We don't touch the ERP's core
The right approach for any integration with a mature ERP (SAP, especially) is "Clean Core": you build side-by-side extensions, via OData, IDoc, or events, and you never modify the core system's code. The core stays updatable by the vendor without breaking your integrations.
The first discovery question decides everything: which version (S/4HANA or ECC), which modules are already active (EWM, Catch Weight Management)? The answer tells us whether we need to build from scratch or just tune what's already there. AI on SAP gets built as a side-by-side extension, through OData, IDoc or events, without touching the core — exactly what SAP itself recommends through Clean Core.
What ERP does your company run?
Choose your system. We'll tell you how it connects and what the first discovery question is.
Accounting has no API — and that's not a blocker
SAGA C desktop, used by many accountants in Romania, has no API and, officially confirmed, won't get one on the desktop side. Real integration happens through direct reads of the native database (DBF or Firebird formats, depending on version) or through the official import/export flow.
We built more than 16 tools for reading, checking and reconciling on top of this system — trial balance, account card, ledger, tax filings, bank reconciliation — all read-only, without ever writing directly into the data files. AI on SAGA means, concretely, reading the native database directly — SAGA C desktop does not have an API and, officially confirmed, never will. The SAP digitalization discovery at FOX is the concrete case behind this approach — Clean Core, without touching the core, with a „Production Brain” as the flagship project. the SAP digitalization discovery at FOX.
The model gets the result, not the raw data
MCP (Model Context Protocol) is the emerging standard through which an AI model connects to an external system via a dedicated server that controls exactly what the model sees. Qlik, for example, launched an official MCP server in February 2026: Qlik does the calculation, the model receives only the already-governed result — raw data never reaches the model directly.
This pattern — calculation and governance in the source system, controlled exposure to the model — is the one we recommend for any serious AI-ERP integration, not just for Qlik. MCP servers for businesses exist specifically for this purpose: a controlled channel through which an AI agent for SAP or SAGA sees only the result, not the raw data in the system. For companies that also require EU hosting on top of access governance, we explain separately the difference between self-hosted and sovereign, so the discovery conversation starts from clear terms.
Prioritizing the integrations
Not every ERP module is worth connecting to AI in the first stage. We prioritize what has fast, measurable impact: 1-click recalculation for certain processes, a dashboard on data that's calculated manually today, alerting on exceptions that go unnoticed today.
A real example of a hidden priority: „giveaway” (product given away unmeasured) can reach a few percent of production on a line with no digital control — often the largest hidden opportunity, invisible until it is measured. AI agents for SAP always start on a single module with clean data — usually procurement or inventory — not on the whole ERP at once, to limit the risk of the first connection. Besides connectors, the system can include AI agents that read from and write to the ERP only with approval.
Who sees what, through AI
A well-built AI-ERP integration respects exactly the same access rules as the original system — an agent shouldn't see data the user behind it doesn't have access to. The central permissions registry and the scope restricted by platform/role are part of the architecture, not a security add-on bolted on later.
For modules with sensitive data (financial, HR), this governance layer is explicitly verified before connecting any model to the system.
From discovery to a side-by-side extension
Every integration starts with a clear map of what comes out of the existing system, before choosing the connection technology.
- 01We map which modules are active and what data comes out of your existing system (SAP/SAGA/other ERP)
- 02We choose the right integration pattern: a BTP extension, a governed MCP connector, or direct data reads
- 03We build and test on real data, without changes to the system's core
- 04We hand over documentation on exactly what the model sees and what stays strictly internal
A fixed price, after a short discovery call.
- audit of active modules
- integration map
- Clean Core recommendation
- 1-2 BTP/MCP extensions
- governed connection to Qlik/SAP/SAGA
- handover + documentation
- complete side-by-side architecture
- agents + reports
- maintenance retainer
Every project has a different context, workflows and infrastructure, so the price is set after a short, paid discovery and does not change along the way.
Frequently asked questions
How does AI integrate with SAP without affecting the existing system?
Through the "Clean Core" approach: you build side-by-side extensions on SAP BTP, via OData, IDoc or events, without ever modifying the core. This lets SAP keep receiving normal vendor updates, without the risk of your integration breaking on every update.
Can AI be connected to SAGA, if SAGA has no API?
Yes — by reading SAGA's native database directly (DBF or Firebird format, depending on the version), under the strict rule of never writing directly into the data files. Every change goes through the official import mechanism.
What is an MCP server, and why does it matter for AI-ERP integration?
MCP (Model Context Protocol) is a standard through which an external system (ERP, BI, database) exposes data to an AI model via a dedicated server, which controls exactly what reaches the model. Qlik, for example, has had an official MCP server since February 2026, through which the model receives only the already-calculated, governed result.
Does raw ERP data reach the AI model?
No, if the integration is designed correctly — the source system (ERP or BI tool) calculates and governs the data, and the model receives only the result filtered through that layer. That's the difference between a safe AI integration and one that needlessly exposes sensitive data.
How much does integrating AI with an existing ERP cost?
The price is set after a short, paid discovery call. We don't publish figures from client projects.
Which companies in Romania have API-less systems worth connecting AI to?
The most common case is accounting (SAGA C desktop, widely used) and older industrial ERPs. The lack of an API doesn't block integration, but it changes how it's done — through direct data reads or side-by-side extensions, not a standard API connection.
Can an AI agent for SAP write directly into the ERP, or does it only read?
It depends on the governance set at discovery — by default, an agent reads and proposes, and writing into the system stays a separate, human-approved step, especially for transactions with a direct financial impact.
Let's see what can be automated in your business.
A free 30-minute session: we'll tell you what can be automated, how long it takes and what it costs, with a fixed price after discovery.