authorization” is.
Transaction pricing also creates a natural bridge from traditional outsourcing. If a company previously paid a business-process outsourcer $4 to process a document, an AI agent that reliably processes the same document may be priced at a meaningful discount while still generating excellent vendor margins. The reference point should not be the model’s token cost. It should be the fully loaded cost, latency, quality, and scalability of the incumbent process.
That said, companies should avoid assuming every agent is a digital employee that can simply be priced at a fraction of salary. Human labor costs include judgment, exception handling, accountability, relationship management, and organizational context. In many agent deployments, humans still review high-risk cases and handle exceptions. Pricing should reflect the work truly automated, not the fantasy of a fully replaced role.
Outcome-based pricing: the destination, but not always the starting point
Outcome-based pricing is compelling because it promises the cleanest alignment: the vendor gets paid only when the customer gets value.
Intercom has become one of the clearest examples. Its Fin AI Agent is priced at $0.99 per outcome. For chat and email, outcomes include a resolution, certain workflow handoffs, a disqualified sales prospect, or—at a higher $9.99 price—a qualified prospect routed according to the customer’s criteria. Intercom defines a resolution as a situation in which no further help is requested after the final AI answer.
Intercom’s product positioning is blunt: “We only charge you if Fin delivers a resolution.” It reports that Fin resolves an average of 76% of conversations across Intercom, with some customers achieving more than 90%.
This is a powerful model because the customer can compare the cost of a $0.99 resolution with the cost of a human-handled interaction—and because the definition of the billable event is reasonably observable. If the agent cannot resolve the issue, Intercom does not charge for the resolution.
Salesforce has also adopted a strictly outcome-based model for its Agentforce Help Agent. Its billing documentation states that an interaction qualifies as a resolution only when it includes at least two turns, is not abandoned, receives no explicitly negative final feedback, and concludes without escalation or with explicitly positive feedback. Unresolved sessions are not billed.
These examples point to the future, but companies should be cautious about declaring all agent pricing outcome-based. The approach works only when five conditions are met:
- The outcome has a precise definition. “Improved customer experience” is not billable. “A support request resolved without escalation within 10 minutes” can be.
- The outcome is measurable in trusted data. Both parties need access to the evidence.
- The vendor materially controls the outcome. Charging per sales closed is risky if the customer’s sales team fails to follow up.
- The vendor can prevent gaming. An agent should not be rewarded for prematurely closing cases or optimizing a narrow metric at the expense of quality.
- The cost-to-serve is bounded. A vendor cannot accept unlimited compute and liability for a fixed outcome price without guardrails.
The lesson is not that outcome pricing is impossible. It is that outcome pricing is a product and operations discipline, not just a commercial term. It requires instrumentation, quality assurance, shared definitions, dispute rules, and often a human-in-the-loop escalation path.
The most resilient agentic AI pricing architecture has three layers.
Charge a recurring subscription for the durable capabilities that make an agent enterprise-ready: integrations, identity and access management, policy configuration, orchestration, data connectors, analytics, testing, audit logs, model routing, and support.
This fee recognizes that enterprise value is not generated solely by model inference. The deployment layer is what makes the agent safe, governable, and repeatable.
Charge a transparent amount for each completed task or transaction. This is the scalable revenue engine and the core unit customers can budget around.
The work fee should be tied to a business-recognizable unit and may have tiers based on complexity. For example:
- $0.75 for a simple support resolution
- $3 for a verified account change
- $12 for an invoice reconciled without exception
- $40 for a software change that passes automated tests and is submitted for approval
The point is not the specific numbers. It is that the customer should understand what they are buying and be able to compare the price with the previous process.
Where value is direct and attributable, add a performance component. This might include:
This layered approach allows the vendor to recover fixed platform costs, monetize real volume, and share in upside without betting the entire business on a metric outside its control.
Pricing agents well begins with a value model, not a competitor price sheet.
The baseline calculation should include four categories:
1. Labor released or avoided. Measure fully loaded time per task, not only wage rates. Include management overhead, training, quality assurance, overtime, and the cost of maintaining capacity for peak periods.
2. Throughput and speed. An agent may create value by handling work outside business hours, reducing backlog, shortening time to resolution, or enabling employees to focus on higher-value exceptions.
3. Quality and risk. Better consistency, fewer errors, improved compliance documentation, reduced rework, and stronger auditability can be worth more than labor savings.
4. Revenue and retention. In sales and service, the value may come from faster response times, more conversions, reduced churn, higher recovery rates, or improved customer lifetime value.
A useful formula is:
Net agent value = labor savings + incremental gross profit + risk and error reduction + capacity value − agent fees − implementation costs − residual human-review costs.
That last term is essential. A vendor should not claim that an agent “replaces” work if a human still spends nearly as much time checking it. Likewise, buyers should not reject an agent simply because it does not eliminate headcount. Capacity can be redeployed to revenue-generating work, service improvements, or higher-quality exception handling.
The commercial rule should be straightforward: price the agent so the customer retains a substantial majority of the verified value, especially early in the relationship. A vendor that captures 80% of theoretical savings before proving reliability will face resistance. A vendor that creates a clear path to a two-to-five-times return on spend will have a much easier renewal conversation.
Agents should not be priced identically at every maturity level.
At Level 1, the system assists a human. Charge per seat, perhaps with a modest usage allowance.
At Level 2, the system completes bounded tasks but requires approval. Charge per task, workflow, or transaction, with lower prices reflecting the ongoing human review.
At Level 3, the agent executes routine work autonomously and escalates exceptions. Charge per verified completed job, and introduce service-level commitments.
At Level 4, the agent directly influences economic results—recovering revenue, converting leads, resolving cases, optimizing operations. Add outcome-based fees or gain-sharing, because the agent’s value is no longer merely activity.
As autonomy rises, the vendor’s pricing power should rise—but so should its accountability. More autonomy requires stronger guarantees around accuracy, reversibility, observability, approvals, and liability. The vendor cannot demand the economics of a managed service while offering the accountability of a lightweight software tool.
The strategic conclusion: sell work, not “AI”
The market will continue to experiment with AI credits, agent seats, conversations, tokens, and digital-worker subscriptions. Many of those models will coexist. But the durable principle is simpler than the packaging suggests.
Customers do not ultimately want an agent. They want customer issues resolved, code shipped, claims processed, revenue collected, records corrected, employees onboarded, and exceptions handled.
That means companies should resist two temptations. The first is to price autonomous agents as generic SaaS because per-seat subscriptions are familiar. The second is to claim pure outcome pricing before the product can reliably measure and control the outcome.
The winning model is a disciplined middle path: a platform fee for enterprise readiness; transparent pricing for completed, verified work; and outcome-based upside where attribution is real.
This structure aligns incentives. It protects the vendor from unbounded compute and customer-side failure. It gives the buyer predictable economics and a credible ROI case. And it allows prices to rise as agents become more capable—not because the technology is more impressive, but because the work is more valuable, more autonomous, and more accountable.
That is the real economic promise of agentic AI: not another software license, and not an artificial employee with a fictional salary. It is a new way to buy and sell trusted business output.