Context
An SAP-heavy manufacturing enterprise with a global footprint across FI, MM, SD, and HCM modules had a chronic operational-load problem. The SAP estate itself was healthy: S/4HANA in the core, ECC in two regions still in migration, proper master data, and a competent in-house SAP team. But a large volume of routine work — three-way matching, order status checks, exception review, operational report generation — consumed specialist SAP time that should have been spent on higher-value configuration and analytics work.
The client had looked at SAP's own AI offerings and at third-party RPA vendors. SAP's roadmap was promising but not available on the timeline the client needed. The RPA tools were brittle — they broke whenever SAP changed screens or a transaction code was updated, and they did not handle the decisioning portion of exception work.
Challenge
The client wanted an agent layer that:
- Connected to both S/4HANA and ECC simultaneously, so the global rollout did not depend on finishing the migration first.
- Handled clean-path work autonomously with full audit.
- Escalated exceptions with the AI's reasoning attached so the specialist could resolve fast.
- Produced continuous operational reporting instead of the daily/weekly batch that the SAP team was maintaining by hand.
Approach
Thoughtwave deployed TWSS eAlliance AI — our production SAP and ERP agentic automation layer — connected to the client's S/4HANA and ECC instances via BAPI and OData. Engagement arc:
- Discovery (3 weeks). Inventoried candidate workflows by module, quantified volume and exception rate, and agreed the first three automations: three-way match close in MM, order-credit-check in SD, and posting reconciliation in FI.
- Integration (4 weeks). Stood up the agent orchestration layer on Node.js with Express and EJS (the existing stack the client's infrastructure team supported), added the SAP connector library, and wired the security controls: Helmet, CSRF, rate limiting, scoped service accounts.
- Rollout (8 weeks). Shipped the three automations in sequence, each running in shadow mode for two weeks before cutover, then live with human oversight for another two weeks. Added the operational dashboards in parallel.
What we built
The production system has four layers:
- SAP connector layer. BAPI and OData clients for S/4HANA and ECC, with scoped service accounts and read/write permission matrices per workflow.
- Agent orchestrator. Node.js + Express service that evaluates trigger conditions, calls the agent planner, and dispatches the SAP calls. Multi-tenant so both S/4HANA and ECC run in the same orchestrator.
- Decision layer. The agent calls a reasoning model for exception classification and routing; for clean-path work, rule-based logic is preferred for auditability.
- Reporting and dashboard layer. Continuous aggregation from the orchestrator's audit log feeds BI dashboards; ops teams see live status instead of waiting on batch reports.
Outcomes
- Routine SAP tasks automated across the three pilot workflows, with materially reduced specialist time spent on clean-path work.
- Faster exception resolution. Exceptions arrive with the AI's analysis attached, so the SAP specialist spends time deciding, not investigating.
- Continuous operational reporting. The SAP team's batch-report work is eliminated; dashboards update live as transactions post.
- Zero vendor lock. The agent layer runs on infrastructure the client controls; SAP roadmap dependencies do not block deployment.
What's next
The next phase extends the agent coverage to HCM (routine employee master updates, org structure changes) and to cross-module scenarios where a single business event touches FI, MM, and SD simultaneously. The operational dashboards are being extended with predictive signals — AI-surfaced indicators that an exception class is about to spike based on upstream data — so operations teams can prevent rather than react.
For deeper context on our SAP and enterprise applications practice, see our Digital Enterprise Applications service.
Why the agent layer approach wins over RPA
The classic RPA (robotic process automation) approach to SAP automation — screen-scraping macros driven by brittle selector scripts — has a well-known failure mode. When SAP upgrades, the transactions move, the screens change, and the RPA bots break en masse. Enterprise RPA teams end up spending more time maintaining bots than building new ones. The agent layer approach solves this two ways: the integration goes through BAPI and OData where possible (stable, versioned interfaces), and where screen-level interaction is unavoidable, the agent's reasoning layer can tolerate minor UI changes that would break a hard-coded RPA script.
There is a second, subtler benefit. RPA bots execute a script; an agent with a reasoning model evaluates the situation. When a three-way match has an unusual line-item variance, an RPA bot either follows its rigid exception rule or fails over to a human. An agent can look at the variance, weigh it against policy and historical patterns, and make a graded recommendation — automate the clean case, flag the edge case with analysis, escalate only the truly novel case. For operational workflows with long-tail exception patterns, that difference compounds.
What clients measure
The three metrics the client uses to evaluate the rollout after the pilot:
- Specialist-time reclaimed from routine keying and status checks, redirected to higher-value configuration and analytics work.
- Exception resolution time — how long between exception surface and resolution, with the AI analysis pre-attached.
- Operational reporting freshness — the move from daily batch reports to continuous live dashboards.