.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.
How to Price for the Age of Agentic Procurement
The buyer is changing
For most of B2B history, pricing has been designed to survive a human buying process.
Suppliers published enough information to generate interest, salespeople explained the offer, procurement collected proposals, and executives negotiated the final agreement. Price was rarely a clean number. It was a combination of licenses, modules, service fees, minimum commitments, volume discounts, implementation costs, contractual protections, and negotiated exceptions.
Agentic AI changes that architecture because the buyer is no longer necessarily human.
Procurement agents are beginning to discover suppliers, draft requests for proposals, normalize bids, analyze contracts, conduct risk checks, recommend awards, create purchase orders, and monitor invoices against negotiated terms. McKinsey estimates that agentic AI could make procurement functions 25% to 40% more efficient. Its research also found that 40% of procurement organizations had already implemented or piloted generative AI. Gartner has gone further, predicting that by 2028, 90% of B2B buying will be intermediated by AI agents, representing more than $15 trillion of spending. (mckinsey.com)
This does not mean algorithms will autonomously sign every strategic contract. High-risk purchases will continue to require legal, financial, technical, and executive judgment. But the initial shortlist—and an increasing share of the commercial analysis behind it—may be created by software.
That distinction matters. A human buyer can be persuaded by a relationship, a brand narrative, or an artfully constructed proposal. A purchasing agent will be more inclined to ask whether the price can be parsed, modeled, verified, and compared.
In agentic procurement, the rate card becomes an interface. The contract becomes executable policy. And pricing ambiguity becomes a measurable purchasing risk.
The central pricing question is therefore no longer simply, “What will customers pay?” It is also, “What price architecture will a buying agent understand, trust, and recommend?”
Traditional B2B pricing was designed around human labor
Per-seat pricing became the dominant SaaS model because it connected software access to an intuitive customer anchor: the employee using the product.
A company with 500 salespeople could understand why it needed 500 CRM licenses. The number of employees roughly represented both the scale of deployment and the value opportunity. The model was predictable for finance, straightforward for procurement, and exceptionally attractive for software vendors because the marginal cost of serving another licensed user was usually low.
Agentic workflows break that relationship.
One employee may supervise ten agents. One agent may perform work previously distributed across dozens of users. Some agents operate continuously without any direct human interaction. If the software is executing the task, rather than helping an employee execute it, the number of human seats becomes a weak proxy for value.
This is the core principle of the Agentic Monetization Spectrum described in Monetizing Agentic AI. The more autonomous an agent becomes, the broader the operational domain it covers, and the more its output value outpaces its compute cost, the further pricing should move from access toward activity, output, or outcome. Per-seat pricing can remain appropriate when AI assists a human. It becomes harder to defend when the human is no longer the primary unit of production. (getmonetizely.com)
Bret Taylor, Sierra’s co-founder, has described the emerging atomic unit of AI productivity as “a process, not a person.” That is also a useful way to frame pricing. If customers are purchasing an automated process, suppliers need to identify what moves when that process creates value: cases resolved, transactions completed, documents reviewed, sourcing events executed, dollars managed, or savings realized. (sierra.ai)
Seats may not disappear. They can remain useful for governance, administrative access, human collaboration, or premium functionality. But in many products, the seat will become what Monetizing Agentic AI calls a cover charge: a base fee that provides access while the principal revenue engine moves toward the work performed by the agent. (getmonetizely.com)
What changes when an AI agent evaluates the purchase
Human procurement teams often struggle to compare commercially dissimilar proposals. One supplier quotes by user, another by transaction, a third by annual spend, and a fourth through a platform fee plus services. Procurement analysts must translate those offers into a common total-cost model, frequently using incomplete assumptions.
Agents can perform that normalization faster and more consistently.
A purchasing agent could model the expected annual cost of each proposal against the company’s actual operating data. It could test low-, expected-, and high-usage scenarios; calculate effective unit rates after thresholds and discounts; estimate overages; identify credits that expire; quantify implementation costs; and adjust for SLA penalties or operational risk.
Its evaluation would therefore extend beyond sticker price:
Expected economic cost = fixed fees + expected variable charges + implementation and switching costs + risk-adjusted failure costs − contractual credits and measurable benefits.
The agent may also detect inconsistencies that human buyers miss. Gartner found that 69% of B2B buyers encounter discrepancies between information on supplier websites and information provided by salespeople. A procurement agent comparing a public pricing page, marketplace listing, proposal, order form, and historical invoice could flag those contradictions immediately. (gartner.com)
This changes the commercial advantage of complexity. Historically, opaque pricing could preserve negotiation leverage, make direct comparison difficult, and allow sellers to price discriminate across accounts. Under agentic evaluation, unnecessary complexity may instead produce an uncertainty penalty. If one offer can be modeled with confidence and another requires several rounds of clarification, the agent may rank the transparent offer more highly even when its nominal price is not the lowest.
The effect will be strongest in repeatable categories with clear specifications and many credible suppliers. Strategic partnerships, novel technology purchases, and deals requiring organizational change will remain less reducible to automated comparison.
That is consistent with current buyer behavior. Gartner reports that 61% of B2B buyers prefer an overall rep-free experience, but buyers still prefer seller involvement when assessing whether an unfamiliar solution fits their company. McKinsey likewise finds that B2B customers continue to divide their interactions among in-person, remote, and digital channels. Agentic procurement is likely to automate information collection and initial evaluation before it replaces the contextual judgment used in complex selection. (gartner.com)
The new pricing models: seat, usage, task, transaction, outcome, and hybrid
There will be no single agentic pricing model. The right metric depends on what the agent does, how reliably it does it, how customers perceive the purchase, and how variable the supplier’s delivery costs are.
The market is already experimenting across this spectrum.
Salesforce offers multiple Agentforce structures. Its Flex Credits are sold at $500 per 100,000 credits, with a standard agent action consuming 20 credits—effectively $0.10 per action. Salesforce also lists conversation pricing at $2 per conversation, per-user options, prepaid purchasing, baseline commitments, and pay-as-you-go arrangements. The menu is significant: even one major vendor does not assume that a single metric can serve every internal and customer-facing agent deployment. (help.salesforce.com)
Intercom’s Fin charges $0.99 when its agent produces a defined resolution. Sierra says customers pay for agreed outcomes rather than tokens or seats, while also acknowledging that blended or consumption-based pricing may be more appropriate for interactions where an attributable result cannot be established. (intercom.com)
Cognition’s Devin illustrates the other side of the spectrum. Its enterprise product measures work through Agent Compute Units, while self-service plans combine subscriptions, quotas, seats, and on-demand credits. Usage depends partly on the number and complexity of the actions Devin takes, including planning, context gathering, execution, and virtual-machine activity. That structure protects the vendor when the amount of compute needed to complete a coding task is unpredictable. (docs.devin.ai)
Vantage provides a spend- and outcome-linked example. Its Autopilot service charges 5% of the cloud savings it realizes. If it creates no savings, there is no performance fee. Its newer FinOps Agent combines savings-based charges for financial commitments with planned token-based charges for conversational usage. (vantage.sh)
These models are not contradictory. They reflect different levels of autonomy, reliability, cost exposure, and attribution.
The governing rule should be: price as close to customer value as the product’s reliability and measurement system can safely support. If a vendor cannot objectively prove the outcome, it should retreat one level—to a completed task, an agent action, a transaction, or measured consumption. As Monetizing Agentic AI argues, “Aspiration belongs in the roadmap, not the rate card.” (getmonetizely.com)
Machine-readable pricing becomes a competitive capability
A price attractive to an AI purchasing agent must be more than publicly visible. It must be structured.
At a minimum, the supplier should make the following elements unambiguous:
A human can interpret “contact sales for enterprise pricing” as an invitation. An agent may interpret it as missing data.
The infrastructure for machine-executable commerce is already emerging. Stripe and OpenAI’s Agentic Commerce Protocol establishes a common language through which agents and businesses can exchange product, pricing, checkout, and order information. Google’s Agent Payments Protocol uses cryptographically signed mandates to record the buyer’s instructions and the exact contents and price of a cart. Visa and Mastercard have introduced infrastructure for authorized agent-initiated payments with user-defined controls. (stripe.com)
These initiatives began largely in consumer commerce, but the principle applies directly to B2B procurement. Buying agents need structured access to catalogs, price logic, inventory or capacity, commercial terms, authorization requirements, and fulfillment status. Stripe explicitly describes the need for machine-readable inventory, pricing, checkout logic, trust signals, return policies, and guarantees rather than agents scraping interfaces built for people. (stripe.com)
For B2B suppliers, this suggests a future commercial artifact: an agent-readable offer manifest. It would provide the canonical version of pricing and terms through an API, marketplace listing, standardized schema, or secure model-context server. The PDF proposal would not disappear, but it would no longer be the only authoritative representation of the deal.
Agentic procurement will intensify competition—but not always on price
Agents reduce the cost of searching and comparing suppliers. They can evaluate more vendors, rerun sourcing events more frequently, and monitor whether incumbent pricing remains competitive. That should put pressure on unexplained premiums and reduce the protection created by difficult-to-compare pricing.
Procurement technology is already moving in this direction. Oracle has documented autonomous sourcing and supplier-negotiation agents. SAP is introducing Joule agents that can execute multistep procurement workflows, including sourcing-event activities. Zip says it has launched more than 50 specialized procurement agents for tasks such as intake validation, risk assessment, document review, and invoice-to-contract compliance. AWS Marketplace has introduced an agent-based discovery experience for software procurement and supports contract, usage, and hybrid pricing for agent products. (docs.oracle.com)
However, easier comparison does not guarantee a race to the bottom.
A procurement agent can compare total economic value more rigorously than a simple price-comparison website. It can assign value to higher reliability, faster deployment, stronger security, lower integration costs, superior support, better contractual protections, and less volatile billing. A supplier charging more may still win if its expected cost distribution is narrower or its failure risk is lower.
Seller agents may also respond dynamically. Google’s AP2 example describes a merchant agent assembling a time-sensitive bundle with a 15% discount in response to a buyer’s requirements. In B2B markets, seller agents could construct packages, offer commitment-based discounts, or alter payment terms within approved commercial guardrails. (cloud.google.com)
The result may be less like a public commodity exchange and more like continuous algorithmic negotiation. Buyers’ agents will optimize total value within budget and policy constraints. Sellers’ agents will optimize conversion, margin, capacity, and account potential.
Opaque pricing will lose some of its protective power, but sophisticated segmentation will not disappear. It will become faster, more data-driven, and more explicitly governed.
How suppliers remain differentiated when agents create the shortlist
When an agent performs the initial evaluation, differentiation must be expressed as evidence rather than adjectives.
Claims such as “enterprise-grade,” “easy to use,” or “best-in-class” are difficult to score. A buying agent will benefit more from measurable signals:
This does not eliminate brand. Brand becomes a prior for trust: a signal that the supplier is likely to remain solvent, protect data, maintain integrations, and honor the agreement. Relationships also remain important in purchases where requirements are ambiguous or value depends on organizational transformation.
But suppliers will need to make their differentiation legible to both humans and machines. That means publishing structured technical documentation, clear commercial definitions, verified performance data, and standardized security information. It also means ensuring that pricing communicated by the website, sales team, marketplace, and contract is consistent.
The strongest differentiation may ultimately come from the agent harness around the underlying model: proprietary workflow orchestration, permissions, integrations, memory, domain context, controls, and feedback data. Foundation models can become cheaper or interchangeable. Embedded operational context and proven execution are harder to replace.
Pricing must protect both customer budgets and supplier margins
Agentic products create an economic problem that conventional SaaS rarely faced: meaningful variable cost.
Every model call, tool invocation, search, generated image, external API request, and minute of execution can increase cost. Microsoft’s Copilot Studio, for example, applies different credit rates to classic answers, generative answers, agent actions, grounding, content processing, voice, and premium reasoning. The cost of an agent interaction depends on what the agent decides to do, not merely on the fact that a user initiated it. (learn.microsoft.com)
At the same time, underlying inference prices are falling quickly. Stanford’s 2025 AI Index found that the cost of inference at approximately GPT-3.5 performance fell more than 280-fold between November 2022 and October 2024. Vendors that pass every compute change directly into their customer-facing metric risk creating a rate card that becomes obsolete whenever model economics change. (hai.stanford.edu)
The answer is usually not pure cost-plus pricing. Customers do not care how many tokens an agent consumed to approve an invoice or resolve a ticket. But suppliers cannot ignore those costs either.
A practical architecture is often:
- A base platform fee or minimum commitment that funds availability, integrations, security, support, and reserved capacity.
- An included usage or outcome allowance that makes the expected bill predictable.
- A variable charge for additional tasks, actions, transactions, or outcomes.
- Volume tiers that share scale efficiencies.
- Caps and alerts that prevent bill shock.
- Premium rates for high-cost models, priority processing, advanced reasoning, or demanding SLAs.
- True-ups or rollover provisions that govern commitment risk explicitly.
This is hybrid pricing not as a compromise, but as risk allocation. The base fee protects the supplier from fixed costs and underutilization. The variable component protects it from heavy usage. Included allowances and caps make the arrangement procurable.
Outcome pricing is most attractive when the result is autonomous, valuable, binary, and attributable.
A support case was resolved. A compliant invoice was processed. A cloud commitment produced a measured discount. A payment was recovered. A qualified meeting occurred.
Procurement outcomes are often less clean. Consider “savings realized.” Was the saving caused by the agent, a fall in market price, reduced demand, a specification change, or a procurement professional’s negotiation? What baseline should be used? Does avoidance of a future price increase count? Who receives credit if several systems contributed?
The further pricing moves from an operational event toward a financial result, the greater the counterfactual problem.
Sierra, despite advocating outcome pricing, acknowledges that the model is operationally, contractually, and financially more complex than seat or consumption pricing. Its conclusion is important: outcome pricing works where the software is highly autonomous and the result is highly attributable. (sierra.ai)
Vendors should therefore distinguish between:
- Operational outcomes: a sourcing event completed, a contract reviewed, an invoice matched, or a case resolved.
- Business outcomes: spend reduced, revenue increased, churn prevented, or working capital improved.
Operational outcomes are generally easier to meter and audit. Business outcomes may support higher prices, but they require shared baselines, attribution rules, dispute procedures, and trusted instrumentation.
The right model is often to charge primarily for the nearest verifiable operational outcome, with a performance bonus or gain-sharing component for attributable financial results.
Procurement leaders should define what agents may discover, recommend, negotiate, and purchase without human approval. Authority limits should cover spend, data sensitivity, contract risk, supplier criticality, and deviation from standard terms.
CFOs should demand scenario-based forecasting for agent usage. An average-cost model is insufficient when autonomous systems can generate fat-tailed consumption. Finance needs visibility into committed capacity, credit burn, overages, supplier margins, and outcome disputes.
Pricing teams should stop treating the metric as a one-time packaging decision. Agentic pricing requires continuous analysis of customer value, reliability, cost distributions, and purchasing behavior.
SaaS vendors should build a machine-readable commercial catalog, instrument usage at the product level, separate expensive actions from inexpensive ones, and ensure that pricing logic can change without rebuilding the billing system.
This is the discipline Monetizing Agentic AI calls Monetization Engineering: the infrastructure connecting CPQ, entitlements, metering, rating, billing, revenue recognition, and ERP. In traditional SaaS, counting seats was largely administrative. In agentic software, determining what happened, what it cost, whether it was billable, which tier applied, and what the customer is entitled to becomes a real-time systems problem. (getmonetizely.com)
Operationalization may ultimately become a larger constraint than pricing strategy. A theoretically elegant outcome model is worthless if the outcome cannot be measured, disputed charges cannot be reversed, or the invoice cannot explain how the total was calculated.
Conclusion: Pricing for a world where software buys software
Agentic procurement does not invalidate the fundamentals of pricing. Suppliers still need to segment customers, understand willingness to pay, quantify value, protect margins, and package offers around differentiated needs.
What changes is the precision required.
When software buys software, every ambiguity becomes data. Every discount requires a rule. Every threshold needs a definition. Every performance claim needs evidence. Every outcome needs a measurement system. Every variable charge needs a forecastable logic.
Per-seat pricing will survive where humans remain the value anchor. Usage pricing will remain appropriate where consumption closely tracks cost or customer activity. Task and transaction pricing will grow where agents perform discrete work. Outcome pricing will command attention where results are autonomous and attributable. Most enterprise vendors will use hybrids that combine commitments and access fees with measured work.
The winning price will not necessarily be the lowest. It will be the price an agent can understand, a procurement team can defend, a CFO can forecast, and a supplier can deliver profitably.
In the age of agentic procurement, commercial transparency is no longer merely a customer-experience decision. It is part of product discoverability. Pricing infrastructure is part of the technology architecture. And the offer itself must be designed for two audiences at once: the human who defines the value and the machine that decides whether the numbers add up.
Selected references
- Monetizely, Monetizing Agentic AI frameworks and published chapter excerpts. (getmonetizely.com)
- McKinsey, research on agentic AI and procurement operating models. (mckinsey.com)
- Deloitte, 2025 Global Chief Procurement Officer Survey. (deloitte.com)
- Gartner, B2B buying preferences and agent-intermediated buying prediction. (gartner.com)
- Salesforce, Agentforce pricing and Flex Credits. (salesforce.com)
- Intercom, Fin AI Agent outcome definitions and pricing. (intercom.com)
- Sierra, outcome-based pricing for AI agents. (sierra.ai)
- Cognition, Devin billing and usage documentation. (docs.devin.ai)
- AWS Marketplace, agent discovery and agent-product pricing models. (aws.amazon.com)
- Google, Agent Payments Protocol. (cloud.google.com)
- Stripe and OpenAI, Agentic Commerce Protocol. (stripe.com)
- Stanford HAI, 2025 AI Index Report. (hai.stanford.edu)