Integrating AI with your CRM and ERP: what works today, what is still in testing and where to start
People searching for AI integration with a CRM usually want one simple thing: to ask in plain language "which Kaufland orders are still undelivered this week" and get an answer from live data, not from last month's Excel export.
The problem is that a distributor's data does not live in one place. The customer and the quote are in the CRM, stock is in the warehouse, the invoice and balance are in the ERP, the retailer's order arrives through EDI. An assistant connected only to the CRM answers only half the questions.
Three ways to connect AI to company data
- An assistant inside the application. It runs in the CRM, sees the screen you are on and answers from the data you are allowed to see. No setup for the user.
- An MCP connector for Claude or ChatGPT. MCP (Model Context Protocol) is an open standard that lets an external assistant read data and run actions in an application, authenticated as the user. You work from the assistant you prefer while the data stays in the CRM.
- AI on documents. Supplier invoice scanning: the system extracts lines, VAT and totals, then matches them against the purchase order and the goods receipt. Here the gain is measured directly in hours of manual entry.
Where CRMconnect is today
Supplier invoice scanning (OCR plus matching against order and receipt) is an available module. The Arvis assistant, dashboards generated from a prompt and the MCP connector for Claude and ChatGPT are in testing, and the Q4 2026 roadmap includes search across company documents and AI agents for purchasing.
How to start, in 6 steps
1Put the data in one place before adding AI
An assistant does not fix scattered data. If stock is in the ERP, orders in email and quotes in Excel, the first project is integration, not AI. CRMconnect connects to your existing ERP (18 documented integrations) and brings orders from the field, the portal, the online shop and EDI into one place.
2Pick the questions, not the model
Write down 10-15 questions the team asks every day that today need someone to search three systems. For example: "which customers have not ordered in 60 days", "which items will fall below minimum stock next week", "which supplier invoices do not match the receipt". They become your acceptance test.
3Access rights are inherited, not rebuilt
The assistant must see exactly what the user sees: a rep sees their customers, the director sees everything. In CRMconnect, AI access follows existing roles and interactions are kept in the audit trail. Check with any vendor: what the assistant shows a rep with limited rights.
4Know where the data goes
Ask which model runs, where, and whether data is used for training. In CRMconnect the model is chosen per customer (Anthropic Claude, Google Gemini or OpenAI GPT) with training opt-out. Company data is not used to train models.
5Read first, act later
For the first months the assistant answers questions and suggests next steps. Actions that change data (an order, a reversal, a price) are approved by a person. Expand only once the team trusts the answers.
6Start with documents, where the gain is visible at once
Supplier invoice scanning is the easiest project to measure: count invoices per month and minutes per invoice before and after. Quantity or price differences against the order surface at approval, not at the next stock count.
Who uses CRMconnect with ERP, warehouse and AI
Pro Trade Impex runs CRMconnect connected to its ERP and warehouse, with orders from several channels and the platform's AI modules.
Questions to ask any vendor
- What data does the assistant see: only the CRM, or also stock, invoices and EDI orders?
- What happens to access rights when the question comes through Claude or ChatGPT?
- Which features are available today and which are still in testing?
- Who approves actions the AI suggests, and where is the history kept?
When it is not worth it yet
With fewer than five users and data that fits in one program, searching by hand stays simpler. AI starts to matter when the same question takes three systems and half an hour.
Last updated:
Frequently asked
What is MCP?+
Model Context Protocol, an open standard that lets an external AI assistant (Claude, ChatGPT) connect to an application with authentication, read data and run actions within the user's rights.
Can I use ChatGPT or Claude with CRMconnect data?+
The MCP connector for Claude and ChatGPT is in testing. Access is authenticated and follows the rights of the connected user.
Is company data used to train AI models?+
No. Models are configured with training opt-out, and interactions are kept in the account's audit trail.
Do I have to change my ERP?+
No. CRMconnect connects to your existing ERP and the AI works on the synced data.
Can I try it before deciding?+
Yes, 14 days, no card required.
Where CRMconnect fits
Want to see CRMconnect in action?
Personalized demo in 30 minutes - no card, no commitment.
Book a demoRelated guides
Warehouse management software for distributors: what it must do and how to choose
How to choose warehouse software for distribution: the ERP stock module, a standalone WMS or a WMS on top of your ERP, plus 8 things to check in a demo.
ComparativStandalone WMS vs. WMS on top of your ERP: which to choose
A separate WMS with migration or a WMS that sits on top of your existing ERP? Compare implementation speed, risk, cost and integration with operations.