Most AI add-ons bolted onto an ERP fail in one of two ways.
They invent numbers. Ask a generic AI layer "which invoices are overdue" and there's often no guarantee the answer came from your actual ledger rather than the model's best guess at what a plausible answer looks like. In accounting, inventory, or pricing, a confident wrong number is worse than no answer at all.
They act without asking. The moment an "agent" gets a create/write tool, the natural next step for most platforms is to let it use that tool autonomously because that's the demo that looks impressive. It's also the demo that goes wrong the first time the model misreads a request and quietly drafts three RFQs instead of one.
OdooNerve AI is built against both failure modes structurally not by prompt instruction, but in the code itself.
- Every number an agent states is computed by Odoo. A pipeline's win probability, an invoice's amount overdue, a product's on-hand quantity, a vendor's price read or deterministically calculated in real ORM code, never estimated by a language model. The AI's job is to search, explain, and recommend against numbers it never touched.
- Every action that would change a record waits for a named human. There is no code path, in any module shipped today, that lets a write-capable tool run itself. Proposing an action creates a pending approval and nothing else. Only a person with the right role, clicking Approve, causes anything to actually happen.
And unlike platforms that want to sell you their own inference, OdooNerve AI never operates AI infrastructure itself. You bring your own OpenAI, Anthropic, Gemini, or xAI credential your data, your bill, your relationship with that provider. OdooNerve AI is the governance layer that sits between that credential and your database, not a reseller of tokens.
How it actually works
Every department-level feature sits on four foundation layers, each depending only on the one before it:
Core → security groups, usage & audit log, budgets, kill switch
Provider → your own AI credential, encrypted, routed through one Gateway
Agent → the Tool Registry, Agent Engine, Ask AI
Automation → the Human Approval Engine, the Cross-Module Workflow Engine
An agent never gets free-form database access. It gets a fixed list of tools narrow, code-defined capabilities, each either read-only or write-capable-and-approval-gated. There's no way, through the UI or the API, to create a tool that runs an arbitrary query.
When a write is genuinely useful drafting a quotation, proposing a follow-up activity, creating an RFQ here's what actually happens:
- The agent's answer includes a proposed action: which tool, which parameters, why.
- That becomes a pending approval record nothing has happened yet.
- A named Approver reviews the full preview: tool, every parameter, target record, stated reason.
- Only that human's click runs the tool. Every outcome executed or failed is logged. Tools rated critical-risk can't execute at all, regardless of approval.
Multi-step processes work the same way. A workflow is a fixed, ordered sequence an administrator authors Query, Rule, AI, Tool, Approval, Notification, Delay, Condition never a sequence a model assembles on the fly. There's no jump/loop construct, so "a workflow cannot run forever" is a structural guarantee, not a runtime check someone has to remember to add.
What this looks like in a real business
Sales & CRM. "What opportunities need attention today?" surfaces stale, open deals with no scheduled activity explained and prioritized, never re-staged automatically. "Draft a quotation for Acme Corp: 20 units of the Pro Widget" creates a real draft with Odoo-computed pricing, held for approval, never auto-confirmed.
Accounts Receivable. Once a day, a rule finds the single overdue invoice with the largest exposure (days overdue × amount outstanding) plain arithmetic over figures accounting already computed. The AI explains the risk in plain language, proposes a follow-up activity, and a named Approver decides. Collections effort lands on the highest-exposure account every morning, with a reason attached, instead of a finance team re-sorting an aging report by hand.
Inventory. "What products are at risk of stocking out?" runs a deterministic forecast moving averages and threshold checks over real stock movement and only ever phrases findings as a recommendation. No write tool exists here at all; turning a recommendation into an actual purchase is a separate, approval-gated step.
Purchase. "Compare vendors for the Pro Widget" reads the real vendor pricelist and lead times. No estimation.
Cross-department: Order Fulfilment Risk. A salesperson about to send a large quotation clicks Analyze Fulfilment Risk. In one run: the customer's receivable standing (Accounting), live stock position (Inventory), and known vendor pricing (Purchase) are pulled together. A rule computes the shortage. The AI recommends Accept, Review, or Procurement Required. If stock is genuinely short, it proposes a draft RFQ to the cheapest known vendor priced by Odoo, not the model and stops at draft, waiting for a human. One click quietly checks credit, stock, and vendor pricing before a big order becomes a fulfilment problem discovered after the fact.
Knowledge. An internal knowledge base answered strictly from retrieved excerpts of your own documents no vector database or OCR stack to stand up first.
The business effect
| Department | Time saved on | What stays 100% human |
| Sales / CRM | Scanning a pipeline for what's gone stale; drafting routine quotations | Which deal to prioritize, confirming any quotation |
| Accounts Receivable | Re-sorting an aging report to decide who to chase first | Every collection contact, every payment decision |
| Inventory | Spotting stockout/dead-stock risk before it costs a sale | The actual reorder decision |
| Purchase | Comparing vendors, drafting an RFQ | Confirming any RFQ into a real Purchase Order |
| Cross-department | Manually checking credit, stock, and vendor pricing before a large order | Whether to actually place that order |
The pattern repeats in every department: OdooNerve AI removes the manual triage and drafting work, and leaves every decision that commits money or relationship capital with a person on the record, with a reason attached.
Built for how real Odoo deployments actually look
Eleven addons, one governance model underneath all of them, source-verified against real Odoo 17.0, 18.0, and 19.0 code not assumed to work on a version just because it worked on another. No Enterprise-only dependency anywhere in the core platform; what you can run depends on your Odoo edition, not on OdooNerve AI.
Getting started
Install odoonerve_core, then odoonerve_provider. Set your AI provider credential, test the connection, and install whichever business packs match the Odoo apps you already run CRM, Sales, Accounting, Purchase, Inventory. Every pack ships a ready-to-use default agent. No demo data, no seeded examples everything an agent answers comes from your own real records, from day one.
Modules link : https://apps.odoo.com/apps/modules/browse?search=odooNerve
OdooNerve AI is developed by Bithost, a technology solutions brand of Zhost Consulting Private Limited. Not developed, endorsed, certified, or maintained by Odoo S.A. "Odoo" is a trademark of Odoo S.A., used here only to describe compatibility. Contact bithost.in or sales@bithost.in.
