Odoo 20 AI Credits Explained: The Real Cost of IAP Credits

Odoo 20 makes AI a normal part of the product — and a separate line in your budget. The release notes say it in one short line: all AI features now need credits, which can be bought via IAP. Odoo's pricing page reinforces the point by listing In-App Purchase credits among the things not included in its plans. So the subscription buys you the software and, on the relevant plan, access to agentic AI. Running the AI is pay-as-you-go.

That distinction surprises teams because Odoo pricing is usually easy to explain: choose a plan, count your users, multiply. AI credits do not work that way. They are consumed by activity, not by headcount. One quiet agent may cost very little. One badly scoped scheduled agent can run all day, whether anyone reads its output or not. This article explains what IAP credits are, where they are consumed, what Odoo does and does not publish about price, and how to budget without inventing numbers.

What an IAP credit is, in plain terms

IAP stands for In-App Purchase, Odoo's mechanism for buying consumable services from inside the database. Odoo has used IAP for services such as SMS, contact autocomplete, lead generation, and scanning vendor bills and expenses; the pricing page names those examples when it says IAP credits are not included in plans. In Odoo 20, AI joins that model. When a feature calls Odoo's AI service, credits are spent.

Treat a credit as a unit of prepaid usage, not as a user licence. Your plan determines which features you are entitled to use — the pricing page lists Agentic AI with the Custom plan — while your credit balance determines how much AI work you can actually run.

The price question Odoo does not answer publicly

Here is the part to state honestly: Odoo does not publish a simple public price list for Odoo 20 AI IAP credits in the release notes, and the pricing page checked for this article lists IAP credits as excluded without stating a per-credit AI price. Third-party blogs quote figures, but those figures are not a substitute for the price Odoo presents in your own database, for your region, at the moment you buy.

That means any spreadsheet that gives you an exact monthly AI cost without first measuring your usage is guessing twice: once about the price you will pay, and once about how many credits your tasks will consume. Do not build an approval on either guess. Run a controlled pilot, record the credit movement against real work, and forecast from your own numbers.

Regional differences make this caution practical. Odoo pricing varies by country, currencies differ, and offers can change. A Canadian business should not budget from a figure copied from an overseas blog, convert it casually, and call the result a plan. Check the purchase flow in your database, note the currency and options shown to you, and date that observation in your budget file.

Where credits are consumed

The safest rule is the release notes' rule: if it is an AI feature, assume it needs credits until you prove otherwise in a test database. The main consumption points follow directly from the verified Odoo 20 AI capabilities.

Activity	Why it consumes credits	What drives the volume
Agent conversations	Asking an agent to answer, summarise, or reason over your data calls the AI service.	Number of conversations, length of context, and follow-up messages.
Agents creating records	Creating records, including from an uploaded PDF or instruction file, is AI work performed on your behalf.	Files processed, records created, and retries after unclear input.
Agents updating records	Reading a record, deciding the change, and applying an update is a billable agent task pattern.	Records touched and how often triggers fire.
Agents in automated actions	An automated action that summons an agent spends credits each time the rule fires the agent.	Trigger breadth, watched-field changes, and duplicate or repeat events.
Agents in scheduled actions	A scheduled action spends credits on every run, including runs that find little to do.	Schedule frequency, records in scope, and runtime per run.
Questions about files	Asking questions about a file while previewing it uses AI to interpret the document.	File size and complexity, and the number of questions asked.
Image generation	Generating images and enriched buttons for websites and mailings is an AI feature.	Images requested, revisions, and discarded drafts.
AI accounting insights	The accounting assistant answers questions and provides insights based on your accounting data.	Questions asked and how often teams use it as a reporting shortcut.
AI website assistance	Creating pages, editing snippets, adding images, and building forms and menus by prompt is AI work.	Pages built, edits requested, and iteration count.
Voice-dictated requests	Dictating a request still sends work to an agent; voice changes the input method, not the billing model.	Dictated tasks and any follow-up agent actions they trigger.

Two amplifiers deserve attention. First, interactivity: agents can ask questions and request permission by buttons, and a person can ask for changes after reviewing an automation chat. Helpful iteration is still iteration; reprocessing sources and rerunning a task consumes more than a single clean pass. Second, limits: when an agent reaches its maximum tool call limit, it prompts for confirmation before proceeding. Approving that continuation is a cost decision.

Budget per agent and per workflow, not per user

Per-user budgeting fails for Odoo 20 AI because usage follows automation design. Ten users who occasionally ask a question may spend less than one scheduled agent reviewing a large record set every hour. Budget around the unit that actually spends: the agent, the workflow, and the trigger.

Use a simple worksheet for each proposed workflow before you enable it in production:

Budget field	How to fill it in
Workflow name and owner	One business name, and one person accountable for reviewing both output and spend.
Trigger	Automated action, scheduled action, or manual use. Record the exact event or schedule.
Expected runs per week	Count from the trigger: records created or updated, or runs on the schedule. Measure the trigger in your data before guessing.
Records per run	Average records the agent will read or touch in one run, plus the largest realistic run.
Pilot credit use	Credits consumed during a controlled pilot, recorded from your database over a stated period.
Cost per run	Pilot credit use divided by pilot runs. Convert to money only using the price shown in your own purchase flow.
Monthly forecast	Expected runs per month multiplied by cost per run, with a separate line for growth and rework.
Stop condition	The balance, error pattern, or review finding that pauses the workflow pending a human decision.

Run a pilot that produces real numbers

Pick one workflow with a countable output. Documents processed, tickets summarised, or records prepared are countable. General helpfulness is not.
Note the starting credit position. Record the balance and any purchase details shown in your database before the pilot begins.
Run it on real but bounded work. Use a narrow domain, one team, or a short schedule. Keep the tool call limit in place.
Log runs next to credits. Track how many times the agent ran, how many records it touched, and how many outputs a person accepted, edited, or rejected.
Do the arithmetic. Credits spent divided by accepted outputs is the first honest unit cost you will have. Label money conversions with the date and price source inside your own purchase flow.
Decide with the owner. Keep, narrow, reschedule, or stop the workflow based on accepted outputs per credit spent — not on how impressive the demo felt.

Controls that keep AI spend boring

Start scheduled, not event-driven. A known cadence makes spend predictable while you learn what a run actually costs.
Narrow the domain before you widen the skill. Most runaway spend is a trigger problem: a rule firing on records nobody needed an agent to touch.
Assign an owner per agent. Someone must read the automation chats, watch for repeated errors, and notice when a workflow keeps reprocessing the same sources.
Review inside the retention window. Agent conversations are stored for up to 30 days. Review before the evidence ages out.
Separate plan decisions from credit decisions. Moving plans to access agentic AI and buying credits to run it are two approvals.
Re-forecast after every change. A new skill, a broader domain, a shorter schedule, or a second team using the same agent all change the burn rate.

What to tell management

Keep the message simple. Odoo 20 AI is not included as unlimited usage. No plan includes the IAP credits the AI features need, the exact public price is not published in the sources Odoo provides for the release, and consumption scales with how often agents run and how much work each run does. The responsible commitment is therefore not a guessed monthly figure. It is a funded pilot, a named owner, a measured cost per accepted output, and a stop condition agreed before the automation was switched on.

That framing protects the project. If the pilot shows an agent preparing work accurately for a defensible cost, expanding it is an easy conversation. If it shows credits draining into rework, you have numbers and a bounded experiment to point to — instead of a surprise balance.