Connecting AI to an ERP System: What to Clarify First

Vendors promise that AI will soon operate the ERP: ask a question, get an answer, no more input screens. For owners of small and mid-sized businesses the decisive question is a different one: what has to be settled in data, permissions and processes before an AI assistant may access the inventory system or the ERP at all? This article briefly explains what MCP is, why the ERP remains the system of record and which points have to be on the table before connecting.

Published ≈ 11 min read

Short answer

AI does not replace the ERP: the ERP or inventory system remains the binding record of stock, orders and documents, and an AI assistant can answer questions about this data through an interface such as MCP (Model Context Protocol) — for owners, purchasing and controlling, while the till, the warehouse and data entry still need a normal interface. Before an assistant is connected, the data source, permissions, limits on reading and writing, logging and data protection have to be settled; the short list follows directly below.

Short list: what has to be settled before connecting

  • Data source: which source is leading for which information, and is the data there correct?
  • Questions: which questions should the assistant answer?
  • Permissions: does the system check the signed-in user’s permissions on every call?
  • Reading and writing: does access start read-only, and does a person confirm every change?
  • Third-party text: which fields contain text written by others, and which tools are available in the same session?
  • Log: is every call recorded with user, tool and parameters?
  • Data path: where does the model run, is there a DPA, is training on your data excluded?
  • Employees: do the works council and data protection have to be involved?
  • Sources: does every number in an answer name its query or report?
  • Kill switch: who switches access off, and how is a token revoked?

Why “AI will operate the ERP” starts in the wrong place

The message now comes from many directions: instead of clicking through screens, you ask the system a question and the AI does the rest. For a beverage producer, a retailer with a product configurator or a craft business this sounds like a shortcut around the software. The idea is not wrong, but it starts in the wrong place. An assistant is only as good as the data it reads and only as safe as the rules under which it reads it. If stock sits in a spreadsheet, orders in a SaaS tool and prices in the owner’s head, even a good model answers questions on the basis of contradictory numbers, only faster and more convincingly worded. And whoever gives an assistant write access before it is settled who may do what shifts control from the business to a language model. The real work therefore comes before connecting: in the data, the permissions, the log and the question of which answers the business actually needs.

What MCP is — and what it does not do

MCP in two sentences

MCP (Model Context Protocol) is an open protocol through which an AI assistant can call defined tools of a system, such as “read open orders” or “look up the stock of an item”. The system defines which tools it offers and which inputs they expect, and the assistant chooses when to call which one (as of November 2026, specification version 2026-07-28 on modelcontextprotocol.io; the protocol changes quickly).

What the protocol does not provide

MCP does not create an interface to the ERP where none exists, and it does not clean data. Someone has to build the tools that access the system and decide what they may do. According to the specification, the protocol itself cannot enforce its security principles; implementations should build consent and authorization flows and access controls. For tools, the specification also recommends that a person can deny calls and that tool usage is logged for audit purposes.

The first question for your software vendor

Does your ERP or inventory system have a documented interface on which such tools can be built, or does the vendor already offer its own MCP server? That determines the effort more than the choice of AI model. If there is neither, the project does not start with AI but with access to your own data.

Can AI replace an ERP system?

The ERP remains the system of record

Stock, orders, invoices and documents need one place that is the binding record, with history and traceable changes. A chat history is not that place. The assistant reads from the system of record and writes, if at all, only through its rules.

Those who analyse and decide

Owners, purchasing and controlling ask questions that span several tables and reports: which customers order less often than last year? Which quotes from the configurator were never ordered? Which jobs are finished but not yet invoiced? Here an assistant can take over the gathering through exports and follow-up questions.

Those who repeat operations

At the till, in the warehouse or at order entry, the same action repeats many times a day. A button is faster, clearer and cheaper for that than a sentence to a language model. A chat is an expensive way to press a button; its value lies in questions that cut across data and calculations.

Calculation is the program’s job

Totals, contribution margins, reorder suggestions and deadlines are calculated by queries and programs that always return the same result for the same input. The language model understands the question, picks the right query and explains the result. If it calculates itself, the result is not reliably reproducible and hard to verify.

The ten points in detail

Where the data lives and whether it is correct

Excel, SaaS tools, ERP: every piece of information must have one leading source. If there are three stock figures for the same item, the assistant may take any one of them, and nobody sees which.

Which questions the business asks

The questions you answer today by phone call, export or gut feeling should be written down. They determine which tools are needed. An assistant without concrete questions is a tool without a job.

The system checks permissions, not the model

Who may see which customers, prices and margins and who may change what must be checked by the connected system on every call against the identity of the signed-in user. This is a requirement on the architecture, not default behaviour: if the assistant is connected under a single service account, every user sees everything that account sees. An instruction to the model such as “do not show salaries” is not access control.

Read-only first

The first access should only read, and even then there needs to be a limit on how many records a single call may return. Changes should appear at most as a draft that a person reviews and confirms in the normal interface. The assistant must not trigger payments, cancellations or postings.

Third-party text in the data

Payment references, order comments, e-mails and supplier price lists are text written by third parties. It can contain instructions to the model (prompt injection). A session that reads such text must therefore not have tools that write or send data outside.

A log of every call

Every call must be recorded: who asked what, which tool ran with which parameters and what came back. It also has to be settled where the log is kept, for how long and who may see it. Without a log you can neither trace an error nor explain a wrong answer.

Where the data goes

It must be settled whether the model runs with a provider in the EU or locally, whether there is a data processing agreement (DPA) and whether it is contractually excluded that your data is used for training. Whether an EU region at a US provider resolves the question of access from third countries, and whether the whole setup holds up under the GDPR in your case, is something to have your data protection officer or a lawyer check.

Employees and the works council

If the system logs who asks what, it creates data about employee behaviour. Whether the works council therefore has to be involved and whether transparency obligations under the EU AI Act apply should be checked before launch.

Every number with a source

Every number in an answer should name the query or report it comes from. A number that cannot be traced back is no basis for an order or a decision.

Kill switch and revocation

Access must be possible to switch off in one step, and every access token must be revocable individually. If you only settle this in an emergency, you settle it too late.

When it is not worth it

If stock and prices sit in spreadsheets that nobody treats as the binding record, and the process lives in the owner’s head, AI access makes the situation worse, not better. The assistant then delivers fluently worded answers based on numbers that nobody trusts even today, and hides the fact that the source is missing. The same applies if it is still disputed who may approve an order or change a price: AI does not settle that dispute, it bypasses it. In this situation the first step is a process review: where is the data created, who maintains it, where does the process break? Only once there is one leading source and clear permissions is it worth asking which answers an assistant should deliver from it.

What we do with it ourselves

  • Own development work: we use MCP in our own development work; we work with AI assistants and MCP tools ourselves.
  • Client projects: we have implemented AI in client projects. Two of them are described as case studies on this site: a jewelry manufacturer (AI integration in partner sales-point systems) and a care service (AI odometer recognition).
  • Access to client systems: we design and build MCP access to client systems. Our approach: read-only first; changes only with confirmation by a person; a log of every call; permissions checked by the system.
  • Pilot: we are building the first pilot on our own product Beistella, a reservation system for restaurants. Access will be read-only. The pilot is being prepared; it is not running in production.

Key takeaways

  • The ERP or inventory system remains the system of record; the assistant is an additional interface for roles that analyse and decide.
  • MCP is a protocol for tool calls, not a finished interface. Permissions, confirmation and logging have to be provided by the implementation.
  • Calculations belong to queries and programs, not to the language model. Every number in an answer needs a source.
  • Read-only first. Changes as a draft that a person confirms; no payments.
  • Third-party text in the data is an attack path. Sessions that read it get no tools that write.
  • Data protection, the works council and the AI Act are checked before launch, not after.
  • If the data sits in unmaintained spreadsheets, the process review comes first, the assistant later.

FAQ

Can AI replace an ERP system?
No, not in the sense that matters to the business. The ERP or inventory system remains the binding record of stock, orders and documents. An AI assistant can answer questions about it, above all for owners, purchasing and controlling. For the till, the warehouse and data entry a normal interface stays faster and clearer.
What is an MCP interface?
MCP (Model Context Protocol) is an open protocol through which an AI assistant can call defined tools of a system, such as reading data or running an operation. An MCP interface to the ERP is a set of such tools that someone has to build and secure. As of November 2026, specification version 2026-07-28.
Is AI in the ERP GDPR-compliant?
That depends on the implementation and cannot be answered across the board. Things to check include: where the model runs (EU or locally), whether a data processing agreement is in place, whether training on your data is excluded, which data the assistant sees at all and how long logs are kept. The legal assessment belongs with your data protection officer or a lawyer.
Should AI be allowed to change data in the ERP?
Not at the start. The first access should only read. Later the assistant can prepare changes as a draft that a person reviews and confirms in the normal interface. It must not trigger payments, cancellations or postings. Permissions must be checked by the system against the signed-in user, not by the model.
Where do we start if our data is in Excel?
With the process, not with AI. First settle which spreadsheet is leading for which information, who maintains it and where the process breaks. The free process check shows first weak points; the process review for €350 ends with a written document on it.

Settle data and permissions first, then connect the assistant.

The free process check runs automatically and shows where your process loses information. In the process review for €350 a person looks at your process; it ends with a written document: where the process breaks and what of it can be automated. You can bring the questions from this article to the first call.