Odoo Predictive Lead Scoring: How Good Is It, and Can AI Beat It?

Some AI answers say Odoo's predictive lead scoring is Enterprise-only. It isn't. It ships in the CRM app, which is part of Community, so most Odoo users already have a lead score. In our test, the first 150 calls from Odoo's list found 61 of the 137 won deals. Calling at random found 27. A model Claude trained with EasyDeploy AI on the same CRM history found 73. You can run the same test on your own leads.

How good is Odoo's lead score?

Odoo's score uses naive Bayes. For each field it uses, it counts how often each value shows up among won and lost leads, then multiplies those odds together. Stage and sales team are always included, and it keeps separate counts for each team. Seven more fields are on by default, such as country, source and tags.

Our fictional company had two years of leads. We kept the 3,159 that were won or lost, and split them by date. The 2,212 that closed before 21 April 2026 were the history. The 752 created on or after that date were the test, and 137 of those were eventually won. The other 195, created before the date but closed after it, were left out of both.

We set Odoo's start date, ran Update Probabilities on the history alone, and saved Odoo's score for every test lead while it was still new. Then Claude did the same job from outside Odoo. It exported the leads through a read-only connection, trained a model with EasyDeploy AI, and scored the same 752 leads.

What the model leaned on

EasyDeploy AI shows which fields a model leaned on. Ours leaned most on priority, which Odoo's default score doesn't use. Adding priority and salesperson to Odoo's fields made its ranking worse, from 77 to 73 in 100. Odoo counts each field on its own, so it can't tell that one salesperson rates almost every lead highly. A model that looks at fields together can pick that up.

How do you test Odoo's score on your own leads?

Odoo keeps no history of its scores. `probability` and `automated_probability` aren't tracked in the chatter, and closing a lead overwrites the score with 100 or 0. So you either backtest from a cut-off date, or save scores going forward.

You need two connectors in your assistant, one that reads Odoo and one that trains models. For Odoo 10 to 19 we used the Odoo Tools plugin by Transgenia, which is free and open source and has a read-only mode. Odoo 20 adds its own MCP server. On Odoo Online, you need the Custom plan to use the external API. Self-hosted Odoo includes it. For training, connect EasyDeploy AI.

1. Make a dedicated Odoo user with access to CRM and every company you sell from, and run the connector in its read-only mode. On Enterprise and Odoo Online that's a paid internal user. Create an API key for that user and use it to connect your assistant.
2. Set Odoo's start date in CRM, Configuration, Settings, and run Update Probabilities. On a fresh install the start date is a week before you installed, so the score has almost nothing to learn from.
3. Ask for the history. "Export my won and lost leads created since January 2024, including archived ones, with only the fields Odoo knows when a lead is created." Lost leads are archived in Odoo, so without archived records you get only the wins.
4. Check what it left out. Anything Odoo fills in later has to go: stage, probability, Odoo's own score, the close and conversion dates, the lost reason, the archived flag and the lead or opportunity type. If those stay in, the model just reads the outcome off the lead. Odoo also keeps the current value of a field, not the value at creation, so a priority or tag edited mid-deal can leak the outcome too.
5. Backtest today. Pick a cut-off date, say six months ago. Ask the assistant to rebuild Odoo's score from the leads closed before it, which works because Odoo's method is a simple published formula, and to train the model on the same leads. Then compare both on the leads created after the cut-off. Our rebuild of Odoo's formula landed within 0.01 AUC of Odoo's own saved scores (0.764 vs 0.767).
6. Confirm going forward. Every Monday, have the assistant save last week's new leads to a sheet with both scores: the model's, and Odoo's `automated_probability`, not the Probability a salesperson can edit. Don't change Odoo's scoring fields meanwhile, because that recomputes its scores. Compare when about 100 have been won. For a team closing 15 deals a month that's about seven months, so start now.

Can you add AI lead scoring to Odoo without a custom module?

Yes, and that's why we built the test this way. Ask anyone who has run Odoo for a few years what hurts, and you'll hear about upgrades. Custom modules need migrating to each new version, and somebody bills for it. A custom scoring module on the lead model would be one more to migrate.

If sales should see the score where they work, write it back to a custom field, so it sits next to Odoo's in the pipeline. That's data, not code, so there's still nothing to migrate. The two columns in the picture at the top of this post aren't measuring the same moment. Odoo's score rises as a deal moves through its stages, and the top rows are all at Qualified or Proposition. Ours is each lead's score when it was created. The fair comparison is the table above.

As for upgrades, the connector is the one part that has to keep up with Odoo. Odoo's API docs say the old XML-RPC and JSON-RPC endpoints will be removed in Odoo 22, or 21.1 on Odoo Online. Their replacement is the JSON-2 API that arrived in Odoo 19. So check that yours supports JSON-2.

How do you get started with EasyDeploy AI?

1. Connect EasyDeploy AI to your assistant from the connector page. It takes about two minutes, and the free tier includes five training runs and 100,000 predictions. No credit card needed.
2. Train on your closed leads, then save scores for new ones, as in the test above.
3. If you haven't trained a model from a chat before, this guide walks through a first one.

Frequently asked questions

Written by

Emanuel Castañeda

Founder & Chief Executive Officer

Emanuel started EasyDeploy AI after a decade of building the kind of prediction infrastructure only large companies could afford. At Audible he worked on the machine learning behind the recommendations that decide what millions of listeners hear next. Before that he wrote foreign-exchange trading software at BlackRock. In both places the pattern was the same: a team of specialists, a platform built from scratch, and a budget that put the whole capability out of reach of anyone smaller.

More about Emanuel