# AI for the ERP You Already Run

**Independent AI development for ERP teams**

Custom AI modules for SAP, Microsoft Dynamics 365, and local ERPs. Built on open-source frameworks, running on your infrastructure, using any model you choose.

## Who this is for

### SAP
*Systems: SAP ECC, S/4HANA on-premise, S/4HANA Cloud, OData · RFC · BAPI*

SAP's flagship AI is reserved for its cloud programs, while most of the installed base still runs ECC or on-premise S/4HANA, facing a costly migration deadline with no on-prem AI on the vendor roadmap. Independent user-group surveys show that where SAP customers do run AI productively, it is overwhelmingly built on non-SAP solutions.

**Get AI on your SAP data now, without waiting for a migration.**

- Connect to ECC and S/4HANA through OData, RFC, and BAPI, the interfaces you already run
- AP automation, order intake, and chat-with-SAP reporting grounded in your actual processes
- Everything stays inside your landscape, in your jurisdiction

### Microsoft Dynamics 365
*Systems: Business Central, Finance & Operations, Dataverse, MCP*

Copilot is real, but it bills per seat plus consumption credits, splits awkwardly across Business Central and Finance & Operations, and struggles to ground on your customizations. Under default settings, inference for newer tenants can even be routed outside the EU during capacity peaks.

**Build the agent you actually need, on your data, with the model you choose.**

- Grounded on your real entities and workflows via Dataverse, APIs, and the Dynamics 365 ERP MCP server
- Costs scale with usage, not headcount. Keep Copilot for what it's good at
- You decide where the model runs and where the data lives

### Local & custom ERPs
*Systems: PANTHEON, SAOP, Vasco, Metakocka, e-računi, Birokrat, Custom & legacy*

If you run a Slovenian or regional ERP, your vendor almost certainly has no AI roadmap. These are small teams without AI R&D budgets. Meanwhile, Slovenia has the EU's widest AI-adoption gap between large firms and SMEs. You're not behind because you chose the wrong ERP. Nobody is building for you.

**No AI from your vendor? You don't need one. Your database and API are all it takes.**

- Most local ERPs expose a REST API or direct database access, and that's the whole integration surface I need
- Invoice capture, plain-language reporting, and forecasting, all on top of the system you keep
- Built and supported in Slovenian, German, or English

## Principles

- **No vendor lock-in:** Built on open-source frameworks like LangGraph and LlamaIndex, and open standards like the Model Context Protocol, which is now a Linux Foundation project. Any developer can pick the codebase up.
- **Your data stays yours:** ERP data is your books, your customers, your payroll. Modules run inside your network or in an EU jurisdiction you choose. Nothing is shipped to a third-party cloud by default.
- **Any model, swappable:** OpenAI, Anthropic, Mistral, or open-weight models on your own hardware. A routing layer keeps providers swappable as prices fall and models improve. Local models are already good enough for many ERP tasks.
- **You own the code & IP:** The codebase is handed over with documentation. Maintain it in-house, extend it with any developer, or keep me on a retainer. Your call.
- **Works with the ERP you have:** SAP via OData, RFC, and BAPI. Dynamics 365 via Dataverse and APIs. Local ERPs via REST APIs or read-only SQL. No migration required first.
- **A project, not a per-seat tax:** One scoped build plus usage costs you control, instead of a subscription that grows with your headcount. Ongoing costs like hosting, model usage, and maintenance are stated up front, not hidden.

## Use cases

- **Invoice capture & AP automation:** Extract invoice fields, suggest GL codes, and match against orders. People approve anything below a confidence threshold. The highest-ROI starting point for most companies, typically piloted in weeks, not quarters.
- **Order intake from email & PDF:** Parse incoming orders from unstructured mail and attachments into sales-order drafts, ready for a human to confirm. Scope depends on how varied your formats are.
- **Chat with your ERP:** Plain-language questions over your ERP data through a semantic layer with read-only access, because raw text-to-SQL alone is not production-safe for finances.
- **Demand forecasting:** Machine-learning forecasts feeding purchasing and replenishment, measured against your current forecast error. Needs clean sales history, which is worth checking before committing.
- **Anomaly & duplicate detection:** Flag outliers, duplicate payments, and unusual postings before the close, with every flag traceable to why it was raised.
- **Master-data cleanup:** Dedupe and normalize customer, supplier, and item masters. Lower risk, clearly scoped, and it improves every other AI use case that follows.
- **Order-status agent:** Answer “where is my order?” for customers or sales from live ERP data, without a support person digging through screens.
- **Document generation:** Draft quotes, RFQ responses, and reports from ERP data plus your templates, reviewed by a person before anything leaves the house.

## How it works

1. **Scope:** A short discovery: which process, which data, what the measurable baseline is today, and whether your data is ready. If it isn't, I'll say so.
2. **Pilot one process:** A fixed-scope build on one high-value workflow, with human approval on anything that posts to the ERP.
3. **Measure against baseline:** The pilot is judged on numbers agreed up front, like cycle time, error rate, and cost per document, not on a demo.
4. **Own & extend:** You get the code, the documentation, and a handover. Extend it with your team, any developer, or with me on a retainer.

## About the consultant

Built and delivered by Rok Benko, an independent AI consultant and engineer from Slovenia, operating via KAPION d.o.o. For his full AI consulting profile, services, and track record see [/ai/index.md](/ai/index.md).

## Contact

- [Book a 30-minute scoping call](https://calendly.com/rokbenko)
- Email: [roksstartups@gmail.com](mailto:roksstartups@gmail.com)

## FAQ

### Why not just wait for SAP or Microsoft to ship it?

They largely have. But it's gated behind cloud contracts, priced per seat or per consumption, and built for standard processes rather than your customizations. If you're on ECC, on-premise S/4HANA, or a local ERP, waiting means paying for a migration first and still getting generic capability.

### Isn't an independent developer riskier than a big vendor?

It's a different risk, and a managed one. You own the code and the IP, built on documented open-source frameworks and open standards, so you're never dependent on me alone. Vendor risk is real too: lock-in, forced upgrades, and pricing you don't control.

### Who maintains this after the project ends?

You do, or whoever you choose. You own a documented codebase on mainstream frameworks that any competent developer can maintain. I offer a support retainer, but you're not captive to it.

### Open-source AI frameworks change fast. Is that a problem?

They do move fast. That's why versions are pinned, tests and monitoring are part of every build, and the foundations are mature components like the Model Context Protocol, which is now a Linux Foundation project. That's more durable than betting on a single vendor's proprietary roadmap.

### What about hallucinations in a financial system?

Modules are designed so AI prepares and suggests while a person approves. Grounding on your data, a semantic layer for reporting, confidence thresholds, and human-in-the-loop approval for anything that posts to the ERP.

### Will our IT and security team approve this?

The architecture is built for that conversation: it runs inside your network with least-privilege access, often read-only, plus audit logging, and your data stays in your jurisdiction. That's usually an easier sign-off than a tool shipping data to a foreign cloud by default.

### Our data isn't clean. Can we still start?

Poor data quality is the most common reason AI projects fail, so scoping includes a data-readiness check. If your data isn't ready, a master-data cleanup is often the right first project. It pays off on its own and improves everything after it.

### We already pay for Copilot. Why add this?

Keep it. It's good at general productivity. A custom module complements it for the ERP-specific workflows it can't reach: your customizations, your posting logic, your local system.

### Is this GDPR and EU AI Act compliant?

The architecture is designed to support compliance: data stays in your chosen jurisdiction, actions are logged, and a human stays in the loop. Compliance itself is a shared responsibility that depends on how the system is operated. I build with that in mind rather than promising a rubber stamp.