AI & Data

MCP in practice: when connecting AI to your systems pays off

The Model Context Protocol brings AI inside your ERP, CRM and database. A guide to deciding where it pays off and where it only adds risk.

PublishedReading time2 minutes

In 30 seconds

  1. 01MCP is an open standard that lets AI query and act on your systems under clear access rules.
  2. 02It pays off when a task spans two or more systems and today depends on someone copying and pasting.
  3. 03Start read-only, with one process and one team. Write access comes later, with human approval.

Every company that has tried generative AI hit the same wall: the model writes well, but knows nothing about order 4,512, yesterday’s stock levels or the customer’s contract. The useful information lives in your systems, and the AI sits outside them.

What MCP changes

The Model Context Protocol1 defines a standard way to expose tools and data to an AI assistant. Instead of a different integration for each model, you write one MCP server for your ERP and any compatible client can use it, within the permissions you define.

In practice, it turns questions like “which of this customer’s orders are late, and why?” into a real query, on real data, answered in seconds.

The gain is not that the AI knows more. It is that it stops depending on someone copying data from one screen to another.

Where it pays off

Across the projects we follow, three patterns show up whenever the return is clear:

01Customer service that checks three systems. The order in the ERP, history in the CRM, delivery at the carrier. The AI pulls it together and the agent decides.
02Reports that are born in spreadsheets today. Recurring leadership questions answered straight from the database, with the query on record.
03High-volume operations with clear rules. Invoice checks, ticket triage, reconciliation. The AI prepares, a person approves.
AI assistant MCP server permissions · logging ERPCRMDatabase
Figure 1. The MCP server sits between the AI and your systems: it decides what can be read and what can be changed, and logs every call.

Where it does not pay off yet

If the process relies on a single system that already has a good screen, the AI becomes just another path to the same place. The same goes for data nobody trusts today: connecting AI to an inconsistent database only produces wrong answers faster.

In practice · quick test

Before approving an MCP project, answer:

  • Does the task span at least two systems?
  • Is the data involved reliable today, without AI?
  • Is there a person who approves the result before it becomes an action?

Three “yes”: a good candidate. Any “no”: fix that first.

The risks nobody shows in the demo

An AI with write access to your ERP is a new user: very fast and with no common sense. Least privilege, logging every call and volume limits are not implementation details: they are the project.2

Where to start

  1. Pick one process, with a known owner and volume.
  2. Expose read-only access and measure the time saved per week.
  3. Allow writes only with human approval and an audit trail.

Footnotes

  1. An open standard published by Anthropic in November 2024 and adopted by several assistants and tools since then. ↩

  2. Treat the MCP server like any exposed API: authentication, per-user scope and alerts for unusual usage. ↩

Keep reading

See all insights →
Contact

Book a diagnostic.

Leave your contact. In 30 minutes, with the people who will do the work, we understand your situation and point out the next step. No sales deck.

  1. 01You send your contact and we reply within 1 business day.
  2. 02We schedule the 30-minute diagnostic.
  3. 03You receive our read of the situation and the next steps.

* Required fields