Adding AI to Your Existing ERP: How to Connect It with n8n and AI, Even Without an API

Your ERP has run reliably for years, your team knows every screen, and a replacement is years away at best. At the same time, you want orders from emails to stop being retyped, supplier confirmations to stop being checked by hand, and customer questions about delivery status to be answered faster. But vendors’ built-in AI features are usually only available in the latest cloud version.

You’re not alone. According to the 2026 investment report of DSAG, the German-speaking SAP user group, 54% of SAP users in Germany, Austria and Switzerland still run SAP ECC, whose mainstream maintenance ends in late 2027. And among the AI applications SAP customers run in production, 77% are not from SAP. Many companies aren’t waiting for their ERP vendor; they’re adding AI alongside the ERP.

The tool gaining ground for this among small and mid-sized businesses is n8n. In May 2026, SAP invested in n8n and plans to embed it in Joule Studio. In April 2026, ERP vendor proALPHA announced a strategic partnership with n8n and is building its integration platform on it. This guide shows how to connect your existing ERP to n8n and AI, which connection path suits which system, and what to do if your ERP has no modern interface at all.

Why add AI instead of replacing your ERP?

Replacing an ERP is one of the most demanding projects a mid-sized company can take on. Industry surveys cite implementation costs of around €6,000 per user and a typical project duration of nine months or more, and a significant share of projects miss their time or budget targets. Add the skills shortage: according to Bitkom, Germany is short around 79,000 IT specialists in 2026.

For most AI use cases, a replacement isn’t necessary anyway. Whether an AI can read orders from emails, check delivery times or draft dunning letters doesn’t depend on how modern your ERP is, but on whether the AI can get to the data and return results. That’s exactly what an integration layer solves.

The principle: n8n as a layer next to the ERP

Instead of rebuilding the ERP, you place a separate layer next to it. Specialists call this the sidecar pattern:

• The ERP remains the system of record. It holds master data, transactions and business logic, and it stays unchanged.
• n8n orchestrates the workflows. It reacts to triggers such as an incoming email, fetches data from the ERP, passes it to an AI model, and writes the result back through the intended path.
• AI is one step in the workflow, not a replacement for the ERP. It reads, classifies, extracts and drafts, but it doesn’t post transactions on its own.

Vendors themselves now follow this pattern: proALPHA is gradually replacing its previous middleware with an integration platform based on n8n, and SAP relies on its Integration Suite and is embedding n8n in Joule Studio.

Why n8n? n8n offers over 1,000 integrations, an AI agent node and built-in human approval steps. It can run in your own data center, which is often decisive for ERP data. According to the company, more than 1,400 enterprise customers use n8n.

Which connection path fits your ERP?

Before looking at individual systems: the check shows you in two minutes which connection path fits your situation best.

The five connection paths at a glance

Many companies combine two paths, for example reading via the database and writing via import files. The order of checking matters: always ask about an official interface first, even if your system is old. There are often licensable integration modules nobody in the company knows about.

Path 1: Modern interface (REST/OData)

The best starting point: your ERP offers a documented REST or OData interface. n8n calls it with the HTTP Request node, and for some systems there are dedicated nodes. Reads and writes go through the ERP’s business logic, with all its validations, number ranges and permissions.

How to proceed:

1. Create a dedicated service account in the ERP with exactly the permissions the workflow needs. Never use the administrator account or a personal one.
2. Use a test system. Most ERPs offer a sandbox or test company.
3. Start read-only, for example with a status query, and add write steps only afterwards.

Good to know: Some ERP vendors now offer their own MCP servers through which AI assistants can access the ERP directly. With its “MCP Client” and “MCP Server Trigger” nodes, n8n can do both: use an MCP server and act as one itself. How this works for analysis is described in our guide to analyzing ERP data with AI.

Path 2: Legacy interface (SOAP, RFC/BAPI)

Many older ERP systems have interfaces, just not modern ones. The good news: you can still work with them.

• SOAP: n8n has no dedicated SOAP node but can call SOAP services via the HTTP Request node, building the XML message within the workflow.
• RFC/BAPI (SAP): n8n doesn’t speak SAP’s RFC protocol directly. The usual approach is to expose function modules as OData services in SAP Gateway, release them via the SAP Cloud Connector, or build a small translation layer. Be careful with custom builds: SAP’s Node.js library for RFC has been archived and is no longer maintained.
• COM and proprietary interfaces: for example the DI API in SAP Business One or older Windows interfaces. These almost always need a translation layer provided by your ERP partner or a service provider.

Path 3: Read the database, write via import

If your ERP runs on-premise on a SQL database, n8n can read directly with its Microsoft SQL, Oracle, PostgreSQL or MySQL nodes. For many use cases that’s entirely sufficient, for example to look up delivery status, inventory or open items for an AI answer.

Three rules are non-negotiable:

1. Read only, never write directly into tables. Writing to the database behind the ERP’s back bypasses validations, number ranges and logs. That leads to inconsistent data and, in the worst case, voids vendor support. Writes go through the ERP’s import function or a person.
2. A replica instead of the production database. Where possible, read from a replica so queries don’t burden the live system.
3. Approved views and a dedicated read-only user. Create database views with understandable field names, without HR or salary data, and a database user that can only read these views.

Check licensing first. Some vendors restrict direct database access contractually or treat access by third-party systems as indirect use for licensing purposes. This applies especially in the SAP world.

Path 4: File exchange via CSV or XML

The underrated classic: almost every ERP can export and import data as files, many automatically into a folder. n8n picks up exports from the file system or via SFTP, processes them with AI and drops import files that the ERP reads in as usual.

Advantages: The ERP isn’t changed, the import function validates data with the normal business logic, and this path is available even for very old systems.

Disadvantages: No real time. The data is only as current as the last export, and errors only surface at import.

How to make it robust:

• Every file gets a unique name and is archived after processing so nothing is imported twice.
• At first, a person checks import files before the ERP reads them in.
• n8n immediately reports via an error workflow if a file couldn’t be processed.

Path 5: Automating the user interface (last resort)

If there’s no interface, no database access and no file exchange, what’s left is automating the user interface: classic robotic process automation (RPA) or an AI that controls screen, mouse and keyboard.

How reliable is it? On the OSWorld benchmark, which measures exactly these kinds of screen tasks, the best AI systems reach a full success rate of around 42 to 44% according to published results. Anthropic itself recommends an isolated virtual machine, no credentials for the model and human confirmation for consequential actions for its computer use, and warns that the model sometimes assumes actions succeeded without checking.

Our assessment: For posting transactions or creating orders unattended, it’s not reliable enough. As a stopgap for clearly structured, recurring inputs with oversight, it can make sense. But every ERP update can break it. More on the difference between RPA and AI agents in our article on agentic AI vs. RPA.

Major ERP systems: how to connect them

As of today, there are no official n8n nodes for SAP, Business Central or DATEV. For SAP, they are on the roadmap according to trade press. Connection is via the HTTP Request node or community nodes. With community nodes, check their origin and maintenance before entrusting them with ERP credentials.

Six workflows you can add with n8n and AI

Read first, write in stages

A staged model has proven itself:

1. Read only: The AI answers questions and creates reports. Risk: low.
2. Suggest: The AI prepares records, a person approves them. n8n has built-in approval for this: before any selected tool, a “human review” step can be added, with approval via Teams, Slack, Outlook or n8n chat. If rejected, the step is aborted and the AI is informed.
3. Write automatically: Only once a workflow has run without errors for weeks is approval removed for clearly defined cases, for example orders from regular customers below a certain amount.

Security: what n8n as a control center means

An n8n server connected to your ERP, email, file storage and AI becomes your company’s control center, and an attractive target. That’s not a theoretical risk.

The most important safeguards:

• Apply updates promptly. n8n ships new versions almost weekly; security updates should be applied within days. Version 2.0 significantly tightened the defaults: the node for executing system commands is disabled by default, code runs isolated, and file access is restricted.
• Not publicly reachable. The n8n interface belongs behind a VPN or secured access. Only the webhooks that genuinely need external access are exposed.
• Dedicated service accounts with minimal rights. Each workflow gets an ERP account that can only do what the workflow needs. For reading, use a database replica instead of production.
• Block dangerous nodes. The NODES_EXCLUDE setting lets you disable nodes such as system command execution or file access entirely.
• Manage credentials centrally and rotate them regularly. n8n’s business and enterprise editions support external secret stores.
• Separate test and production. Workflows are developed on a test instance against a test ERP.
• Data protection: For any external AI service, you need a data processing agreement. For high protection needs, the language model can also run locally, as our local AI vs. cloud AI comparison shows.

Limits: when n8n isn’t enough

n8n is an orchestration tool, not accounting software. Three limits to be aware of:

1. No transactions across multiple steps. If a workflow creates an order, then adds a line item, and fails at the third step, n8n doesn’t automatically roll back the first two. Odoo, for example, explicitly documents that each API call is its own transaction. The fix: bundle related changes into a single call, make every write idempotent so a retry doesn’t create a duplicate, and catch errors with a dedicated error workflow.

2. Core logic belongs in the ERP. Pricing, posting rules and availability checks belong in the ERP, not in an n8n workflow. Otherwise you end up with a second, undocumented business logic next to the ERP.

3. Bulk data and special formats. For certified EDI, very high data volumes or complex SAP integrations, specialized integration platforms or an ERP partner are often the better choice. n8n can orchestrate them but doesn’t replace them.

What does it cost?

n8n itself (per its pricing page, as of October 2026, billed annually):

License: n8n is under the “Sustainable Use License”. Using it for your own internal business purposes is allowed, including in the free Community Edition. Offering n8n as a paid service to third parties is not.

Additional costs:

• a server for n8n, or a cloud subscription
• AI models, usually billed by usage
• licenses or fees for ERP interfaces, such as paid DATEVconnect extensions or integration modules for older ERP systems
• building the connection and first workflows, plus ongoing operation

Published fixed prices from consultancies for an initial feasibility check are in the low four figures; for a first production workflow, usually in the low to mid five figures. That’s a fraction of an ERP replacement.

Your first AI workflow in five steps

1. Pick a workflow: a frequent, clearly defined process with measurable effort, such as orders from emails or delivery status inquiries.
2. Check the connection path: use the check above and clarify with your ERP partner which interfaces are available and licensed.
3. Set it up securely: n8n in your own network or in the cloud, a service account with minimal rights, a test system, and a data processing agreement with the AI provider.
4. Start read-only and with approval: read first, then suggestions with human approval. Measure error rate and time saved.
5. Expand: once the workflow runs stably, relax approval for clearly defined cases and tackle the next workflow.

Checklist: Is your ERP ready for AI integration?

Frequently asked questions

Conclusion: keep the ERP, add the AI

In 2026, adding AI to an ERP is no longer a question of ERP generation. With n8n as a layer next to the ERP, even older systems can be connected: through a modern interface, a legacy interface with a translation layer, a database replica, or classically via files. SAP investing in n8n and proALPHA building its integration platform on it show where things are heading.

Success comes down to three things: the right connection path for your system, clean operations with updates and minimal rights, and a staged approach that starts with reading and allows writing only after approval. Then your proven ERP gains new capabilities without you having to launch a major project.