8. September 2026 · AI
SAP Now Prices the Action, Not the User
Why per-action pricing for Joule Agents changes the business case, why SAP’s headline value figures cannot carry it, and how to build one that can
Executive Summary
Strategic context
Situation. SAP is moving from AI that answers questions to AI that executes work. Its learning material describes Joule Assistants that coordinate specialised Joule Agents across five autonomous domains, and the commercial model is changing with it. According to SAP’s own course on the subject, Joule Assistants are “included with your SAP subscription,” while “the Joule Agents that each assistant coordinates are what carry the commercial cost.”
Complication. SAP plans to introduce a new Premium AI model in a phased rollout during Q3 2026, in which agents “are commercialized through AI Unit consumption” and “every agent action consumes 0.02 AI Units.” Per-user packages are to be phased out. At the same time, the value figures SAP publishes for agents are, by SAP’s own labelling, indicative, modelled on an assumed company, or taken from individual customer stories. A cost that scales with work performed is being justified with benefits that do not describe your work.
Question. How should an SAP customer build a business case for agents when cost is per action, and what can SAP’s published material contribute to it?
Key findings
1. The unit of cost has moved from the licence to the work. SAP states that “the value of AI is no longer determined by how many users have access, but by how much work AI actually performs,” and prices agents accordingly. So what? Agent cost becomes a variable operating cost that grows with process volume and exception rates. Forecasting it requires knowing how many actions an agent takes per process instance, which SAP says “depends on the agent and the work it performs” and which the material read does not quantify.
2. SAP’s headline value figures are hypotheses, and SAP labels them that way. The “up to 75%” reduction in multi-step workflow time carries the footnote “numbers assumed for consumer products company with €1 billion in revenue and 2,000 employees.” Assistant-level figures are marked “indicative” and sourced to “SAP Value Advisory industry benchmarks, 2026 BCG and KPMG analysis.” The three supply chain figures on SAP’s SCM page are each footnoted to a single customer story. So what? These numbers are useful for choosing where to look. They are not evidence for your business case, and a steering committee that approves on them is approving someone else’s process.
3. The catalogue is still moving, including the unit being priced. A figure of 55 assistants and 245 agents circulates on LinkedIn. SAP’s learning lesson says “more than 50 Joule Assistants and 200 specialized Joule Agents are being made available.” SAP’s public assistant catalogue listed 52 assistants on 14 September 2026. The same Supplier Management Assistant coordinates “5 Agents” in one SAP lesson and “nine agents” on SAP’s Spend Management page. So what? Plan against a dated catalogue and contract terms, not against a headline. The number of agents behind an assistant is also the number of cost carriers behind it.
4. Custom agents run free until the end of 2026, which removes the cost signal from pilots. SAP describes design time at no cost and a “time-bound promotion” under which “custom agents run at no additional cost through the end of 2026,” after which runtime is “billed on actual resource consumption across compute, storage, transactions, and AI tokens.” So what? A pilot started now can scale without anyone seeing its run cost. Measure consumption during the promotion as if it were billed, or the first real cost figure arrives after the decision to scale.
5. Per-action pricing makes process knowledge a cost prerequisite. SAP’s Autonomous Enterprise page advises to “start with outcomes, not processes.” Under per-action pricing that advice is incomplete: the outcome says what you want, but only the process tells you how many instances, how many exceptions and therefore how many billable actions it takes. So what? Measure the baseline of a process, its volume, variants and manual effort, before an agent touches it. Without that, the ROI dashboard SAP offers has no “before” to compare against.
Critical recommendations
| Priority | Recommendation | Effort | Timeline |
|---|---|---|---|
| High | Obtain in writing: the definition of an “agent action,” the expected actions per process instance for your target agents, and your contractual price per AI Unit | Low | Before any commitment |
| High | Measure baseline volume, variants and manual effort for two or three candidate processes before activation | Medium | One quarter |
| High | Track consumption of custom agents during the 2026 promotion as if it were billed | Low | Immediate |
| Medium | Rebuild the business case per process from your own figures; use SAP benchmarks only to shortlist | Medium | Per process |
| Medium | Review how current per-user AI packages transition, and what changes in your contract | Low | Before renewal |
| Low | Record the catalogue version and date that each plan is based on | Low | Ongoing |
Bottom line
Per-action pricing is a fair model: you pay for work done, not for seats nobody uses. It is also unforgiving of guesswork, because cost follows your real process volume while the published benefits follow someone else’s. The business case has to be built from your processes upward, and the measurement has to exist before the agent does.
Where the conversation currently sits
A widely shared LinkedIn post recently summarised SAP’s direction well: the shift from “AI that answers questions to AI that actually gets work done,” with assistants coordinating agents, data and business context towards action under human oversight. Its advice to SAP leaders was to start not with “where can we add AI?” but with which processes consume the most manual effort, where humans must remain in the loop, and “how do we measure ROI from an AI agent?”
That last question is the one this article takes up, because SAP’s own material has quietly changed what the answer depends on. The post pointed to six SAP resources. Reading them, together with the SAP Learning lessons they lead to, shows that the commercial model for agents is being rebuilt around consumption, and that the published value figures are not built to carry a customer’s business case.
Finding 1: From seats to actions
SAP’s learning course “Introducing Joule” contains a lesson on the commercial model that is more explicit than any product page read for this article. Its core distinction is between “what is included as Base AI, and what scales as Premium AI through consumption-based pricing.”
The change it describes has four parts:
| Element | Today, per SAP | Planned, per SAP |
|---|---|---|
| Premium AI packaging | “Per-User-Per-Month (PUPM) packages for different business areas such as Finance, HCM, Supply Chain, Spend, and Customer Experience” | PUPM packages “will be phased out” |
| Generative AI features | Largely inside PUPM packages | “Most generative AI features … move to Base AI” at no additional cost |
| Agents | Carry “the commercial cost, within the Premium AI capabilities packages” | “Agents become the core of Premium AI and are commercialized through AI Unit consumption” |
| Unit price | Not applicable | “In 2026, every agent action consumes 0.02 AI Units” |
SAP’s rationale is stated directly: “the value of AI is no longer determined by how many users have access, but by how much work AI actually performs.”
Two things are not stated in the lesson. It does not define what counts as an “agent action.” And it does not give a price per AI Unit, which is a matter of contract. The formula that matters for a business case therefore has two unknowns that only SAP and your contract can fill:
Agent cost per period = process instances × agent actions per instance × 0.02 AI Units × price per AI Unit
The structure of the formula is itself the finding. Cost scales linearly with volume and with the number of actions per instance. A process with many exceptions, many variants or many retries costs more to automate per instance than a clean one, even if both are “the same process” on a slide.
The decision this creates. Before committing, obtain three figures in writing: the definition of an agent action, the expected actions per instance for each agent you intend to use, and your price per AI Unit. Without all three, any agent cost forecast is an assumption.
Finding 2: The benefits are someone else’s process
SAP publishes a great many value figures for agents and assistants. Read with their footnotes, they fall into three categories.
| Figure as published | Where | What the footnote or wording says | What it is |
|---|---|---|---|
| “Reduce time spent on multi-step workflows by up to 75%” | Joule Agents product page | “Numbers assumed for consumer products company with €1 billion in revenue and 2,000 employees … Source: SAP Value Management” | A modelled value for an assumed company |
| Closing cycle time reduced by 10 to 20 percent; cash cycle time reduced by 50 to 70 percent; and many more | SAP Learning, assistant and agent lessons | Worded as “expected,” “targeting,” “projected”; “All figures are indicative … Sources: SAP Value Advisory industry benchmarks, 2026 BCG and KPMG analysis” | Indicative benchmarks |
| 5 to 15% maverick spend reduction; 40 to 50% FTE productivity in contract management | Spend Management page | “Validated from BCG/SAP Study 2026 on Agentic AI Impact for Spend Management” | A study co-authored with the vendor |
| 4x data visibility; 80% fewer hours on supply chain operations; 30% lower days inventory outstanding | Supply Chain Management page | Each footnoted to one SAP customer story: Microsoft, Embraer, SLB | Three individual customer results, not agent outcomes |
None of this is concealed. SAP labels its figures honestly in the small print, and its value lesson states that “your results will depend on your landscape, your adoption approach, and how deeply Joule is integrated into your processes.”
The problem is how such figures travel. On a slide in a steering committee the footnote is gone, “up to” becomes “about,” and a modelled figure for an assumed consumer products company becomes the expected benefit for a machinery manufacturer. Under per-user pricing that mistake was bounded by the licence cost. Under per-action pricing, cost follows your real volume while the benefit was never yours.
The decision this creates. Use published figures to shortlist processes where agents plausibly help. Then discard them and build the benefit side from your own baseline.
Finding 3: Counting a moving catalogue
The LinkedIn post that prompted this article cited “55 Assistants + 245 Agents across five autonomous domains,” from an image attributed to SAP. That exact figure does not appear on the pages read. What does appear:
- SAP Learning: “More than 50 Joule Assistants and 200 specialized Joule Agents are being made available across these domains.”
- Another lesson in the same course: “more than 200 specialized agents execute well-defined business processes continuously,” in the present tense.
- The public assistant catalogue on sap.com, counted page by page on 14 September 2026: 52 assistants.
- The Supplier Management Assistant: “across 5 Agents” in the SAP Learning lesson, “Rely on nine agents” on the Spend Management page.
- On availability, SAP itself says it “spans general release, early access, and planned releases across 2026.”
None of this is surprising for a portfolio in rapid build-out, and none of it is a criticism. It matters for two practical reasons. First, a plan written against a headline count will not match what can be contracted on a given date. Second, under the new commercial model the agent is the cost carrier. When the number of agents behind an assistant changes, the cost profile of using that assistant may change with it.
The decision this creates. Tie every plan to a catalogue version and a date, and ask for the agent composition of each assistant you intend to use, as it stands at contract signature.
Finding 4: The free runtime problem
For agents a customer builds, SAP describes a separate model. Design time in Joule Studio is “provided at no cost,” within fair-use limits. Runtime is “billed on actual resource consumption across compute, storage, transactions, and AI tokens, purchased through BTPEA, CPEA, or PAYG.” And: “Under a current time-bound promotion, custom agents run at no additional cost through the end of 2026.”
The promotion removes a real barrier to experimentation, which is its purpose. It also removes the only signal that tells a pilot team what its agent will cost in production. A custom agent built in autumn 2026 can be scaled across business units before its first invoice exists.
There is a second consequence. A make-or-buy comparison now spans two different cost regimes: SAP’s agents at AI Units per action, custom agents at resource consumption. They are not directly comparable unless both are measured in the same process under the same volume.
Experience, not sourced. Consumption-priced cloud services are routinely underestimated in their first year, because pilot volumes are small and production volumes arrive with the exceptions nobody tested. No figure is offered here because none was read for this article.
The decision this creates. Instrument custom agents during the promotion as if they were billed: record compute, tokens and transactions per process instance. That is the only way the 2027 budget is based on data.
Finding 5: Outcomes need processes to be priced
SAP’s Autonomous Enterprise page advises: “Start with outcomes, not processes, by partnering with AI to remove busywork and apply human judgment where it creates the most value.” The LinkedIn post advised almost the opposite: start with the processes that consume the most manual effort.
Both are right about something. Outcomes define what success means. But per-action pricing settles which one comes first for a business case, because every term in the cost formula is a property of the process: how many instances, how many variants, how many exceptions requiring additional agent actions.
SAP’s own portfolio contains the instruments to measure this. SAP’s Joule Agents page describes SAP AI Agent Hub as including “a KPI dashboard that offers insights into governance, adoption, and ROI,” and its agent mining capability measures “variants, path completion and conformance, and activity coverage.” Those tools can show what an agent did and at what volume. What they cannot show is the process before the agent arrived, unless someone measured it.
The decision this creates. Before activating an agent in a process, capture the baseline: instance volume, variant count, exception rate and manual effort per instance. Process mining is the natural instrument where it exists. Without a baseline, ROI after activation is a number without a comparison.
Implications for Executives
Agent cost belongs to the process owner, not only to the licence manager. Under per-user pricing, cost was set at contract signature. Under per-action pricing, it is set every day by process volume and quality. Budget responsibility should follow.
Process quality now has a price tag. A process with many exceptions was always expensive in human effort. Automated with agents billed per action, the same exceptions show up as consumption. Cleaning up a process before automating it has become a directly quantifiable saving.
The steering committee needs a different kind of slide. Instead of vendor benefit figures, the case for each agent should show measured baseline, expected actions per instance, contracted unit price and the resulting cost range. That slide is harder to produce, and it is the only one that survives the first invoice.
Recommendations
| Recommendation | Rationale | Owner | Effort |
|---|---|---|---|
| Obtain the definition of an agent action, expected actions per instance and the AI Unit price in writing | Two terms of the cost formula are not in the published material | Procurement with enterprise architecture | Low |
| Measure baseline volume, variants, exceptions and effort per candidate process | Cost and benefit both depend on it; ROI needs a “before” | Process owner with process mining team | Medium |
| Instrument custom agent consumption during the 2026 promotion | Otherwise the first cost figure arrives after scaling | IT and BTP platform owner | Low |
| Rebuild each business case from own figures; use SAP benchmarks to shortlist only | Published figures are modelled, indicative or single-customer | Business case owner | Medium |
| Map the transition of current per-user AI packages in your contract | Packages are to be phased out; features move to Base AI | Procurement | Low |
| Record catalogue version and date per plan | Agent composition of assistants is still changing | Enterprise architecture | Low |
Executive Dashboard
| Question | What SAP’s material provides | What you must supply | Confidence |
|---|---|---|---|
| What does an agent cost? | 0.02 AI Units per action in 2026; consumption depends on agent and work | Actions per instance; price per AI Unit | High for the mechanism, unknown for your figure |
| What counts as an action? | Not defined in the material read | Written definition from SAP | High that it is undefined |
| What benefit can we expect? | Modelled, indicative and single-customer figures, labelled as such | Measured baseline per process | High |
| How many agents are available? | “More than 50” assistants and “200” agents “being made available”; 52 assistants listed on 14 September 2026 | Dated catalogue at contract signature | High |
| What do custom agents cost? | Free build; free run until end of 2026; then resource consumption | Consumption measured during the promotion | High |
| Can ROI be measured afterwards? | KPI dashboard and agent mining in AI Agent Hub | Baseline captured before activation | Medium |
Appendix: Methodology
Research type. Document analysis of vendor commercial and value material, with a derived cost structure.
Sources and how they were read. The six links in the LinkedIn post that prompted this article resolve to five distinct SAP pages; two of the links lead to the same page. All five were read in full on 14 September 2026 (the Joule Agents page in its global version, while the post links the India variant), together with SAP AI Agent Hub’s product page and four lessons of the SAP Learning course “Introducing Joule,” including the lesson on the commercial model. Several sap.com pages refuse automated retrieval and were read in a browser. The assistant catalogue was counted by paging through all five catalogue pages. Footnotes were read alongside every figure. Quotations are verbatim.
Frameworks applied. SCQA for framing. Value chain thinking in reduced form: the cost formula decomposes agent cost into process-level drivers. No full value chain analysis was performed, because no customer cost data was available.
Frameworks considered and not applied. Porter’s Five Forces: the question concerns a customer’s business case, not industry structure. BCG Matrix: no portfolio decision. PESTEL: not a macro-environmental question. Technology Adoption Curve: no reliable adoption data for SAP agents was available, and SAP’s own customer counts are vendor statements. Scenario analysis: without a price per AI Unit or actions per instance, scenarios would have been built on invented inputs.
Evidence status, stated plainly.
- Sourced from SAP material read directly. The commercial model, the 0.02 AI Unit figure, the promotion, the value figures and their footnotes, the catalogue statements.
- Own count. 52 assistants in the public catalogue on 14 September 2026.
- Derived. The cost formula, which is arithmetic on SAP’s stated mechanism.
- Not verified. The “55 Assistants + 245 Agents” figure circulating on LinkedIn; its original source was not found on the pages read.
- Experience claims, not sourced. First-year underestimation of consumption-priced services. Labelled where it appears.
Confidence. High for Findings 1 to 4, which rest on verbatim SAP statements and a direct count. Medium for Finding 5, which combines SAP statements with an argument about sequence.
Limitations.
- The commercial model is described in a learning lesson aimed partly at partner sales roles. Contract terms may differ and take precedence.
- SAP describes the new model as planned for a phased rollout during Q3 2026. Details may change.
- No price per AI Unit was available, so no cost example with real figures could be given. An invented example was deliberately avoided.
- The LinkedIn post that prompted this article is not quoted as evidence; only its advice is discussed.
What could not be answered. How many actions typical SAP agents take per process instance, and what an “agent action” is precisely.
Update trigger. Revisit when SAP publishes a definition of an agent action or consumption estimates per agent, when the Q3 2026 rollout details are final, or at the end of the custom agent promotion.
Frequently Asked Questions
Is per-action pricing bad for customers?
Not inherently. Paying for work performed rather than for seats is fair and removes the cost of unused licences. It does require knowing your process volumes, which many organisations have not measured.
What is an AI Unit worth in euros?
The material read does not say. It is a contractual figure. Ask for it, together with the expected consumption of the agents you intend to use.
Can we use SAP’s value figures in our business case?
Use them to decide where to look. SAP labels them as modelled, indicative or based on individual customers. The benefit in your case has to come from your own baseline.
Are Joule Assistants free?
SAP describes assistants as part of the Joule experience layer “included with your SAP subscription,” and states that the agents they coordinate carry the commercial cost. The commercial model lesson qualifies the inclusion: Joule Base requires a qualifying SAP cloud subscription and a one-time acceptance of SAP’s AI terms. Activating an assistant is therefore not the same as incurring no cost.
Should we build custom agents while runtime is free?
Experimenting during the promotion is sensible. Scaling without measuring consumption is not, because the first real cost figure would then arrive after the decision.
How many SAP agents are there?
SAP Learning says more than 50 assistants and 200 agents are being made available. The public catalogue listed 52 assistants on 14 September 2026. The figure of 55 assistants and 245 agents could not be verified on the pages read.
Why was no scenario analysis included?
Because two of the four inputs to agent cost, actions per instance and price per AI Unit, are not published. Scenarios on invented inputs would look precise and mean nothing.
Sources
- SAP, “Joule Agents and Joule Assistants,” https://www.sap.com/products/artificial-intelligence/ai-agents.html, retrieved 14 September 2026.
- SAP, “Joule Assistants” catalogue, https://www.sap.com/products/artificial-intelligence/ai-assistant.html, retrieved and counted 14 September 2026.
- SAP Learning, “Introducing Joule,” lesson “Discovering Joule,” https://learning.sap.com/courses/introducing-joule/discovering-joule_e055e58d-18c3-48fd-9137-80a96e4431e7, retrieved 14 September 2026.
- SAP Learning, “Introducing Joule,” lesson “Understanding the Value Provided by Joule,” https://learning.sap.com/courses/introducing-joule/understanding-the-value-provided-by-joule_ea9c9612-5bc7-427f-a518-f71d47061707, retrieved 14 September 2026.
- SAP Learning, “Introducing Joule,” lesson “Getting to Know AI Assistants in Joule,” https://learning.sap.com/courses/introducing-joule/getting-to-know-ai-assistants-in-joule_f6252c97-c6fd-46e9-b0c5-2467efd1f1c9, retrieved 14 September 2026.
- SAP Learning, “Introducing Joule,” lesson “Getting to Know Joule Agents,” https://learning.sap.com/courses/introducing-joule/getting-to-know-joule-agents_f9d18ca0-1021-4c1a-a044-4c00ec8c2898, retrieved 14 September 2026.
- SAP Learning, “Introducing Joule,” lesson “Understanding the Commercial Model,” https://learning.sap.com/courses/introducing-joule/understanding-the-commercial-model_a778010e-ba84-4378-ab3f-6c4b31d360eb, retrieved 14 September 2026.
- SAP, “The Autonomous Enterprise,” https://www.sap.com/products/business-suite.html, retrieved 14 September 2026.
- SAP, “Supply Chain Management (SCM) Software Solutions,” https://www.sap.com/india/products/scm.html, retrieved 14 September 2026.
- SAP, “Spend Management Software Solutions,” https://www.sap.com/india/products/spend-management.html, retrieved 14 September 2026.
- SAP, “SAP AI Agent Hub,” https://www.sap.com/products/artificial-intelligence/ai-agent-hub.html, retrieved 14 September 2026.
