TL;DR
SAP AI agents are AI systems that plan and carry out multistep business tasks across SAP and non-SAP systems, and most come from SAP as prebuilt Joule Agents. Joule Assistants coordinate those agents, while Joule is the interface people talk to. Adopt prebuilt Joule Agents when your finance, procurement or HR process is close to standard. Build your own in Joule Studio when the workflow is unique, and accept that you then own its full lifecycle. Connect a third-party agent only through an SAP-endorsed path, because SAP’s API Policy restricts other AI-driven API use. Whichever path you choose, start in read-only or recommend mode and add approved write actions later.
Key Takeaways SAP AI agents interpret a goal, pick tools and act inside SAP processes, which is more than a chatbot or a fixed workflow can do. Prebuilt Joule Agents suit standard processes and come with SAP-managed updates. Joule Studio suits unique workflows, but you take on full lifecycle responsibility for anything you build. Third-party agents must use SAP-endorsed paths such as the Agent Gateway, because API Policy section 2.2.2 restricts direct agent API use. An agent’s effective permission is the overlap of the user’s rights and the agent’s rights, checked at every hop. Pilots should begin read-only, prove permission isolation and audit trails, and only then move to approved writes. Watch on YouTube
Why AI Agents Fail in Production
A short Kanerika explainer on why agents that look finished in a demo break once they meet real data, real permissions and real users.
The Agent That Passed Every Demo
The demo went perfectly. An accounts payable agent built for an SAP finance team matched invoices to purchase orders and flagged a price variance. It even drafted a polite note to the supplier. So everyone in the room wanted it live by the end of the quarter.
Production did not go that way. Three weeks in, the same agent was clearing exceptions that a clerk would have escalated. The model was the same one from the demo. The trouble, though, was its access. It ran under a broad service account that could post what no AP clerk could. It also trusted a supplier note claiming the variance was approved.
The story is illustrative and anonymised, but each failure in it maps to a control SAP now documents for SAP AI agents. So the useful question is which kind of agent to run, through which door and with which rights.
What Are SAP AI Agents?
Joule Agents are “AI agents with business process expertise that reliably execute multistep tasks across SAP and non-SAP systems,” according to SAP’s AI agents page . In plain terms, an SAP AI agent receives a goal, breaks it into steps, calls the right tools and then reports back or acts.
The term reaches past SAP’s own catalog. It also covers custom and third-party agents that work on SAP data. According to SAP’s FAQ, these agents:
choose from Joule skills, other agents and third-party applications act on a plan, reflect on the results and make recommendations stay grounded in SAP Knowledge Graph and SAP Business Data Cloud Joule Agents vs Joule Assistants vs Joule
Three names sit close together, so it helps to separate them. Joule is the conversational interface that users type or speak to, covered in our guide to SAP Joule . Behind it, Joule Assistants “use deep role and process context to coordinate Joule Agents and execute complex workflows,” in SAP’s words.
The agents are the workers underneath. They carry out the multistep tasks and reach other systems through the Model Context Protocol (MCP) and the Agent2Agent (A2A) protocol.
So a finance user asks one question, and an assistant quietly hands it to two or three agents.
SAP AI Agents vs Chatbots, Workflows and RPA
The difference is easiest to see side by side. Our post on the types of AI agents covers the broader taxonomy, while the table below stays inside SAP.
Kanerika Service
Agentic AI Services
Kanerika designs, builds and deploys goal-driven AI agents with planning, memory and tool use, and integrates them into ERP, CRM and cloud environments.
Explore Agentic AI Services Approach What it does How it decides Controls it needs
Assistant or chatbot Answers questions and retrieves records Responds to each prompt Data access rules and answer quality checks Workflow or RPA bot Runs the same fixed steps every time Follows rules written in advance Change control and exception queues SAP AI agent Interprets a goal, picks tools, acts across systems Plans, checks results and adjusts Scoped identity, approvals for high-risk actions, full audit log
That last column is the real dividing line. Once software can choose its own next step, the rights it runs under matter as much as the model behind it. Our comparison of AI copilot vs AI agent explains where assistance ends and agency begins.
SAP’s Agent Portfolio in 2026: What Exists and What Is Still Rolling Out
SAP’s portfolio grew quickly through 2026, and announcements often ran ahead of availability. At Sapphire on May 12, 2026, SAP unveiled the “Autonomous Enterprise,” as CIO reported . Its Autonomous Suite deploys more than 50 Joule assistants that orchestrate 200-plus specialized agents.
Autonomous Suite and Agents by Line of Business
The suite spans Autonomous Finance, Spend, SCM, HCM and CX, plus Industry AI, according to the SAP Architecture Center . On its product page, SAP also groups Joule Agents by line of business. The groups are financial management, spend, supply chain, human capital, customer experience, as well as migration and modernization.
So the first question for any team is whether an agent for its process already exists. When it does, building a custom one is usually the slower path.
Joule Studio, AI Agent Hub and the Model Layer
Joule Studio is where customers build their own agents. CIO also reported that Joule Studio 2.0 launched at Sapphire free through year-end. It supports Python, Claude Code and Cursor, and then deploys to a managed runtime.
Later, in October 2026, SAP News called Joule Studio “soon to be generally available,” with over 11,800 prebuilt business accelerators. The same release said SAP AI Agent Hub is “now generally available” to govern every agent, LLM and MCP server across SAP and non-SAP environments.
Underneath sits the model layer. SAP and Anthropic have announced plans for Claude to become a primary reasoning capability across SAP’s portfolio. Teams that also host their own models next to these agents usually look at SAP AI Core as the runtime.
Availability Check: Announced vs Generally Available
Dates matter, because a pilot built on a preview feature tends to slip. So the table records status as SAP stated it during 2026, with the source month in each row.
Component Role Status (as stated by SAP)
Joule Agents Prebuilt agents by line of business Sold as Premium AI, priced in AI Units Joule Studio Build custom agents, low-code or pro-code “Soon to be generally available” (October 2026) SAP AI Agent Hub Inventory and govern agents, LLMs and MCP servers “Now generally available” (October 2026) Joule Work and Joule A2A Cross-agent work and agent-to-agent calls General availability planned for Q4 2026 Bidirectional third-party agents via Agent Gateway External agents calling into SAP and back “Not yet supported” (Architecture Center, October 2026)
The Q4 2026 target comes from the Sapphire 2026 Innovation News Guide . It also notes that AI Agent Hub was once limited to SAP LeanIX customers. Recheck every row before a contract or a project plan depends on it.
Where SAP AI Agents Create Value: Use Cases by Process
SAP AI agents earn their keep on multistep work full of exceptions, where people still copy data between screens. SAP’s value model claims agents can “reduce time spent on multi-step workflows by up to 75%,” based on a footnoted assumption. That footnote describes a modelled consumer products company rather than a customer, with EUR 1 billion in revenue and 2,000 employees. So treat it as a vendor estimate rather than a measured result.
Finance: Invoice and Dispute Exceptions in Accounts Payable
Accounts payable is the most common starting point. An agent gathers the invoice, the purchase order and the goods receipt, then explains a mismatch and proposes the next step. Our guides to accounts payable automation and AI in finance cover the wider picture.
Procurement and Spend: PO, Goods Receipt and Supplier Exceptions
SAP reports that Mota-Engil removed manual goods receipt entry with a procure-to-pay agent on SAP BTP, Business Data Cloud and Joule. The agent also sped up invoice clearance, while Uniper used Joule Studio to build a custom procurement agent. Both examples come from SAP’s marketing pages, so read them as vendor claims.
Supply Chain and Order Status
Order status questions are a natural fit, because the answer sits across sales orders, deliveries and carrier data. An agent assembles that picture and flags orders at risk, and then a planner decides what to expedite. Our post on AI in supply chain also has more examples.
HR and Customer Service Case Routing
Case routing is where SAP’s Bosch example sits. SAP says Bosch replaced rule-based workflows with a Joule Classification Agent that routes customer cases in real time. HR requests follow the same pattern, since an agent reads the request, classifies it and sends it to the right queue.
Kanerika has seen the same matching problem in a project that did not involve SAP. For one expert-network client, a context-aware agent replaced manual searching across three disconnected systems and cut mismatch tickets by 80%.
Process Agent task SAP data it needs First permitted action Human checkpoint Metric
Accounts payable Explain invoice exceptions Invoices, POs, goods receipts Recommend a resolution Clerk approves any posting Exception cycle time Procurement Chase missing receipts and supplier data POs, supplier master Draft supplier messages Buyer sends or edits Open PO aging Order status Answer “where is my order” Sales orders, deliveries Read-only answer Service agent reviews escalations First-response time Case routing Classify and route cases Service cases, customer master Assign to a queue Supervisor audits a sample Misrouted case rate
The rows are illustrative patterns, not a list of SAP-delivered agents, and each one starts with a low-risk action. Our post on AI in ERP also covers ERP use cases beyond agents that act on their own.
How SAP AI Agents Work: Identity, Data, Tools and the Agent Gateway
Every SAP AI agent needs three things before it can do useful work. It needs business context to reason over, tools to act with and an identity that limits what it may touch. The Agent Gateway ties those together, because all agent traffic flows through it.
Grounding: Knowledge Graph and Business Data Cloud
Grounding is what stops an agent from guessing. SAP Knowledge Graph tells the agent what business objects mean and how they relate, such as an invoice that belongs to a supplier. Business Data Cloud then supplies the governed data products behind those objects.
If that data is wrong, though, the agent simply acts on wrong data faster. That is why bad data quality turns into a production risk once agents can act.
Tools and Protocols: Joule Skills, MCP and A2A
Without tools, an agent can only talk. In SAP’s design, they include Joule skills and other agents, as well as external applications reached through standard protocols. MCP exposes APIs and data as tools an agent can call, as our explainer on the Model Context Protocol describes.
A2A, by contrast, lets one agent hand a task to another, even when the two come from different vendors.
Identity: The Intersection Rule
Identity gets the least design attention, yet it decides what an agent can break. The SAP Architecture Center describes two operation models in SAP Cloud Identity Services.
In the first, an agent acts within a user’s context, with agent-specific limits on top. In the second, an autonomous agent runs with its own technical identity, tokens and audit trail.
Either way, the effective permission at runtime is the intersection of user permissions and agent permissions. The Agent Gateway then enforces that at every hop. So an agent working for an AP clerk can never post more than the clerk could, even when the agent itself holds wider rights.
SAP adds further guards on its side as well. Each agent gets a unique identity and reauthenticates at every step, and it runs in a hardened sandbox. High-risk processes also require explicit approval, and every action is logged. Those defaults would have stopped the agent in the opening story, because its broad service account would have been cut down to the clerk’s rights.
Adopt, Build or Connect: Choosing Your Path
This is the decision behind most searches for SAP AI agents. SAP says to use prebuilt Joule Agents when your needs match standard processes in finance, HR or procurement. When workflows are unique, build in Joule Studio, “keeping in mind that you take on full lifecycle responsibility for anything you build.”
AI Assessment
Is Your Organization Ready for AI Agents?
Score your data foundations, generative AI and AI agent readiness in a few minutes and get recommendations before you commit to a pilot.
Start Your AI Assessment → Most organizations “adopt a hybrid approach,” SAP notes. In practice, that means prebuilt agents for the standard core and a few custom agents where the business really differs.
A third-party agent joins only when it brings something SAP’s agents lack.
Adopt Prebuilt Joule Agents
Adopting is the fastest route to value. That is because SAP manages updates, the agent already knows the standard process and it runs under SAP’s security model by default. The trade-off is fit, since a prebuilt agent follows SAP’s view of the process rather than yours.
Build in Joule Studio, Low-Code or Pro-Code
Joule Studio has two flows, according to the Architecture Center:
Low-Code Flow: a visual builder.Pro-Code Flow: developers work in their own IDE with the Joule Studio CLI and a coding agent through MCP.Both flows deploy to the SAP-managed runtime and register with Joule automatically. So users reach custom agents the same way they reach prebuilt ones.
Bring Your Own Agent with SAP Cloud SDK for AI
For teams that already run an agent framework, SAP supports custom pro-code agents built with SAP Cloud SDK for AI in Python, Java or JavaScript and TypeScript. These agents run in a customer-managed subaccount and then connect to Joule through A2A. Supported frameworks include LangGraph, CrewAI, AG2, Smolagents, Google ADK and Pydantic AI.
Connect a Third-Party Agent Platform
Platforms such as Google Cloud, Microsoft Azure, AWS and IBM Cloud have two routes, per the Architecture Center. Either they consume Joule Agents through the Agent Gateway over A2A, or Joule orchestrates them. SAP Integration Suite also provides an MCP Gateway for governed access to SAP and non-SAP APIs. Still, the next section shows why this route is narrower than it looks.
Path Best fit What you need Who owns the lifecycle Limits today
Adopt Joule Agents Standard finance, HR, procurement and service processes Premium AI entitlement and AI Units SAP Follows SAP’s standard process design Build in Joule Studio Unique workflows grounded in SAP context Joule Studio, runtime consumption, builders You Joule Studio GA still pending (October 2026) Bring your own agent Teams with an existing agent framework SAP Cloud SDK for AI, customer-managed subaccount You You run, secure and update it in your own subaccount Connect third-party Agents that live in another platform Agent Gateway, Integration Suite MCP Gateway or Joule orchestration You and the platform vendor Bidirectional Agent Gateway access not yet supported Plain automation (one row) Stable rules with no judgment Workflow or RPA tooling You Breaks when inputs vary
When Plain Automation Is the Better Answer
Not every SAP task needs an agent. When the steps never change and the inputs are clean, a deterministic workflow is cheaper and easier to test and audit. Our guide to SAP automation covers those options, and our RPA services team often pairs bots with agents in one process.
Case Study
80% Fewer Mismatch Tickets with Context-Aware AI
A context-aware AI agent replaced manual searching across three disconnected systems to match survey requests to the right experts, with 40% better mapping accuracy and 22% bandwidth savings.
Read the Case Study → Whichever path you pick, an agent that needs core modifications is a warning sign, because extensions should leave SAP’s clean core untouched.
Third-Party Agents and SAP’s API Policy: What Is Allowed
Many teams assume that if an agent can reach an SAP API, it may use it. SAP’s API Policy says otherwise, and its 2026 version (v.4.2026a) spells it out. Before any third-party agent touches SAP, procurement, architecture and legal should read the clause below together.
What Section 2.2.2 Says
Section 2.2.2 of the SAP API Policy (version 4.2026a) reads as follows.
“Except through and within the limits of SAP-endorsed architectures, data services, or service-specific pathways expressly identified and intended for such purposes, SAP prohibits API use for: (a) interaction or integration with (semi-) autonomous or generative AI systems that plan, select, or execute sequences of API calls, and (b) scraping, harvesting, or systematic and/or large-scale data extraction or replication.”
Read clause (a) twice. It describes exactly what an AI agent does.
So a third-party agent that calls SAP APIs directly, outside an endorsed path, falls into the prohibited category. The policy also says customers must not bypass SAP’s API controls.
Endorsed Paths: Agent Gateway, Integration Suite and Business Data Cloud
The endorsed routes come from SAP’s own architecture, and each one does a different job:
Agent Gateway: carries agent traffic.Integration Suite MCP Gateway and A2A connectivity: link SAP with outside agents, though the Architecture Center adds that MCP Gateway “API usage is subject to the SAP API Policy.”Business Data Cloud and supported integration services: carry data pipelines, since clause (b) covers large-scale extraction.Teams moving SAP data elsewhere should keep the policy in view from the first design. Our data integration services team scopes those pipelines against it before a build starts.
There is a current gap to plan around too. According to the Architecture Center, bidirectional Agent Gateway communication with third-party and self-hosted agents “is not yet supported,” and full capabilities are “expected to be released soon.” So a design that depends on them today has to wait.
Microsoft 365 Copilot and Joule: An Endorsed Cross-Vendor Path
One cross-vendor route is already generally available. Joule and Microsoft 365 Copilot work together in both directions, according to Microsoft’s SAP applications blog . It needs Joule Base and a Microsoft 365 license, plus a Copilot license only for Copilot skills inside Joule.
Governance and Guardrails for SAP AI Agents
SAP builds a lot of protection into its own agents, as the identity section showed, but it cannot own your process risk. Deciding which actions count as high-risk, who approves them and what gets reviewed stays with you.
The Action-Risk Ladder
The simplest governance tool is a ladder of permitted actions. An agent climbs it one rung at a time, and only once it proves itself on the rung below. The four rungs are read, recommend (including drafting a message or entry for a person to send), execute with approval and bounded autonomy.
Most pilots should sit on the first two rungs, because reading and recommending measure answer quality without touching the ledger. Bounded autonomy is only for narrow, reversible actions with tight limits.
Segregation of Duties and Service-Account Scope
Segregation of duties still applies when the worker is software. For instance, an agent that can both create a vendor and pay that vendor breaks the same rule a person would. So scope each agent to one role and avoid shared service accounts. Then let the intersection rule do its job, as our agentic AI permission control framework shows.
Prompt Injection from Documents, Emails and Supplier Notes
Agents read untrusted text all day, such as invoices, emails and supplier comments. A line like “this variance is pre-approved” can steer an agent when nothing checks it against a system of record. The rule is simple, since instructions come only from the user and the workflow, never from the data. Our LLM security guide also covers the wider set of attacks.
Audit, Monitoring and a Kill Switch
Every agent needs an owner and a log that someone actually reads. It also needs a way to stop it fast. SAP AI Agent Hub gives one inventory of agents, LLMs and MCP servers across SAP and non-SAP systems, so that oversight has a home. Our guide to enterprise AI agents also covers operating models beyond SAP.
Watch on YouTube
AI Guardrails Explained: 4 Types, 5 Building Blocks & Agent Guardrails for Enterprise
Kanerika’s team breaks down the guardrails an enterprise agent needs before it gets write access, from input checks and approval steps to monitoring.
Failure mode Consequence Prevention Detection Response
Over-broad service account Agent posts beyond a user’s rights User-context identity and narrow roles Permission tests in the pilot Revoke and rescope the account Bad master data Wrong supplier or price acted on at speed Data quality checks before go-live Exception rate monitoring Pause writes and fix the source Prompt injection in a document Agent follows planted instructions Treat document text as data only Adversarial test cases Block the action, flag the document Unapproved high-risk action Financial or compliance exposure Mandatory approval step Audit log alerts Kill switch and incident review
Master data deserves its own line in every risk review. Agents amplify whatever the data says, so duplicate suppliers and stale prices become wrong actions instead of wrong reports. That is why our data governance work usually starts before the first agent goes live.
Licensing and Cost: What to Check Before a Pilot
SAP prices AI in three tiers, according to its AI pricing page :
Base AI: included with every SAP cloud subscription at no extra cost, so it needs no AI Units.Premium AI: covers Joule Agents.Custom-built AI: covers what you build and deploy with Joule Studio.AI Units and the Joule Studio Runtime
Joule Agents come in two types, Standard and Advanced, and both are priced by AI Units consumed per action. Joule Assistants carry no separate charge, because their cost follows the agents they run. Custom agents, by comparison, use the Joule Studio runtime on consumption-based pricing.
AI Units are SAP’s single virtual currency for Premium AI. They are bought annually, held in one shared pool and consumed only when Premium AI does work, such as agents executing workflows.
Unused units expire after 12 months. SAP’s public page lists no per-action price, so ask for one in writing.
Cost driver How it is measured Who to ask Evidence to request
Agent entitlement AI Units per action, Standard or Advanced SAP account team Units per action for your chosen agents Custom agent runtime Joule Studio runtime consumption SAP account team A consumption estimate for your pilot volume Integration work Project effort for gateways, data and identity Your implementation partner A scoped estimate with assumptions
ECC and On-Premise Customers
Customers still on ECC or on-premise S/4HANA face a conditional answer. At Sapphire, SAP CEO Christian Klein said most Joule assistants and agents would also come on-prem “as an interim solution,” The Register reported . That offer applies only to customers who sign the Max Success Plan, so check your commercial position first.
Checklist
Enterprise Agentic AI Checklist
A step-by-step checklist for the data, security, governance and ownership work an agentic AI pilot needs before it goes live.
Get the Checklist → Pricing is only one of the checks a pilot needs. Data readiness, ownership and security sign-off also belong on the same list.
A Staged Rollout for SAP AI Agents
The safest rollout moves in small, testable steps. Each stage has an exit test, and the agent moves on only when that test passes.
Pick one bounded workflow and baseline it. Choose a process with clear volume and a measurable pain, such as AP exceptions, then record today’s cycle time and error rate.Confirm entitlement, interfaces and data readiness. Check AI Units, the endorsed pathway and the quality of the master data the agent will read.Pilot in read-only or recommend mode. Test normal, edge and adversarial cases, including planted instructions in documents.Add approved writes with narrow tools and logging. Give the agent one write action at a time, with a named approver and a full audit trail.Scale after operational acceptance. Expand to new volumes or processes, and retest whenever the model or the SAP release changes.The pilot should not end on a feeling. Instead, it ends when five checks pass, namely answer quality, permission isolation, action correctness, audit completeness and a named escalation owner. Our guide to moving from AI pilot to production also explains how to set those thresholds.
Timelines vary too much by system setup to promise a number. Human review also stays in the loop longer for high-value postings, the pattern behind human-in-the-loop AI agents .
How Kanerika Helps Enterprises Put SAP AI Agents to Work
Kanerika’s approach starts with the workflow, and the agent comes later. First we assess the process and the SAP estate, and then we fix the data foundation. After that we design the agent with human checkpoints, run a governed pilot and hand over a run model.
Engagements follow Kanerika’s IMPACT framework, and our agentic AI services team builds goal-driven agents that integrate into ERP, CRM and cloud environments. On the SAP side, our SAP ABAP practice also covers AI-assisted development on S/4HANA, RAP, CDS and clean core. It also supports developers working with SAP Joule and Copilot under human approval.
The review-by-design pattern comes from our own agent work, which was not an SAP project. For a global expert-network firm, Kanerika built an AI compliance agent that gathers evidence and maps findings to criteria with citations. As a result, expert vetting ran 3x faster, backlog cases fell 70% and event delays dropped 40%, shifting teams “from research to review.”
Kanerika is a Microsoft Solutions Partner for Data and AI and a Databricks Consulting Partner. We also hold ISO 27001, ISO 27701:2019 and SOC 2 Type II, and our AI governance services set approval rules before the first write action. Poor master data, broad service accounts and early write paths are the risks we check first. To scope a first SAP agent with them covered, book a conversation with our team .
Case Study
3x Faster Expert Vetting with an AI Compliance Agent
An AI agent that gathers evidence and maps findings to criteria with citations, cutting backlog cases by 70% and event delays by 40%.
Read the Case Study → Adopt, Build or Connect: The Short Answer
SAP AI agents are now a real part of SAP’s platform, but they are only as safe as the rights, data and pathways around them. So start with one bounded workflow, keep the agent read-only until it proves itself and confirm entitlement and API terms in writing.
Then make the call by process. Adopt Joule Agents where your process is standard, build in Joule Studio where it is yours and route third-party agents through an endorsed path.
Frequently Asked Questions
What are SAP AI agents?
SAP AI agents are AI systems that take a business goal, plan the steps, call tools and act across SAP and non-SAP systems. Most are SAP’s prebuilt Joule Agents, grouped by line of business, but the term also covers custom agents built in Joule Studio and third-party agents that reach SAP through endorsed pathways.
How are SAP AI agents different from Joule?
Joule is the conversational interface people type or speak to. Joule Assistants sit behind it and coordinate work, while Joule Agents carry out the multistep tasks. So one question asked in Joule can trigger an assistant that routes the job to several agents, each acting within the permissions of the user who asked.
Which SAP AI agents are available today?
SAP offers Joule Agents across finance, spend, supply chain, HR, customer experience, and migration and modernization, sold as Premium AI. SAP AI Agent Hub became generally available in October 2026, while Joule Studio was described as soon to be generally available. Check SAP’s current release notes before planning a pilot around any single agent.
Can I build my own SAP AI agent?
Yes. Joule Studio has a low-code visual builder and a pro-code flow with its own CLI, both deploying to an SAP-managed runtime. Teams with an existing framework can also build agents with SAP Cloud SDK for AI and connect them to Joule through A2A. Either way, you own the full lifecycle of what you build.
Can third-party agents like Copilot or Claude-based agents access SAP data?
Only through SAP-endorsed pathways. Section 2.2.2 of SAP’s API Policy prohibits API use by autonomous or generative AI systems outside endorsed architectures. The Agent Gateway, Integration Suite MCP Gateway and Business Data Cloud are the endorsed routes, and the Joule and Microsoft 365 Copilot integration is generally available. Confirm your contract terms with SAP.
Do SAP AI agents work with ECC or on-premise S/4HANA?
Partly, and only on conditions. At Sapphire 2026, SAP said most Joule assistants and agents would also come to ECC and on-premise S/4HANA as an interim solution. The offer is limited to customers who sign the Max Success Plan, so the answer depends on your commercial agreement with SAP rather than on your technical setup.
How are SAP AI agents priced?
Joule Agents are Premium AI, offered as Standard and Advanced types and priced by AI Units consumed per action. Joule Assistants carry no separate charge. AI Units are bought annually, pooled and expire after 12 months. Custom agents run on the Joule Studio runtime with consumption pricing, so ask SAP for per-action figures in writing.
Which process should we start with?
Start with one bounded, high-volume process full of exceptions, such as accounts payable invoice exceptions or order status questions. Baseline today’s cycle time and error rate first. Then run the agent in read-only or recommend mode, prove permission isolation and audit trails, and add approved write actions one at a time.
