AI can help your ERP tam move faster, but it should not be the final decision-maker when a choice affects money, production, customers, or compliance. For Canadian manufacturers and operationally complex businesses, labour shortages, rising input costs, and customer expectations can turn small approval delays into larger operational problems.
At Kodershop, we see the strongest results when AI gathers information, spots patterns, and prepares a clear recommendation, while accountable people retain authority over meaningful changes. Human-in-the-loop AI keeps the speed benefits of automation without giving up oversight.
Put Human Judgment at the Centre
Odoo AI agents can reduce the time spent searching through records, summarizing exceptions, and routing work to the right person. That does not mean every recommendation should become an automatic ERP update.
Some actions may be safe to run automatically when they are low risk and tightly defined. Creating a follow-up activity, assigning an exception for review, or preparing a draft communication are common examples. Other actions need human approval because they can affect margins, supply commitments, financial records, or customer relationships.
Set a clear line between:
- AI-generated recommendations that inform a person
- Low-risk actions permitted to run automatically
- Changes that require approval before execution
- Actions that should remain manual because the risk is too high
Odoo 20 consolidates access rights and record-level filtering into a unified access rights model. Instead of separate record rules, domains are defined directly on access rights to determine which records users can access. External AI tools can add language interpretation, reasoning, summaries, and task coordination, but they do not replace Odoo’s own security controls and permissions.
Know Where Odoo Ends and External AI Begins
Odoo provides the ERP foundation. It holds your records, user permissions, workflow controls, and automation logic. Its AI MCP server can expose selected Odoo capabilities to an external AI client through the Model Context Protocol, or MCP.
That connection should be treated as controlled access, not an open door into your ERP. An external agent may retrieve approved context and invoke only the server actions Odoo makes available to it. The AI client supplies the conversational and reasoning layer. Odoo remains the system that decides what data and operations are permitted.
Server actions are especially important in this design. They define specific operations inside Odoo, such as preparing a quotation for review, assigning an exception record, creating a follow-up activity, or updating an approved workflow status. The “Available in MCP” setting determines which server actions an MCP client can access.
We treat that setting as a governance control, not a convenience feature. If an action is not appropriate for an AI client to invoke, it should not be exposed through MCP.For a closer look at the integration layer, read our guide to the Odoo 20 MCP server and connecting external AI to your ERP.
Before planning a production workflow, verify the Odoo version and edition, hosting arrangement, subscription plan, and availability of the required AI features. Check the official Odoo release notes and documentation for version-specific requirements, including IAP credit requirements for AI features. Confirm these details before designing the workflow, as they can affect which capabilities are available and how the integration should be implemented.
Build a Controlled Approval Path
A useful AI workflow begins with a real operational trigger. It could be a supplier delivery arriving late, a production order exceeding planned material cost, or a sales quote that needs an exceptional discount.
Once the trigger appears, an external AI agent connected through Odoo’s MCP server can retrieve the relevant ERP context using the tools exposed to it. Depending on the case, that may include supplier performance, inventory availability, customer order history, margin information, open production commitments, and related activities.
The agent’s job is to prepare a proposal, not immediately change the record. A well-formed proposal should state what it recommends, why it recommends it, and which records could be affected. For example, it may suggest an alternate supplier, draft a revised purchase order, prepare a production reschedule, or write a customer update for review.
A controlled workflow can follow this pattern:
- Identify the exception or trigger in Odoo.
- Retrieve the relevant ERP data through read-only tools.
- Prepare a recommendation or draft action, supported by evidence from the available records.
- Route the proposed change to the designated approver.
- Let an authorized user execute the approved action in Odoo, then verify the result.
Odoo 20 supports AI-assisted workflows through its MCP server and server actions. The MCP server can expose selected tools to an external AI client, while the Read-only Tool setting indicates that a tool does not modify existing data and can be invoked without explicit user approval. This setting is a signal to the client, not a substitute for enforcing permissions and controlling what the tool can do. See the official Odoo MCP server documentation.
For workflows that require human approval, a safer design is to have the AI prepare a draft or recommendation rather than execute a consequential change directly. An authorized employee can then review the proposed changes and trigger the relevant action through an approval-controlled button in Odoo. Odoo Studio supports approval rules for button-triggered actions and lets teams designate the users or groups responsible for approval. Odoo Studio is an Enterprise feature, so confirm that the deployment includes the required edition and functionality. See the Odoo Studio approval rules documentation.
Do not assume that a Studio approval rule automatically governs a server action invoked through MCP. The approval mechanism must be verified for the specific implementation, including the way the action is exposed, invoked, and authorized. Where approval enforcement has not been confirmed, keep the final action in the hands of an authorized user rather than allowing the AI client to execute it directly.
For example, a purchasing manager can review a proposed supplier change and confirm it through an approval-controlled button. A production planner can review a suggested rescheduling before applying it, while an account executive can check a customer-facing message before sending it. In each case, the AI assists with analysis and preparation, but the appropriate employee remains responsible for approving and executing the consequential action.
After the action is completed, verify the resulting change in the relevant Odoo record. This final check helps confirm that the intended update occurred and that the ERP reflects the approved decision.
Keep Permissions and MCP Exposure Narrow
Access controls define which models, records, and operations a user or connected service account can access. AI workflows should operate with deliberately limited permissions rather than broad administrator access.
This matters when your ERP holds employee information, customer details, supplier records, financial data, and production plans. An AI integration should only see the information required for its assigned task.
The Read-only Tool checkbox in Odoo’s MCP configuration tells compatible AI clients that a tool should be treated as read-only and can be called without explicit user approval. It is a signal to the client, not a security boundary that independently prevents data changes. For example, read-only tools can retrieve inventory levels, order status, production capacity, or invoice details. Exposed server actions, by contrast, can enable specific operations or changes, depending on how they are configured and what the connected Odoo user is authorized to do. Review the permissions and behaviour of each exposed tool or action before making it available to an AI client.
Both controls matter. Minimal read access reduces unnecessary data exposure. Specific, approval-aware server actions reduce the risk of unexpected changes. Teams should also define approval thresholds, audit needs, exception handling, and a clear process for disabling the integration if its behaviour becomes unreliable.
Measure Results and Turn Them Into Policy
The best measure of Odoo AI agents is whether your team spends less time waiting, searching, reworking, and handing work from one person to another.
Before introducing an AI-assisted workflow, establish a baseline for manual processing time, approval wait time, missed handoffs, rework volume, and the time required to find the information behind a decision. Then compare those measures after the workflow is in use.
For manufacturers and distributors, useful measures may include:
- Purchase order approval cycle time
- Production exception response time
- Quote turnaround time
- Inventory shortage resolution time
- Delay between reporting a quality issue and assigning it
Validation should remain part of the operating policy. Teams need to confirm that an approved server action created the intended record change and did not affect unauthorized records. Approval history should remain available for review, and regular sampling can compare AI recommendations with human decisions.
Practical goals start with a narrow, high-friction process where delays are visible and approval ownership is clear. When AI provides timely context, people make the consequential choices, and Odoo executes only approved actions, your ERP can become faster without losing accountability.
Turn AI Insights Into Accountable ERP Results
Kodershop helps organizations assess where Odoo AI agents can reduce workflow friction while preserving the approval controls your teams need. We focus on practical use cases that support faster operations, clearer decisions, and reliable execution across your ERP. To discuss the right opportunity for your business, contact us for a focused conversation with our team.