Guide

SAP AI architect: the decisions the build team cannot take alone

As of 2026-10-07

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

A SAP AI architect decides which layer of SAP's AI portfolio a use case belongs to, where its data comes from, what an agent may do, which regulation applies and who pays for each call. SAP groups the portfolio as the SAP Business AI Platform, announced at Sapphire in May 2026, in three pillars: Build, Contextualise and Reason, and Govern. On 7 October 2026 the Analytics Legends contract corpus held 62 postings whose titles combine an architect word with AI wording, 35 freelance and 27 permanent, while 95 named Joule, 87 of them permanent. The title is common; the content varies from one posting to the next.

Choose the layer before the tool

Five placements cover most cases: a feature SAP ships inside an application, a standard Joule agent that is activated rather than built, a custom agent in Joule Studio, a custom application calling models through the generative AI hub, and a pro-code agent on BTP reached over A2A. Embedded scenarios in S/4HANA have their own promotion gate, separate from AI Core's deployment step. The first duty is to say which capabilities should not be built because SAP already ships them.

Two cautions apply. A standard agent's release status, whether beta, early adopter, restricted or generally available, is itself a contract condition. And SAP's figure of more than 200 agents announced in 2026 is a roadmap ceiling, not an availability count. For discovery, SAP catalogues over 240 pre-valued use cases on its Discovery Center; the architect filters them against data readiness and risk class instead of starting from a blank whiteboard. Fewer than five quick wins after two workshops usually means the filter was too narrow.

Data and grounding: three paths, one common mistake

Datasphere grounds Joule in three different ways: navigational answers on Datasphere's own metadata, analytical insights over analytic models exposed through the catalogue, and Joule Studio agents anchored in the SAP Knowledge Graph. Document grounding on the vector store is a fourth, separate channel. Confusing them explains most tickets saying Joule does not know the client's data. A private, ungoverned view is invisible to Joule by design.

SAP's phrase AI-ready data bundles three problems: quality, semantics and lineage. Score them as three lines, each with its own owner. Zero-copy sharing of data products to Databricks or Snowflake solves duplication, not ambiguity or access control. On HANA Cloud a hybrid pattern is available: master data in the graph engine, transactions in SQL over data products, text through vectors. Do not confuse the SAP-managed Knowledge Graph with the HANA Cloud graph engine, the triple store you populate yourself.

Agents and protocols: smaller than the slide suggests

Default to one agent with three to five tools, and split into subagents or A2A peers only for a concrete isolation, ownership or reuse reason. A subagent shares its parent's guardrail profile; an A2A peer is a separate trust boundary with its own agent card, authentication and audit trail. The hop rule is simple: person to agent through the interface, agent to SAP data through MCP, agent to a foreign agent through A2A.

Two SAP constraints shape the design. SAP's API policy allows agentic use of SAP APIs only through endorsed routes; hosting an MCP server over business APIs is best done as an Integration Suite artefact, limited to 30 tools, rather than as a custom proxy. And approval design follows reversibility: confirm before irreversible writes, sample reversible high-volume work, and register every agent and server in the Agent Hub inventory.

Regulation: the classification is a design decision

The EU AI Act applies in stages. Transparency duties apply since 2 August 2026. The Digital Omnibus, in force since 27 July 2026, moved high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I. The architect's first scoping question is provider or deployer: building or repurposing an agent can make the customer a provider. Two SAP cases already reach Annex III: Joule used for performance or promotion decisions, and supplier scoring that feeds payment terms. Fines for most operator obligations reach EUR 15 million or 3 % of turnover.

The European Commission's impact assessment, as cited in our material, puts documentation at 200 to 800 person-days per high-risk system, with technical documentation alone at 60 to 120 days for a medium Joule agent integrated with S/4HANA and Datasphere. Existing ISO 27001 and SOC 2 certification cuts the lower bound by 30 to 40 %. Separately, ask three privacy questions per engagement: which model and region, whether masking covers the languages in use, and whether the contract restricts general R&D use of customer data. Not retained by the provider is a retention commitment, not a residency guarantee.

Cost: three layers, kept out of the day rate

SAP prices AI in three layers: Base AI included in eligible subscriptions, Premium AI paid in AI Units, and custom-built AI billed at runtime. In 2026 a delivered agent consumes 0.02 AI Units per action, and Joule for Consultants is the one per-user offer left, at 35 AI Units per user per month including 22,900 requests. SAP's own example: 5 users for 12 months is 2,100 AI Units, or EUR 14,700 at the EUR 7 per unit used in that example, read on 23 September 2026 and not a list price.

The rule for a statement of work is to present three lines separately: SAP's consumption cost, passed through; the consultant's day rate; and delivery effort in days. Never blend AI Units into a day rate, and add a re-pricing clause given how often SAP changed its model in 2025 and 2026.

Decision table: who decides what

A feature SAP already ships — Decision: activate, do not build · Evidence: scope item, licence, release status · Failure: duplicated build.

Custom application on models — Decision: orchestration deployment, resource groups, cost labels · Evidence: promotion gate passed · Failure: proof of concept promoted as is.

Questions over reporting data — Decision: which models are exposed and how they are described · Evidence: semantic readiness · Failure: good data, poor answers.

Agent acting across systems — Decision: tool scope, approval points, protocol per hop · Evidence: agent registered, owner named · Failure: wider access than purpose.

People or credit decisions — Decision: provider or deployer, Annex III classification · Evidence: documentation plan · Failure: high-risk system treated as a pilot.

What we cannot assert

No source we hold publishes a day rate or a certification for the SAP AI architect role. The compliance effort range is an estimate from the impact assessment cited in our material, not an observed project cost. Corpus counts mix SAP-specific and generic AI postings.

Frequently asked

What is the difference between a SAP AI architect and a SAP AI consultant?

The consultant delivers within a chosen layer. The architect decides the layer, the data route, the controls and the cost model before delivery starts. On small scopes one person can do both.

Which certification is relevant?

C_AIG_2604, the Generative AI Developer associate (3 hours, system-based, 76 % pass mark), covers the hub layer. We found no SAP certification for the architect role as such.

Does the AI Act apply to a Joule deployment?

Yes, in stages. Transparency duties apply now; high-risk duties for Annex III systems apply from 2 December 2027. Classification depends on the use case, not on Joule.