Odoo 19 is the first version with native AI at the core of the ERP. Agents, tool calling, RAG, AI fields — it's all there. And the architecture behind it is solid: provider-agnostic by design, agents are configurable, and tools connect directly to your business data.
But if you work with AI daily and are used to how other tools handle it, you'll notice the implementation leaves opportunities on the table. The foundation is there — but it can be better. And we don't want to wait for v20.
What I Learned from Airtable
At my previous company, we used Airtable extensively. One of the features that stuck with me most: AI Fields. A column that thinks. You provide a prompt, reference other columns as input, and the field automatically generates a value for each record.
What Airtable did so well:
- Per-field model selection — a simple task runs on a cheap model, a complex analysis on a more powerful one
- Automation — you decide whether the field runs automatically on changes, or whether you regenerate manually
- Respect for human input — if you manually adjust an AI-generated value, the AI leaves it alone
It's an elegant balance between automation and control.
But Airtable isn't open source. It's not your ERP. And your data sits in a silo, separate from your CRM, your invoicing, your inventory. The thought that stayed with me: what if you brought the best of Airtable to Odoo?
Improvement 1: More AI Providers
Odoo 19 supports two AI providers: OpenAI and Google Gemini. That's a start, but it's not enough.
Vendor lock-in at the AI level is a real risk. Prices change, models improve unevenly, and some tasks call for a specific model. You want to be able to choose.
We've added Anthropic Claude as a full third provider. Not as a workaround or standalone integration, but as a native extension of Odoo's existing AI infrastructure. The same UI, the same agents, the same tools — just an extra option in the dropdown.
Three models, each with their own sweet spot:
- Claude Haiku — fast and affordable, ideal for simple tasks like categorisation or translation
- Claude Sonnet — the best balance between speed and intelligence for daily use
- Claude Opus — the heavy artillery, for complex analysis and nuanced tasks
The choice for Anthropic is deliberate. Claude scores consistently high on benchmarks, excels at following instructions, and Anthropic visibly invests in safety and reliability. But the principle is broader: as an organisation, you shouldn't depend on a single AI vendor. That's why Mistral — European and open-weights — is next on our roadmap.
Improvement 2: Web Search per Agent
Odoo's AI agents work with internal data only by default. That makes sense for many tasks — you want an agent to search your CRM data, not the entire internet. But sometimes you need current, external information.
We've added a web search toggle per agent. One switch in the agent configuration, and the agent can search the web. It runs on Anthropic's native web search — no external service to configure, no extra API keys.
What makes this powerful is the combination. An agent has access to your internal Odoo data and can simultaneously consult the web. A few examples of what that delivers:
- "Write a competitive analysis for this lead based on their company profile in our CRM and their latest press releases"
- "Check whether this product's specifications comply with current EU regulations for this product category"
- "Summarise what has appeared in the news about this customer over the past month"
Internal context plus current external information — in a single agent call.
Improvement 3: AI Fields, Airtable-Style
The standard AI Fields in Odoo work, but they're limited. The model is hardcoded to OpenAI. There's no choice in which model you use. And there's no automation: you click, you wait, you get a value.
We've fundamentally improved AI Fields, inspired by how Airtable does it. This is the biggest change.
Each Field Gets Its Own AI Agent
Each AI field can be linked to its own AI agent — and thereby to its own model, temperature, and settings. This sounds like a technical detail, but it's a game-changer for cost and quality.
A field that categorises products? Run that on Haiku — fast, cheap, more than capable enough. A field that writes a nuanced market analysis? Put Opus or Sonnet on that. You optimise per field, not with a one-size-fits-all setting.
Auto Fill and Auto Update
Two toggles that determine how the field behaves:
Auto Fill runs as a background process. It finds records where the AI field is still empty and fills them in automatically. No manual clicking, no starting batches — the agent runs as a cycle in the background.
Auto Update goes a step further. When a source field changes — a product description is updated, a customer profile is modified — the AI value is automatically regenerated. Your data stays consistent without anyone having to think about it.
The agent runs as a continuous cycle in the background. That's fundamentally different from how it works out of the box, where you have to click manually per record.
Smart Indicators and Protection
Don't have Auto Update enabled, but something changed in the source data? A subtle "Inputs changed" indicator appears on the field. You know the value is outdated, but you choose when to regenerate. No surprises.
And then the feature I'm most proud of: human-edit protection. When an employee manually adjusts an AI-generated value, that field becomes protected. The AI won't overwrite it again — not during Auto Fill, not during Auto Update, not during a bulk regeneration.
This sounds like a small detail, but it's crucial for adoption. Users need to trust that their deliberate adjustments won't be silently rolled back by a background process. Without that guarantee, nobody will trust AI fields.
Smooth User Experience
Finally: the AI computation runs in the background. The UI doesn't freeze. You see a spinner on the field while it works, but the rest of the screen remains fully usable. That sounds obvious, but it's a noticeable difference from the standard implementation.
What Does This Make Possible?
The real power lies in the combination: internal Odoo data, AI processing with the right model, and optionally live web information. A few scenarios:
Product enrichment — Automatically generate multilingual product descriptions based on technical specifications, category, and market trends. One source record, multiple AI fields, each with the right language and tone.
Lead intelligence — Combine what you know about a prospect in your CRM with recent news about their company. Automatically write a personalised outreach message that shows you've done your homework.
Compliance monitoring — Link product data to current regulations for your target market. Let an AI field continuously check whether your products still comply — and flag when they no longer do.
Customer insights — Summarise the complete interaction history and enrich with public information. One field that brings an account manager up to speed on a customer relationship in 30 seconds.
The common thread: your ERP data isn't just stored, but actively enriched and analysed. Automatically, in the background, with the right model for the right task.
What's Next
We're rolling this out internally at Pantalytics first. Eating our own dog food. We use it daily in our own Odoo environment — to enrich leads, generate product information, and improve customer communication. That's the best way to stay sharp: when you use your own tool every day, you quickly notice what can be improved.
Once it's proven in our own operations, we'll make it available to customers. The module is open source and available on GitHub — transparent and auditable.
Our sincere hope is that Odoo itself moves in this direction with v20. The architecture is ready for it. Multi-provider support, better AI Fields automation, web search as a configurable option — the building blocks are there. It would be good for the entire ecosystem.
But until then, we're not waiting. And neither do our customers have to.
Pantalytics specialises in AI and Odoo. Want to see what AI Fields can do for your Odoo environment? Get in touch for a demo.