Odoo & ERP

AI Agents for ERP Systems: A Practical Guide

Updated 8 min read

AI Agents for ERP Systems: A Practical Guide

An AI agent can work with any ERP that offers three things: an API to read and write records, a permission system that can restrict the agent to its job, and an audit trail that shows what it did. Odoo, SAP, Oracle NetSuite and Microsoft Dynamics 365 all provide these in their own ways. The agent design is largely the same across them. What changes is the integration work and the vendor's rules.

Key takeaways

  • An ERP agent is a language model that calls a small set of defined tools, each of which performs one controlled operation through the ERP's API.
  • The ERP stays the system of record. The agent reads from it and writes drafts to it. It does not hold its own copy of the truth.
  • Readiness depends more on API access, clean master data and a test environment than on which ERP you run.
  • Some ERP work should stay deterministic or human: period close, payments, payroll and tax decisions.

Where this guide fits: This guide is ERP-neutral. For Odoo specifically read the Odoo AI agents guide and how to connect an AI agent to Odoo.

What an ERP agent is

An ERP holds the records a business runs on: customers, suppliers, products, orders, stock, invoices and ledgers. People interact with it through screens, forms and reports. An ERP agent offers a second way in. It accepts a request in plain language or an incoming document, works out which records are involved, and uses the ERP's API to look them up or prepare new ones.

It is different from the automation ERPs have always had. Workflow rules and scheduled jobs are deterministic: when a condition is met, a fixed action runs. An agent handles the cases where the trigger is a message, a PDF or a question, and something has to interpret it before any rule can apply.

Four things an agent needs from any ERP

1. An API

The agent must reach records programmatically. Modern ERPs expose web APIs for this. Check four points for yours: which objects are available through the API, whether it supports writing as well as reading, what rate limits apply, and whether your licence or hosting plan includes API access.

2. Scoped permissions

The agent should connect as its own integration user or service account, with a role that grants only what its tasks require. If the ERP supports restrictions by company, department or record, use them. The principle is least privilege: an agent that drafts sales quotations has no reason to see payroll or change system settings.

3. An audit trail

You need to be able to answer who changed what and when. ERPs generally record the user and time on created and changed records, and many keep field-level change history. Because the agent has its own user, its work is distinguishable from a person's. Add your own log on the agent side covering the request, each tool call and the result.

4. A place to test

Never develop against live data. You need a sandbox, a test company or a recent copy of the database where an agent can create and change records without consequences. If your ERP plan does not include one, arranging it is the first task.

The same architecture across ERPs

Whatever the system, a sound design has the same layers.

  1. Channel. Chat, email, a messaging app, or an event from the ERP.
  2. Agent service. Holds instructions, calls the model and runs the tool loop.
  3. Tool layer. Narrow functions such as "search customer", "get open orders", "create draft order" and "post note". Each validates its inputs and applies business limits in code.
  4. ERP connector. The code that speaks the specific ERP's API. This is the only layer that differs much between systems.
  5. Approval and logging. A review step for anything financial and a record of every action.

Odoo, SAP, NetSuite and Dynamics 365 at a high level

The table below describes each platform in general terms. Products, editions and interfaces change, so confirm the details against the vendor's current documentation for the version you run.

ERPGeneral characterWhat that means for an agent project
OdooModular suite with one shared database across apps, in Community and Enterprise editions, hosted or self-hosted. External API over XML-RPC and JSON-RPC, extendable with Python modules.One API reaches most business data. Access rights and record rules apply to the API user. Options depend on hosting.
SAPA family of ERP products widely used by large enterprises, with web service APIs and a mature authorization model.Integration is usually governed by an internal IT or basis team. Expect formal change control and plan for it.
Oracle NetSuiteCloud ERP with web service APIs, role-based permissions and its own scripting and customization platform.A dedicated integration role is the natural home for the agent. API usage is subject to the platform's governance limits.
Microsoft Dynamics 365A set of business applications for finance, operations, sales and service, with web APIs and security roles, part of the wider Microsoft ecosystem.Identity and access are typically managed through the organization's Microsoft tenant, so involve whoever administers it.

Several vendors also offer their own built-in AI assistants. Evaluate those first for standard tasks. A custom agent makes sense when you need behaviour specific to your process, a connection across several systems, or control over the model, data flow and approval rules.

Use cases that work on any ERP

  • Questions about data. "Which orders for this customer are still open?" answered from live records.
  • Order intake. Emailed purchase orders turned into draft sales orders.
  • Supplier bill preparation. Bills read, matched to orders and entered as drafts.
  • Customer and supplier replies. Status, delivery date and document requests answered from ERP data.
  • Exception review. Daily lists of late orders, stock shortfalls or overdue invoices, summarized with suggested actions.

If you run Odoo, our guide to AI agents for Odoo goes through these app by app.

Where agents should not be used

  • Period close and consolidation. These need deterministic, repeatable procedures.
  • Payments and payroll runs. Money movement requires human authorization and your bank's controls.
  • Tax and compliance determinations. An agent can gather facts. A qualified person decides.
  • Configuration. No agent should alter users, roles, workflows or the chart of accounts.
  • Anything a simple rule already handles. If native workflow automation does the job, adding a model adds cost and risk for no gain.

A readiness checklist

  1. API access is confirmed for our edition and plan, including write access where needed.
  2. We can create a dedicated integration user with a custom, limited role.
  3. We have a sandbox or test copy with realistic data.
  4. Customer, supplier and product records are reasonably free of duplicates.
  5. The target process is written down, including who approves what.
  6. Someone owns the agent's review queue and its ongoing improvement.

Rolling out in stages

Stage one is read-only. The agent answers questions and produces summaries. This proves the connection, the permissions and the data quality at low risk.

Stage two is drafts with approval. The agent prepares orders, bills or replies, and a person confirms each one.

Stage three is selective autonomy. For categories that have been accepted unchanged for a sustained period, and where a mistake is cheap to reverse, let the agent complete the step and report it. Everything else stays in stage two.

Related guides

Frequently asked questions

Which ERP is best for AI agents?

There is no single best one. Any ERP with a usable API, flexible permissions and a test environment can support an agent. The quality of your data and the clarity of your process affect the outcome more than the brand of ERP.

Do we need to upgrade or migrate our ERP first?

Usually not. If your current version exposes an API and you can create a restricted user, you can start. An upgrade matters only if the version you run lacks the access the agent needs.

Should we use the ERP vendor's built-in AI or a custom agent?

Try the built-in option first for common tasks, since it needs the least work. Choose a custom agent when the process is specific to you, spans several systems, or needs approval and logging rules the built-in feature does not provide.

Sources and further reading

Primary documentation and research behind this guide.

The ERP you already run is very likely ready for a first, read-only agent. Growmerz builds AI agents for Odoo and ERP along the lines described here. To find the right starting point for your system, book a free AI audit.

Keep reading