Contents8
Every ERP publisher announces AI. Some uses save hours every week; others are mostly demos. The criterion that sorts them is simple: is the task repetitive, and is its result easy to check?
In short
Solid uses: reading supplier invoices, bank reconciliation, sorting incoming requests.
Useful uses with review: natural-language questions about the data, draft reminders.
Uses to handle with care: forecasts and decisions without review.
Many ERP automations need no AI at all: a rule is enough.
Data sent to a model must be chosen, and the model hosted where it should be.
The uses that hold
Reading supplier invoices
Extracting supplier, amount, date, lines and VAT from a received PDF, then proposing the purchase document. The accountant validates instead of typing. Odoo offers this reading in its accounting app, and an ERP without it can get it through an external workflow. With e-invoicing, part of these invoices will arrive already structured, with no reading needed.
Proposing bank reconciliation
Matching an incoming transfer to the right invoice, even when the label is incomplete. Classic rules cover simple cases; a model helps with ambiguous labels. Validation stays human.
Sorting incoming requests
Classifying emails from a shared address (order, complaint, quote request) and creating the right document in the ERP. Errors show quickly and are easy to fix.
Useful uses, with a human
Asking a question in plain language about your data (“which customers ordered less this quarter than last?”) and getting an answer, rather than building a report. Protocols such as MCP let you connect an assistant to the ERP, with limited read rights. We detailed the principle for a CRM in Running your CRM in natural language; it applies to an ERP.
Draft payment reminders, adapted to the customer’s history, fall into the same category: AI prepares, a person reviews and sends.
What can still wait
Sales and stock forecasts “by AI” are only as good as the historical data, and they often hide a classic statistical method. They help thinking, not deciding alone. Any action that commits the company (a supplier order, a credit note, a payment) must keep human approval.
Not every automation needs AI
AI everywhere
A cost per call, even for a simple rule
Results that vary from one run to the next
Human review needed everywhere
Rules first, AI next
Simple cases handled by rules, with no usage cost
AI on free, variable text
Human review where it counts
Creating an invoice on delivery, chasing an invoice overdue by 15 days, alerting when stock falls below its minimum: these are rules, and a rule is cheaper, faster and more predictable than a model. AI everywhere, without a real need, multiplies costs and makes projects drift. The right order is to automate with rules, then add AI where text is free and variable.
The data to protect
An invoice contains names, amounts, sometimes bank details. Before sending data to a model, ask three questions: which data leaves, to which provider, and is it retained? A model hosted in Europe, or on your own infrastructure for sensitive data, settles most cases. The European AI regulation also frames some uses; the European Commission’s AI Act page sets out the framework.
Measuring the gain
An AI use is judged like anything else: hours saved, errors avoided, model cost. If invoice reading saves three hours a week but requires two hours of corrections, the real gain is one hour. Measure corrections, not just processing time.
How to go about it
Choose a task, measure it before (hours, errors), connect AI on a limited scope with human approval, then measure again after a few weeks. If the gain is clear, extend; otherwise, go back to a simple rule. This method avoids AI projects that cost more than they bring. The complete ERP guide describes the ERP these uses rely on, and our AI at the right cost offer helps choose the model and hosting according to the task and the data.
Frequently asked questions
Does Odoo have built-in AI?
Yes. The publisher promotes AI features, including document reading and, in its Custom plan, agentic AI. Their availability depends on the edition and plan. For Dolibarr, these uses rather go through modules or an external workflow connected to the ERP through its API.
Do you need a special model for accounting?
No. General-purpose models read invoices and statements well. What matters is the frame: data limited to what is needed, results proposed and validated by a person, a log of what was done. The model is chosen on cost, language and hosting.
Where should you start?
With supplier invoice reading if its entry takes time, or with sorting incoming requests. Both are easy to measure: entry hours before and after, correction rate. If the gain is clear over a few weeks, extend to other tasks.
Sources
- AI regulatory framework, European Commission (new tab)digital-strategy.ec.europa.eu
Jim · founder of Brain Plus AI
Also in erp: odoo and dolibarr
Practical guide5 min The topic guide
ERP for a company of 20 to 500 people: the guide
An ERP brings sales, purchasing, stock and invoicing into one database. This guide explains when it becomes useful, how to choose between Odoo and Dolibarr, and what makes the project succeed or fail.
Practical guide4 min
Purchasing in an ERP: from request to supplier payment
Purchasing is often a company's least equipped flow, between emails, quotes as attachments and invoices typed by hand. What an ERP changes, from purchase request to payment, and the approval rules that protect cash.
Choose and price4 min
Choosing an ERP integrator: the questions that sort them
Every integrator promises a successful project. A few precise questions, on method, data, customisation and aftercare, make the difference between two proposals.