Connect AI to your ERP, without letting it inventThe MCP protocol links the model to your real data; a knowledge graph gives it a reliable memory.

A general-purpose AI assistant knows neither your customers, nor your stock, nor your rules. Connected to your ERP, it becomes useful: it answers from your real data, cites its sources and prepares work your teams validate. This article describes the method we use, from the connection protocol to opening the assistant to end users.

Your situations

Why an AI "connected to the web" is not enough

It answers confidently on data it does not have: it invents.

It ignores your business rules, your statuses, your naming.

It does not cite its sources, so nobody can verify it.

Pasting your data into a chat exposes it and does not scale.

Plain text search misses the links between customer, order and invoice.

What we deliver

What the method produces

An MCP connector to your ERP

The model reads and writes in your ERP through declared, framed actions, with your access rights, never in direct access to the database.

A knowledge graph as memory

Your business entities, customers, products, orders, are linked in a graph the assistant queries to answer on connected facts, not on approximate text.

Continuous enrichment

The graph updates as data changes, so answers stay correct without manual reindexing.

A controlled opening to users

Once proven internally, the assistant opens to your teams then your customers, with per-person rights and a trace of every exchange.

How it works

The method, step by step

  1. 01

    A use case, not a technology

    We start from a frequent question or a time-consuming task, measurable. Technology comes after the need.

  2. 02

    The MCP connector

    We expose the ERP's useful actions to the model: read a record, find an order, prepare a quote. Each action is declared, tested, logged.

  3. 03

    The knowledge graph

    We model the entities that matter and their links, then connect continuous feeding from the ERP and your documents.

  4. 04

    The guardrails

    The AI proposes, the human validates; every answer cites its source; no personal data is sent without necessity. We measure quality on a sample.

  5. 05

    Production and opening

    We deploy, operate, watch cost and quality, then widen access to teams and customers.

What the MCP protocol is

MCP, for Model Context Protocol, is an open standard published by Anthropic. It defines a clean way to connect a language model to a business system: the system declares "tools", precise actions the model can call, and the model uses them instead of guessing.

In practice, your ERP exposes actions such as "find a customer", "list overdue orders" or "prepare a credit note". The model never touches the database directly: it goes through these actions, with your access rights, and every call is logged. The reference documentation is at modelcontextprotocol.io.

What a knowledge graph is

A knowledge graph stores information as nodes and links: a customer is linked to their orders, each order to its lines, each product to its supplier. It is close to how a human connects facts.

This structure changes the quality of answers. A classic text search retrieves passages that "look like" the question. A graph follows the links: to the question "which customers have an overdue invoice on an out-of-stock product", it walks the relations rather than hoping to hit the right paragraph. That is what separates a reliable answer from a merely plausible one.

Why enrich the graph continuously

Your data changes every day. A frozen graph ages and the assistant starts answering wrong without warning. Continuous enrichment connects the graph to the same flows as your ERP: a created order, a changed status, an added document update the graph right away. The assistant stays aligned with reality, without weekend reindexing.

The guardrails, non-negotiable

  • The AI assists, it does not decide. It prepares, proposes, computes; a person validates before an action counts.
  • Every answer cites its source. The user can trace back to the original data and its date.
  • Personal data stays protected. Nothing is sent without necessity, in line with the French data authority's AI guidance and the European AI Act.
  • Quality is measured. A sample of answers is reviewed regularly to catch drift, especially after a model change.

Opening the assistant to your users

Once the tool is proven by your teams, the same architecture opens to your end customers: each queries the assistant on their own data, with strict rights and a trace of every exchange. This is the shift from an internal tool to a service, and it is where AI connected to the ERP creates value visible to the customer.

What we know about this architecture

We design and operate this architecture in production: an assistant connected to an ERP through MCP, backed by a continuously enriched knowledge graph, with the guardrails above. That is what lets us talk about it concretely, call costs and edge cases included, rather than in theory.

References

Proof

What this method rests on

  • An assistant + MCP + knowledge graph architecture operated in production, with continuous enrichment and opening to end users in preparation.

  • Odoo ERP integrations for SMEs and a local authority, with declared actions connected to external services.

  • Operating AI systems over time: tracking call cost, latency and answer quality.

Frequently asked questions

Questions we are asked

What is MCP, in one sentence?

MCP, or Model Context Protocol, is an open standard published by Anthropic that lets a language model call precise, framed actions in a business system instead of guessing, with your access rights and a trace of every call.

Why a knowledge graph rather than plain search?

Because a graph follows the links between your data, customer, order, invoice, product, where a text search only retrieves similar-looking passages. On a question crossing several entities, the graph gives a reliable answer, search only a plausible one.

Do we need to change ERP to connect an AI?

No. The method is added on top of your existing ERP, Odoo, Dolibarr, Sage or a custom build, through the MCP connector. We do not replace your management tool, we add a framed access layer for the model.

Can the AI change my data on its own?

Only if you allow it, and always under human validation for actions that matter. By default the assistant reads, proposes and prepares; a person validates before a write takes effect. Every action goes through the connector, with rights and a trace.

How long for a first useful version?

A few weeks for a first use case in production, starting from a precise need. We deliver that case, measured, before widening. You keep the code, the documentation and the access.

A use case in mind?

Thirty minutes to start from a concrete task and see whether the method applies.

Book a free diagnostic

Nous utilisons des cookies pour analyser le trafic et améliorer votre expérience. En savoir plus