How to Make SAP and ERP Data Ready for AI Agents, Per Workflow

Emily Winks

Data Governance Expert

Updated:Nov 6, 2026

Published:Nov 6, 2026

16 min read

See Context Agents Live Get the AI Context Stack

Key takeaways

SAP data is ready for an agent only for a named workflow, with meaning, owner, lineage and access rules attached.
A parked FI document is a BKPF row with BSTAT = 'V'. The agent sees the V, not which statuses count as posted.
Released CDS views carry business semantics for SAP's standard model. Your definitions and owners are your job.
Classify each failure as data, process, access, model or missing context before you add anything.

How do you make SAP and ERP data ready for AI agents?

Pick one agent workflow and list every SAP table, CDS view and API it touches. For an agent outside SAP, prefer views and APIs released for remote use, which carry SAP's business semantics and a stability contract. For each object, record the company-specific meaning, a named owner, lineage to the agent's read point and the access the agent runs under. Then run real questions and classify each failure as a data defect, a process or integration gap, an authorization problem, a model error or missing context.

Is your data estate ready for agents?

You make SAP and ERP data ready for an AI agent one workflow at a time: name the workflow, list the SAP objects it reads, and send each object’s meaning, owner, lineage and access rules with the rows. Atlan’s context layer brings SAP’s own descriptions, F1 field help and CDS view lineage into one graph, where your owners add the company-specific meaning an agent needs.

Map Where Context Stops Across Your SAP Estate

Give it the systems your agents read, from SAP ECC and S/4HANA to the warehouses and BI tools they feed, and what governs each today. It scores coverage per system and names the gap that breaks first. Read the skill.

Paste into a new chat

Use the skill at https://atlan.com/skills/catalog-coverage-gap-map.md to map which catalog or control plane covers each system our agents read across our SAP, warehouse and BI estate, and where that coverage stops. Ask me for whatever it needs.

Run once in a terminal

curl -fsSL --create-dirs \
  -o ~/.agents/skills/catalog-coverage-gap-map/SKILL.md \
  https://atlan.com/skills/catalog-coverage-gap-map.md

For an agent

curl -fsSL https://atlan.com/skills/catalog-coverage-gap-map.md

In SAP, an agent can read a correct row and draw a wrong conclusion, because the rule that gives a row its meaning often lives in a status code, a custom object or an analyst’s head. SAP documents the semantics of its standard model; your team records the rest in an enterprise context layer. It is the test behind what makes data AI-ready, applied to SAP.

Dimension	For SAP and ERP agents
Systems in scope	SAP ECC, S/4HANA, BW, Business Data Cloud, and the warehouses they feed
Where meaning hides	Status fields, technical names, custom fields, F1 field documentation
Key timelines	Business Suite 7 mainstream maintenance to end of 2027, optional extended to end of 2030 (SAP Support)
Unit of readiness	One named agent workflow, never the whole SAP estate

Why do AI agents misread SAP data that works for dashboards?

AI agents misread SAP data because the rules that interpret a row live outside it, often with the analysts who built the reports. A dashboard has those rules built into its queries. An agent on the same tables starts without them, one reason AI projects stall on data rather than models.

A status lives inside a field

According to SAP Help on held, parked and completed documents (SAP, 2026), a parked FI document is stored in BKPF, the accounting document header, and is “identified as parked for field BKPF-BSTAT = ‘V’”. Its lines go to parking tables, not BSEG, and transaction figures stay unchanged. The agent can see the V; it lacks the rule for which statuses count as posted. Counting BKPF headers as posted documents includes the parked ones: every row is correct and the count is wrong.

Custom code and local variants build context debt

Faram Medhora, Principal Analyst at Forrester, calls this context debt: “an accumulation of missing, conflicting, or inaccessible business context that prevents work from completing correctly.” And: “A 20-year-old SAP ECC estate exposes context debt through custom code, duplicate records, and local process variants” (Forrester, 2026). According to a vendor-sponsored ASUG survey (2026), 53% cite handling ECC customizations as their leading S/4HANA migration challenge.

Enterprise schemas break agents before SAP adds its own codes

In the original Spider 2.0 paper (Lei et al., 2024), the best agent tested solved 21.3% of enterprise text-to-SQL tasks, against 91.2% on Spider 1.0, with no SAP codes involved. SAP codes come on top of that enterprise text-to-SQL gap. Gartner’s May 2026 press release (Gartner, 2026) names the cause: “Traditional schema-based data models alone no longer suffice for agentic AI because they lack business context and data meaning.”

Where the meaning lives	Documented SAP example	What the agent needs with the data
A status value inside a table	Parked FI document: BKPF row with BSTAT = ‘V’, no transaction figures updated	The status meaning and the counting rule
Short technical names	Classic names such as BKPF and BSTAT; VDM names such as C_GoodsMovementQuery_2 read as business terms (SAP Help)	A business description for each object
Field documentation	F1 field help (data element documentation)	The documented meaning linked to the column
Custom fields	A renamed custom field appears in a data product as a removed field plus a new one (SAP Help)	Lineage, and an owner who knows they are one field

Each row is one of the context problems data teams hit when building agents: meaning an analyst carried in their head.

What does SAP’s own semantic layer already cover?

SAP supplies business semantics for its standard model through the virtual data model (VDM) in S/4HANA; your team supplies the company meaning each workflow needs. According to SAP Help on the virtual data model (SAP, 2025), CDS views expose business data stored in database tables “in a format that is based on business semantics and therefore easier to consume.” Only a subset is released for reuse.

Released views carry stability contracts (SAP, 2025): C0 for extensions, C1 for system-internal use, C2 for remote API use. A contract keeps the interface stable, one part of what data contracts for AI cover, and the VDM acts as SAP’s semantic layer for its own model. An agent outside SAP reads C2 objects exposed as APIs; C1 covers use inside the system. On ECC, list the tables and APIs the workflow reads directly.

SAP also documents:

SAP Business Data Cloud auto-detects custom fields from S/4HANA Cloud Public and Private Edition and adds them to data products (SAP Help), the way you package data products in a data catalog.
In Joule Studio, “Granting a user access to an agent does not expand their data permissions”; the agent uses the invoking user’s credentials, which SAP calls principal propagation (SAP Help).
Integration Suite can expose OAuth-authenticated OData and REST APIs as an MCP server with tools and policies you choose (SAP Help). See when to use MCP versus an API.

On a clean-core S/4HANA Cloud estate running SAP’s agents on released views, start from SAP’s standard semantics. Forrester recommends that tech leaders “keep company-specific definitions, policies, and decision criteria under enterprise control” (Forrester, 2026), and Gartner “advises data and analytics (D&A) leaders to establish a context layer as a core component of their D&A infrastructure” (Gartner, 2026). A collections agent also reads the CRM and the warehouse, and the company meaning has to follow it there.

Readiness need	What SAP documents for its standard model	What your team records for each agent workflow
Business meaning	CDS views based on business semantics; VDM naming	Your definitions, status rules and decision criteria
Interface stability	C0, C1 and C2 stability contracts on released views	Which released views and APIs this workflow may use
Agent permissions	Principal propagation in Joule Studio	Which agent may run which workflow, for whom

SAP’s model says what a field means in SAP’s terms. Whether it fits your collections agent, and who vouches for it, is your call.

How do you test whether SAP data is ready for one agent workflow?

To test SAP data for one agent workflow, list the objects it reads and run six checks on each before go-live. According to Gartner’s AI-ready data guidance (Gartner), “There is no way to make data AI-ready in general or in advance. The readiness of data for AI depends on how the data will be used.” McKinsey (2026) asks teams to define what “good enough” means for each use case.

Check	Question for this workflow	Evidence it is met	Who signs off
Scope	Which tables, CDS views and APIs does the agent read or call?	A listed set, released views preferred	AI product owner
Meaning	Does each object carry a business description and status rules?	Descriptions and field documentation	SAP process owner
Ownership	Who answers when the agent’s number looks wrong?	A named owner on each object	Data owner
Lineage	Where did the value come from, across SAP and downstream?	Column-level lineage to the agent’s read point	Data platform team
Access	Whose authorizations apply, on SAP and on any warehouse copy?	User propagation for live reads, row rules on copies, a scoped user for scheduled agents, a short tool list	SAP security
Failure call	Is a wrong answer a data, process, access, model or context failure?	A failure log with each case classified	Process and data owners

The checks take the AI-ready data checklist and an AI readiness assessment down to the SAP object, in line with the context layer evaluation criteria; sign-off is where owners and stewards certify AI-ready data.

The same BKPF table can pass for an agent that applies the status rule and fail for one that does not: the line between AI-ready and analytics-ready data, hidden when you grade the table.

What does AI-ready SAP data look like in real agent workflows?

Three SAP agent workflows show the test at work: finance questions, API actions and a mixed ECC and S/4HANA estate. The first and third are illustrative.

A finance agent answering questions about postings

Challenge. Parked documents and local status conventions change which postings count; only analysts know the rules.

Solution. The agent reads released CDS views exposed as APIs. The process owner writes the status rule into each object’s description and certifies it; the team tests real questions against an agreed SAP report, as in a talk-to-data agent blueprint and AI-ready data for talk-to-data.

What passing looks like. Counts reconcile to that report, and a reviewer sees which rule applied: the job of a context layer for financial services.

A process agent acting through OData business APIs

Challenge. A writing agent can call a valid endpoint with an invented code or at the wrong step.

Solution. Give it typed API metadata, valid code lists and the process model, and only the tools the workflow needs: the scoping an AI agent harness enforces and AI agent governance audits. Writing agents also need an approval step and a check that the record landed.

Outcome (published study). In Khan and Farea (2026), an agent using OData metadata and BPMN process models for 42 enterprise services completed 38.7% of 81 multi-step queries, against 10.3% for the paper’s baseline. Valid code lists eliminated invented codes but left invalid keys in 46% of write runs; full typed context cut invalid calls to 3% (3 of 101 runs). Most tasks still failed, so keep write access narrow when you give AI agents access to enterprise data.

An agent reading a mixed ECC and S/4HANA estate

Challenge. Mid-migration, one entity lives in ECC tables, S/4HANA CDS views and copies. In the same vendor-sponsored ASUG survey (2026), 67% were fully or partially migrated to S/4HANA, and another 23% expect to finish within 24 months.

Solution. Lineage for AI runs from the SAP object into the warehouse and BI layer the agent reads. One owner answers for each entity in both systems, and a cutover rule names the system of record per company code and period (AI-ready data lineage covers the mechanics).

What passing looks like. The agent can say which system a figure came from and who owns it, as when you implement an enterprise context layer for AI across a migration.

What goes wrong when teams point agents at SAP data?

Four mistakes recur when teams point agents at SAP data, one of them treating a data defect as a context problem.

Reading raw tables when a released CDS view exists. The agent inherits every status trap. For agents outside SAP, use C2 views exposed as APIs.
Giving the agent a broad technical user. Run it as the invoking user, as principal propagation does, with few tools; a scheduled agent gets a user scoped to one workflow.
Adding context to cover a data defect. Forrester advises classifying failed executions, data defects among them (Forrester, 2026). A duplicate customer record returns two credit exposures, and no description fixes that. Fix duplicates in the data: data quality for AI agents plus SAP Master Data Governance or its enterprise alternatives.
Declaring the whole estate ready at once. Certify one workflow, then expand.

How Atlan supports AI-ready SAP and ERP data

Atlan’s context layer puts SAP’s descriptions, field help and lineage in one graph with your other systems, where owners add company meaning and certify it.

SAP context ingestion. SAP ECC and SAP S/4HANA context ingestion is generally available, with CDS view column-level lineage and Fiori apps traced from app to CDS view to tables (Atlan release notes, 2026).
The S/4HANA crawl covers tables, CDS views, ABAP programs, transaction codes, Fiori apps and DataSources, among others; table, view and CDS view descriptions come from SAP’s DDTEXT (Atlan docs). SAP ECC is crawled too.
An optional setting, off by default, builds a glossary from SAP’s F1 help and links terms to columns. Each run refreshes the F1 text; descriptions your team adds sit in a separate field and survive the refresh.
The SAP Analytics Cloud connector is in Public Preview, following SAP data “from source to story”; the SAP Datasphere connector is in Private Preview.
Owners and certificates. Owners answer for an asset; people mark it Verified, Draft or Deprecated, and a message can name the workflow.
Context Agents. Context Agents draft a description where an asset lacks one; one crawled from SAP counts and is kept (Atlan docs).
Atlan MCP is “a hosted server, available on every Atlan tenant with nothing to install,” and “every action respects the permissions you’ve already set in Atlan,” with tools for “search, lineage, metadata, glossary, governance, and data quality” (Atlan docs).

Keep the agent’s SAP reads on SAP’s access path, such as principal propagation; Atlan permissions govern the context MCP serves. SAP objects sit in the Enterprise Data Graph beside warehouse tables and BI models. Documented lineage runs inside SAP from tables through CDS views to Fiori apps, so confirm it reaches the copy your agent reads. MCP delivers that context to agents, and teams with an MCP-connected data catalog can extend it for AI readiness: context engineering applied to the system of record.

How do you get started making SAP data ready for agents?

Start with one workflow your SAP process owners know well, then expand.

Inventory one workflow. List every table, CDS view and API it touches, marked standard or custom.
Move the read path to released objects. Use released CDS views and APIs where they exist (C2 for agents outside SAP) and record each stability contract.
Record meaning and an owner. For each object, write the description, status rules and decision criteria, and name the owner who certifies it.
Test with real questions. Classify each wrong answer as data, process, access, model or context before adding anything.

The sequence holds for any agent: see how to prepare enterprise data for AI agents, the AI readiness guide and what enterprise-ready AI agents need. On a new stack, build AI-ready data from day zero.

Why SAP readiness is earned one agent workflow at a time

SAP’s standard model carries real semantics for its own objects. The rest is your company’s: definitions, owners, lineage out of SAP, and the scope one agent may touch. Record it per workflow, test with real questions, and fix what the failure log points to: data, process, access, model or context. Comparing tools? Hold the best AI-ready data platforms to the same six checks.

FAQs about making SAP and ERP data ready for AI agents

1. Can an AI agent read SAP tables directly?

Yes, but the rows carry codes whose meaning lives elsewhere. A parked FI document is a BKPF row flagged BSTAT = ‘V’ that updates no transaction figures, so an agent that counts every header as posted overcounts. For an agent outside SAP, released CDS views exposed as APIs are the safer read path.

2. Do released CDS views make SAP data AI-ready?

Released CDS views give SAP’s standard model business semantics and stability contracts. Readiness still depends on the workflow: your definitions, status rules, an owner per object and the agent’s scope are recorded alongside them.

3. Does SAP Business Data Cloud cover custom fields?

SAP Business Data Cloud auto-detects custom fields from S/4HANA Cloud Public and Private Edition and adds them to data products. A renamed custom field appears as a removed field plus a new one, so someone must record that they match.

4. How do AI agents get access rights to SAP data?

In SAP Joule Studio, giving a user access to an agent does not expand that user’s data permissions; the agent reads SAP data with the invoking user’s credentials, called principal propagation. For an agent reading a warehouse copy, set access rules on that copy as well.

5. Does SAP ECC need to move to S/4HANA before agents can use its data?

No. Readiness is set per workflow, so an ECC workflow can pass the test today. SAP’s mainstream maintenance for Business Suite 7 core applications runs to the end of 2027, with optional extended maintenance to the end of 2030. Atlan crawls SAP ECC as well as S/4HANA.

6. How is AI used in ERP systems?

Agents answer questions over ERP finance data or update records through business APIs such as OData. Both need code meanings, process order and workflow-scoped access.

Sources

Share this article

Atlan is the Context Layer for AI. It translates business knowledge, including data definitions, working procedures, and governance policies, into context AI can actually use. This knowledge lives in a single Enterprise Data Graph that every team and AI agent can reach.

In Atlan's AI Labs benchmark, adding this context improved AI's text-to-SQL accuracy by 38%.

Atlan is recognized as a Leader across multiple Gartner reports and Forrester Waves, and is trusted by over 400 enterprises representing $10T+ in market cap, including Mastercard, Workday, General Motors, CME Group, HubSpot, FOX, Virgin Media O2, and Elastic.

Get a Custom Demo