.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 Should We Price a Procurement Agent per Seat per Action or per Outcome
Procurement agents are becoming harder to price with the logic that governed traditional SaaS. A purchasing system once created value because employees logged into it. An agent can now monitor supplier messages, draft follow-ups, analyse purchase-order changes, retrieve information, complete tasks and trigger workflows while the procurement team remains the same size. Microsoft makes that shift visible in its own commercial model: as of 13 August 2026, Dynamics 365 Supply Chain Management is still sold by user, but its Procurement Agent consumes Copilot Credits, and Microsoft says supplier-communication runs carry both a fixed consumption charge and variable consumption tied to work such as emails and attachments.
Other enterprise vendors are moving in the same direction. Salesforce prices a standard Agentforce action at 20 Flex Credits, equivalent to $0.10 at its published $500 per 100,000-credit rate. Workday says its Flex Credits charge for work completed rather than employee count, with one published example consuming one credit for information retrieval and five credits for autonomous task completion. SAP lets premium AI consume AI Units against units such as requests, users and records.
Monetizely's position is clear: per action should be the primary pricing metric for a procurement agent in 2026. Outcome pricing ranks second and should be confined to unusually clear, auditable results; seat pricing ranks last for the agent itself. The commercial architecture should pair an annual platform commitment with a pooled allowance of weighted procurement actions and transparent action-based overage.
Procurement agents decouple software value from buyer headcount
The pricing decision becomes clearer when we treat it as one part of a larger design rather than asking, in isolation, whether a seat, action or outcome sounds most modern. Monetizely's 5-Step Pricing Framework, developed more fully in Monetizing Agentic AI, starts with Customer Segmentation: decide which customers generate different value and have different willingness to pay. Packaging then determines which capabilities and levels of autonomy each segment receives. Pricing Metric decides what unit causes the bill to grow. Rate-Setting determines how much each package or unit costs. Operationalization makes the design workable through entitlements, metering, billing, sales rules and governance. The sequence matters here because a procurement organisation allowing an agent to draft supplier emails is buying something materially different from one allowing the agent to execute procurement work with limited human review.
The pricing-metric step carries unusual weight because the human user is steadily becoming a weaker proxy for value. Microsoft's current model is a useful case. Dynamics 365 Supply Chain Management was listed at $210 per user per month and Premium at $300 per user per month on 13 August 2026, yet agents require Copilot Credits; Premium includes 1,000 credits per user per month. The human licence provides access to the system, while additional agent work has its own consumption economics.
For a dedicated procurement agent, we would rank the three candidate meters as follows.
| Rank | Metric | Monetizely's verdict | What to do in practice | Core failure mode |
|---|---|---|---|---|
| 1 | Per completed action | Best primary meter | Sell an annual action commitment, pool credits across workflows and weight actions by work and value | Poorly designed actions can become as opaque as tokens |
| 2 | Per outcome | Use selectively | Apply only where success is objectively defined and attribution is strong | Savings, risk reduction and negotiated value invite attribution disputes |
| 3 | Per seat | Do not make it the agent's main meter | Keep seats for builders, administrators or human-facing workflow access | Agent output can rise without adding a single employee |
The table makes the core choice explicit: procurement software may retain seats around the platform, but the agent should monetise the work it performs.
Seat pricing fails first on expansion. Suppose ten buyers deploy an agent to 200 suppliers, then extend it to 2,000 suppliers without increasing procurement headcount. The customer's value and the vendor's compute workload can rise sharply while seat ARR remains almost flat. The vendor has priced the organisation chart rather than the product's output.
Outcome pricing fails for the opposite reason. It moves so close to financial value that the vendor begins billing for variables it may not control. A supplier price reduction can reflect the agent, a buyer's negotiation, lower commodity prices, reduced volume, a changed specification or a competing bid. Charging a percentage of "savings" requires agreement over the counterfactual: what would the buyer have paid without the agent?
Per action sits between those extremes. A supplier follow-up was either completed or it was not. A bid comparison ran or it did not. A purchase-order change was analysed or it was not. Those units are close enough to customer value to make intuitive sense, yet close enough to the software's work to meter reliably.
The agentic profile points to usage before outcomes
Agent pricing requires one additional question: how independently is the software actually doing the job?
The Agentic Monetization Spectrum, or AMS, answers that through three dimensions. Zero-Human Ability measures how much of the work still requires a person: small when humans perform most of the work, medium when they delegate and review, and large when the agent largely performs the job itself. Operational Domain asks whether the agent handles one narrow task, a workflow inside one function, or work that crosses functions. Output/Cost Curve compares the value produced with the cost required to produce it, ranging from roughly linear economics through an inflecting relationship to cases where output value far outruns compute cost. In Monetizely's view, these dimensions identify when a human seat remains a sensible anchor, when usage should take over and when an attributable outcome becomes mature enough to bill.
A 2026 procurement agent lands in the middle of that spectrum rather than at the fully autonomous edge.
| AMS dimension | Procurement-agent score | Evidence and interpretation | Pricing implication |
|---|---|---|---|
| Zero-Human Ability | Medium, moving upward | Microsoft's Procurement Agent can monitor supplier communications and perform impact analysis, but purchasers still make consequential decisions on changes. | Seat pricing is already weakening, but full outcome pricing is premature |
| Operational Domain | Medium | Supplier communication, PO updates and downstream impact analysis form a multi-step procurement workflow rather than one isolated prompt. | Charge for meaningful business actions, not prompts |
| Output/Cost Curve | Inflecting | Microsoft's billing itself separates fixed execution from variable resources consumed, while enterprise agent vendors increasingly abstract model usage into work units. | Weighted actions can protect gross margin without exposing token economics |
The AMS reading reinforces the ranking. Procurement agents have moved too far beyond human assistance for pure seat pricing, yet the work still contains too much human judgment and external causality for broad outcome pricing.
The strongest meter is therefore completed procurement work.
The market evidence matters because pricing metrics become easier to sell when buyers encounter familiar logic elsewhere. By August 2026, several large enterprise vendors had begun moving away from headcount or raw model consumption and toward credits that represent actions, tasks or other recognisable units of work. Outcome-priced products remain important comparators because they show where the model can go once the result becomes highly verifiable.
The evidence is unusually consistent.
| Product | Published meter and price structure | Date | What it teaches a procurement-agent vendor |
|---|---|---|---|
| Microsoft Dynamics 365 Procurement Agent | Copilot Credits; supplier communications have fixed consumption per run plus variable consumption based on emails and attachments | 10 Aug 2026 | Direct procurement precedent for metering agent work rather than only users. |
| Salesforce Agentforce | $500 per 100,000 Flex Credits; standard Agentforce action uses 20 credits, or $0.10 | 15 May 2025 | A business action can be the customer-facing unit while credits absorb technical complexity. |
| Workday Flex Credits | Annual credit purchase; example skills use 1 credit for information retrieval and 5 for autonomous task completion | Current 13 Aug 2026 | Different kinds of work can carry different weights without exposing tokens. |
| SAP Business AI | Premium AI can use AI Units; SAP says conversion rates may be based on request, user, record or another metric | Current 13 Aug 2026 | A common credit pool can span several enterprise workflows while the underlying units vary. |
| HubSpot Data Agent | $0.10 per answer through 10 credits; other HubSpot agents use resolutions or leads | Current 13 Aug 2026 | One credit currency can support action and outcome units side by side. |
| Intercom Fin | $0.99 per outcome; one billable outcome per conversation even when Fin takes several actions | Current 13 Aug 2026 | Outcome pricing works when the vendor can tightly define success and cap ambiguity. |
| Zendesk AI Agents | Automated resolutions; resolution tiers introduced 18 May 2026 distinguish assisted, contained and verified results | 18 May 2026 | Even outcome pricing needs verification and value differentiation. |
Five of these examples illuminate the winning action or consumption logic directly: Microsoft's Procurement Agent, Salesforce Agentforce, Workday Flex Credits, SAP AI Units and HubSpot's Data Agent. Intercom and Zendesk show the higher bar required before an outcome can become the billable unit.
Workday provides perhaps the cleanest analogue. Its current Flex Credits page explicitly says the company moved away from tying AI value to employee count. Credits are consumed when eligible capabilities are used in production, and specific skills carry distinct rates. Workday lists one credit per information-retrieval action and five credits for autonomous task completion.
That design translates well to procurement. We would define the billable unit above infrastructure but below corporate financial results. A practical action catalogue could include:
- a supplier follow-up completed;
- a purchase-order change reviewed and analysed;
- a sourcing event or bid set compared;
- a contract or statement-of-work draft completed;
- an invoice exception investigated and prepared for resolution.
Those actions should not all cost one unit. A short supplier follow-up may involve little reasoning, while a multi-document bid comparison or downstream PO impact analysis can require far more retrieval, computation and tool use. A rate card of, for example, 1, 5 and 20 credits for progressively heavier classes of work lets the vendor recover that difference while giving the buyer a unit that can still be understood at budget time.
Tokens should remain underneath the meter. Buyers do not receive business value because an LLM consumed 30,000 tokens instead of 10,000. They receive value because a procurement task was finished.
Three first-party pricing changes are particularly instructive. They show that the difficult part is not merely abandoning seats. Vendors also need to avoid choosing a usage metric that is too crude.
These resets point toward a subtle rule: the meter has to represent the work, not the interface through which the work happened.
A procurement agent may read five supplier emails, inspect two attachments, query an ERP, compare an amended delivery date against inventory and prepare one recommendation. Billing for seven messages is arbitrary. Billing for one "conversation" is equally weak. Billing for a completed PO-change analysis is understandable.
Salesforce's transition is especially relevant. The company said in May 2025 that enterprises had difficulty planning Agentforce usage at scale, then introduced Flex Credits that could move across teams and use cases. One standard action consumes 20 credits, while customers can still buy other constructs for specific deployments.
Procurement vendors should learn from that sequence rather than repeat it. Start with a durable business action before autonomous volume makes a weak metric painful to unwind.
Weighted actions capture value without turning savings into a billing dispute
Outcome pricing appears attractive because procurement is associated with measurable money. A vendor might propose charging 1 per cent of negotiated savings, a fee for each supplier saving achieved, or a percentage of recovered spend.
The difficulty is visible when we compare how each metric behaves under the same customer. The following scenario isolates the sensitivity of the three models rather than proposing market rate levels.
The pattern is the argument. When automated work doubles without extra headcount, a seat model captures no expansion. When external market conditions lift reported savings without additional agent work, an outcome model can double the vendor's upside for value it may not have caused. The action model moves when the agent's actual work moves.
Outcome vendors succeed by spending considerable effort on definition. Intercom charges only one Fin outcome per conversation even when the agent takes multiple actions. HubSpot defines Customer Agent pricing around a resolved conversation and, as of April 2026, charges $0.50 for that result. Zendesk's May 2026 design goes further: it uses a 72-hour window and LLM verification to distinguish contained from verified resolutions.
Support resolution has a relatively observable end state. Procurement savings often do not.
Microsoft's own Procurement Agent illustrates the attribution problem. The agent can detect supplier changes and analyse downstream effects on inventory, production and customer orders, yet the purchaser still decides how to respond. Our inference is that charging the software provider a share of the eventual economic "outcome" would require separating the agent's contribution from the purchaser's judgment and from later operational events.
Outcome pricing can still earn a small role. An agent that independently identifies and recovers an objectively duplicated payment, for example, may support a success fee because the recovered amount and causal chain are auditable. That exception should not dictate the base model for the wider procurement platform.
Monetizely's preferred commercial design therefore has four concrete properties:
- The annual commitment buys pooled action capacity. Buyers receive budget predictability while the vendor earns committed ARR.
- Each procurement action has a published credit weight. A simple lookup can cost less than an autonomous multi-system workflow.
- Overage uses the same action currency. Buyers can forecast incremental cost instead of negotiating a new SKU whenever a new agent goes live.
- Outcome fees are exceptional. They apply only to contractually defined, independently auditable events rather than broad measures such as "savings influenced".
That architecture has a named primary meter: per action. The platform fee supplies predictability; the action meter supplies expansion.
The larger strategic risk is waiting too long. Microsoft already charges its procurement agent through consumption on top of the underlying Dynamics licence. Workday explicitly says AI work should be charged based on completed activity rather than employee count. Salesforce has created an action currency that can follow agents across use cases. By August 2026, action-based enterprise AI pricing is no longer an experimental edge case.
Procurement-agent vendors should resist the temptation to jump directly from SaaS seats to a percentage of savings. Moving one step at a time produces a stronger business: first establish trustworthy task completion, then develop a stable action catalogue, then prove how those actions affect business KPIs. Outcome pricing becomes possible later where causality is strong enough to survive scrutiny from procurement, finance and legal.
Monetizely's position for 2026 is therefore not merely that actions beat seats. The winning design prices autonomous procurement work at the narrowest unit a CFO can recognise and an engineering team can meter reliably. Today, that unit is the completed procurement action.
The practical priorities follow from that position:
Instrument autonomous work before optimising price levels. Record completed actions, human review, failures, retries and workflow complexity so pricing decisions start from observed behaviour rather than token estimates.
Package customers by the autonomy they permit, not by the size of their procurement team. An organisation that allows unattended execution should sit in a higher-value package than one that requires a human to approve every material step, even when both have 50 buyers.
Set margin guardrails by action class and revisit them as model costs fall. Customers should see stable business units while the vendor absorbs changing inference economics underneath them.
Prove value with metrics that are deliberately separate from billing. Track procurement cycle time, manual work avoided, exception rates, compliance quality and verified savings to support willingness to pay without forcing every KPI into the invoice.
Assumptions
The billing-sensitivity exhibit is a model built solely to compare how the three meters respond to changing headcount, activity and reported savings; its $150 seat price, $0.50 action price and 1.5 per cent outcome fee are not vendor benchmarks or recommended rate levels. The AMS ratings describe our 2026 procurement-agent archetype rather than a measured score for any single product. Current vendor pricing observations were checked on 13 August 2026 unless a historical date is shown explicitly.
Footnotes
https://www.amazon.com/Monetizing-Agentic-AI-Handbook-Transformation/dp/B0H7Z13VKJ/
https://www.getmonetizely.com/blogs/how-do-companies-decide-on-their-pricing-model
https://www.getmonetizely.com/monetizing-agentic-ai-book-saas/the-agentic-monetization-spectrum
https://www.microsoft.com/en-us/dynamics-365/products/supply-chain-management/pricing
https://learn.microsoft.com/en-us/dynamics365/supply-chain/procurement/procurement-agent-supplier-com-overview
https://learn.microsoft.com/en-us/dynamics365/supply-chain/procurement/procurement-agent-overview
https://learn.microsoft.com/en-us/dynamics365/supply-chain/procurement/procurement-agent-impact-analysis-overview
https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/
https://www.salesforce.com/in/news/press-releases/2024/09/12/agentforce-announcement/
https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio
https://learn.microsoft.com/en-us/microsoft-copilot-studio/billing-licensing
https://www.workday.com/en-us/artificial-intelligence/ai-flex-credits.html
https://www.sap.com/products/artificial-intelligence/pricing.html
https://www.sap.com/products/artificial-intelligence/ai-units.html
https://www.hubspot.com/products/artificial-intelligence
https://www.hubspot.com/company-news/hubspots-customer-agent-and-prospecting-agent-now-you-pay-when-the-task-is-complete
https://www.intercom.com/pricing
https://www.intercom.com/blog/from-resolutions-to-outcomes-evolving-how-fin-delivers-value/
https://support.zendesk.com/hc/en-us/articles/9570369117338-About-automated-resolution-tiers