How to add AI to your ERP without switching systems

When someone suggests “adding AI to the company”, the first question is usually whether you’ll have to switch ERP. It’s a reasonable question: you’ve had that system for years, it holds your customers, your invoices and the way you work, and the idea of migrating it is daunting.

The short answer is no. The AI connects to your ERP as one more piece, and the ERP remains the place where the real data lives. What you do have to decide is how it connects, and that decision sets the cost and the risk of the project.

Your ERP stays where it is

Think of the AI as a very fast new employee who doesn’t have a user account in the ERP. To do anything useful, it needs to be able to read what’s inside and, in some cases, leave something written. Everything else (your screens, your reports, your invoicing) stays the same.

You can see this clearly in Cerrajería Carlos Rodríguez’s management software. Next to the item and description fields of each quote there’s a button. When pressed, the system sends that text to an AI model (Mistral, today) and shows the original and the proposal side by side. The person accepts, rejects or asks for another version. Nothing was migrated and no screen changed: a button appeared where the friction was.

Three ways to connect AI to your ERP

Through the ERP’s API

An API is a documented door the vendor opens so other programs can read and write data without touching its internals. Many cloud ERPs, such as Holded or Odoo, publish one. If your ERP has one, it’s almost always the best route: the vendor guarantees that what you write goes through its validations, and an ERP update is much less likely to break your integration.

What to check: what it lets you read and write (sometimes the API falls short exactly where you need it), whether it costs extra and what usage limits it imposes.

Through the database

If there’s no API, or the one there is doesn’t go far enough, you can work directly against the ERP’s database. It’s fast and gives access to everything. For example, EPMTIC’s customer support application reads and writes directly to the database of their ERP, E3M, for customer and contract data. It isn’t an AI project, but the connection route is the same.

There’s a catch in that example: I develop E3M myself, so I know what each table means and which rules have to be respected when writing. With a vendor’s ERP things change. Reading from its database is usually safe. Writing to it while bypassing the program’s logic can leave inconsistent data the ERP can’t handle, and an update can change the tables without warning. My rule: reading from the database, yes; writing, only if there’s no other option and with the vendor in the loop.

Through files and exports

The humblest route, and the one that almost always works. The ERP exports a file (a CSV, a spreadsheet, an XML), the AI works with it and leaves another file the ERP imports. Email counts here too: an inbox where supplier invoices arrive, the AI extracts the data and prepares the import file.

It’s slower and less elegant, but it asks nothing of the ERP beyond what it already knows how to do. For an old ERP without an API it’s often the sensible place to start.

Where the AI goes

There’s one detail worth demanding from whoever builds it for you: the call to the AI model goes on the server, not in the browser.

In the locksmith’s case, the button calls the management software’s backend, and the backend is what talks to Mistral. That has three very practical advantages. The API key never reaches the user’s computer. You control exactly what data leaves, which in this case is only the text of the field, without the customer record. And changing model or provider is done in one place. That software has already changed once: it started with a model on their own server and moved to Mistral in August, and the screen didn’t even notice.

Which use cases work well

The ones that work best have something in common: a repeated task, clear rules and a person who can review the result in seconds.

• Drafting or polishing texts that end up in an ERP document: quotes, job descriptions, replies to customers.
• Extracting data from documents (invoices, delivery notes, orders arriving by email) and preparing their entry into the ERP.
• Querying in natural language: “which of this customer’s orders are still pending?”, without going through three screens.
• Feeding the ERP from other channels, such as WhatsApp or the phone, so that a request becomes an order or a ticket.

What works worse: decisions with a lot of judgement and few rules, or processes the company itself isn’t clear about. AI doesn’t tidy up a messy process; it automates the mess.

The risks to watch

Writing without anyone looking. In testing after the model change at the locksmith’s, a description with no price at all came back with an invented one: €186. It was the format example I had put in the instructions, and the model copied it as if it were data. It was fixed, but the real defence is the design: the AI proposes and a person approves before saving. Always start that way.

Personal data. Your ERP is full of third parties’ names, addresses and amounts. Send the model only what’s strictly needed, choose a provider that signs a data processing agreement and check that it doesn’t use your data for training. In the Mistral account used by the locksmith’s software, for example, that use is switched off. If your data can’t leave your infrastructure, the model has to run on your servers, and that’s a separate project with its own budget (more in how much it costs to integrate AI).

Not knowing what it did. Everything the AI proposes or writes has to be logged: what went in, what came out and who approved it. When something fails, and at some point it will, it’s the first thing you’ll need.

When you really do need to change the ERP

If your ERP was already holding you back before anyone mentioned AI (it doesn’t allow integrations, nobody dares touch it, everything goes through parallel spreadsheets), AI won’t fix that. That’s a different conversation, and I cover it in modernizing a legacy ERP: when to migrate and when not to.

But if your ERP works and you only want to take repetitive work off your people, leave it alone. Before asking for a quote, ask your ERP vendor three questions: does it have a documented API? What does it let you write? Does access cost extra? With those three answers you’ll know which of the three routes is yours. If you’d like to look at it together, start with the diagnosis.

Resources

• Applied AI for companies: use cases, process and prices.
• How much does it cost to integrate AI into a company: the budget items and what drives them.
• AI agent or chatbot: which one you need
• Case study: AI in Cerrajería Carlos Rodríguez’s management software
• Custom development services