Contents8
- 01What is an AI credit?
- 02How much does one AI action cost, vendor by vendor?
- 03What the included allowance really covers
- 04What would the same work cost paid straight to the model?
- 05When to keep the vendor’s AI, and when to take it out of the tool?
- 06How do you price a tool’s AI before switching it on?
- 07Where to start?
- 08Frequently asked questions
Open the pricing page of any major business software product: AI is no longer simply “included”. It is metered in credits, a currency each vendor defines for itself, with a monthly allowance bundled into the subscription, a rate card per action and a price for every credit beyond the allowance. Odoo followed suit with version 20, released in September 2026: “All AI features now need credits”.
The principle is defensible: the vendor pays for the models it calls on your behalf. But each one has its own currency, rate card and rollover rules, and without a conversion nobody knows what an answer, an enriched record or a resolved conversation actually costs. We did that conversion using the price lists published as of 14 October 2026.
In short
A credit is not worth the same from one vendor to the next: $0.010 at HubSpot, $0.005 at Salesforce, $0.008 to $0.01 at Microsoft, €1 at Odoo.
Brought down to a single action, in-app AI costs from $0.02 (a generative answer at Microsoft) to $2 (an agent conversation at Salesforce).
The included allowance resets every month with no rollover, and spending caps are not always on by default.
For a few hundred actions a month, built-in AI remains the simplest option. For thousands of repetitive tasks, calling a model from your own automations costs a fraction of the price in credits.
What is an AI credit?
Definition
AI credit
A unit of account specific to one software product. Each AI action consumes a number of credits set by the vendor; the subscription includes a monthly allowance, and credits beyond it are billed at a price the vendor also sets.
Three numbers tell you what you will pay: the price of a credit, the number of credits per action, and what happens once the allowance is used up. The vendor can move the last two. HubSpot says so plainly on its credits page: beta features “may consume credits in the future, and credit rates may change as features evolve”.
Credits pay for a real service: the model call, but also reading your records, integration into the screen your team already uses, and hosting. For occasional use, it is the simplest way to get AI without building anything. The question is the volume at which that convenience becomes expensive.
How much does one AI action cost, vendor by vendor?
What one AI action costs at published rates, in dollars, as of 14 October 2026
Per action or per answer
Per conversation or per outcome
These amounts are our own calculation, based on what each vendor publishes: the number of credits per action, multiplied by the price of a credit. At HubSpot, an AI action in a workflow uses 10 credits and a conversation resolved by the customer agent uses 50, at $0.010 per credit beyond the allowance. At Salesforce, an action costs 20 Flex Credits, sold at $500 per 100,000. At Microsoft, a generative answer costs 2 credits and an agent action 5, at $0.01 per credit pay as you go; a prepaid pack of 25,000 credits at $200 a month, billed annually, brings the credit down to $0.008. Intercom and Zendesk, for their part, bill the outcome directly.
Two vendors do not lend themselves to the calculation. At Odoo, a credit costs €1, sold in packs of 10 to 1,000 with 20 bonus credits, but no rate card per action is published. Consumption, says the Odoo AI service page, “varies based on the specific AI feature, task complexity, and the size of your request”. These in-app purchase credits are not part of the subscription, and agentic AI only exists in the Custom plan. So there is no way to forecast a monthly bill without having paid one first.
On n8n Cloud, credits only pay for the assistant that helps you build workflows: 1,600 a month on Starter, 4,000 or 9,600 on Pro, with no rollover. Running a workflow consumes none, and the model calls your workflows make go through your own key or through separate prepaid credits. Self-hosted, there are no credits at all: the assistant runs on your key.
What the included allowance really covers
Take HubSpot’s Sales Hub Professional. It includes 3,000 credits a month. At the published rates, that means 300 AI actions in workflows, or 60 conversations resolved by the customer agent, or 30 prospects for whom the prospecting agent suggests an approach, at 100 credits each. A smart property, which fills a record field using AI, consumes credits every time it runs, “even when a value isn’t filled for the property”.
Three rules in the catalogue change the maths:
- credits reset every month, and whatever is left is lost;
- included allowances do not add up across products: two subscriptions do not double the allowance;
- once the allowance is used up and no pack has been bought, pay-per-use features stop until the reset.
Finally, an additional credit pack can only be cancelled or reduced at contract renewal. The extra you needed for one busy month can therefore end up being paid for all year.
Watch out
Spending caps are not always set by default. At HubSpot, “Credit limits aren’t turned on by default”: only buying a pack triggers an account cap, set 50% above the monthly total, and alerts go out at 75, 85 and 90% consumption. At Microsoft, agents are switched off at 125% of prepaid capacity. On n8n Cloud, auto top-up has a monthly cap, on by default.
What would the same work cost paid straight to the model?
No vendor publishes how many tokens an action consumes, so an exact comparison is impossible. An order of magnitude is not. Take a short task, classifying an email or summarising a record, of roughly 1,000 tokens read and 100 written (that is our assumption). At the prices published by model providers, the same task costs:
- $0.00015 with GPT-6 Luna, OpenAI’s budget model ($0.10 per million tokens read, $0.50 per million written);
- $0.0002 with Mistral Small 4 ($0.15 and $0.60);
- $0.0011 with Google’s Gemini 3.8 Flash ($0.75 and $3.75, until 31 December 2026);
- $0.003 with Anthropic’s Claude Sonnet 5.5 ($2 and $10).
Even with ten times more text, the task stays around $0.03 on a mid-range model. Batch processing halves these prices again at OpenAI, Google and Mistral, for tasks that can wait.
The gap with the $0.10 of an action paid in credits does not prove the vendor is overcharging: it pays for integration and hosting, and spares you any development. It shows where the line sits. At 20,000 tasks a month, for example sorting every incoming email of a 100-person company, HubSpot’s rate card works out at 200,000 credits in workflow AI actions, close to $2,000 a month beyond the allowance. The same tasks handed to GPT-6 Luna by an automation cost a few dollars of model usage, plus building the workflow and monitoring it. Our guide AI costs in business: tokens, routing and forecasting walks through this calculation.
From the field
On Palmora Property, a real estate portal in six languages, translation goes through an AI model called directly, with no vendor credits. Switching to an AI model cut the translation cost by a factor of 60, and 95% of segments are served from the cache without calling the model again.
These prices move too: Google doubles the price of Gemini 3.8 Flash on 1 January 2027. Our AI model comparison tracks these releases and changes.
When to keep the vendor’s AI, and when to take it out of the tool?
Keep the AI built into the software
A few hundred actions a month
A task that lives in the tool: summarising a record, drafting an email from the CRM
Nobody available to maintain an automation
An included allowance that is enough, with a cap set
Call a model from your own automations
Thousands of repetitive tasks: sorting, extracting, translating
The model chosen task by task, and the freedom to switch
Sensitive data kept on a model hosted on your own infrastructure
A meter, a budget and a cap per use
The second route has its own cost: you have to build the automation, monitor it and keep up to date the table that says which model handles which task. It pays off at volume, or as soon as the data must not leave your hands.
The choice of software matters too. Twenty, the open-source CRM we deploy, also uses credits in its cloud version, but there a credit is worth one dollar of usage, “1 credit = $1 of usage”, and its settings show the price of each model per million tokens. Hosted on your own server, it calls the provider of your choice, OpenAI, Anthropic, Google, Mistral or xAI, on every plan, including without a paid subscription. Our three-year CRM cost breakdown runs the same three-year calculation for licences.
How do you price a tool’s AI before switching it on?
- List the actions.
The AI features your team will actually use, and their monthly volume: enriched records, drafted emails, handled conversations.
- Convert using the rate card.
Credits per action multiplied by the price of a credit, minus the included allowance. With no published rate card, ask the vendor for a written commitment.
- Set caps before switching on.
One cap per tool, and alerts received by a named person, without waiting for the first invoice.
- Compare with a direct call.
The same task done by an automation that calls a model, build and monitoring included.
- Review every quarter.
Rate cards, model prices and volumes change: yesterday’s decision does not necessarily hold.
Key takeaway
One credit cannot be compared with another. Only the cost per action, converted at the published rate and multiplied by your volumes, lets you decide.
Where to start?
Start with the bill you already pay: pull last quarter’s credit consumption from each tool and spot the two or three actions that weigh the most. If AI already plays a part in your processes, our AI at the right cost pack (€2,900 to €5,400) runs this audit, compares models on your own content and sets a key, a budget and a cap per use. If you do not yet know where AI is worth it, our digital and AI maturity audit (€3,500) starts from your processes before talking about tools. The overall approach is in our guide Integrating AI in business: private and at the right cost.
Frequently asked questions
Do unused AI credits roll over to the next month?
At HubSpot, no: credits reset every month and whatever is left is lost. The same rule applies to the n8n Cloud assistant and to Microsoft Copilot Studio’s monthly packs. Extra credits bought at n8n remain valid for twelve months, and Odoo’s do not expire. Before buying a pack, also check its term: at HubSpot, it can only be cancelled at contract renewal.
Does an AI credit cost the same from one software product to another?
No, and that is the main trap. As of 14 October 2026, a credit is worth $0.010 at HubSpot, $0.005 at Salesforce, $0.008 to $0.01 at Microsoft depending on the plan, and €1 at Odoo. The number of credits charged for the same action varies just as much. Only converting per action, at the published rate, lets you compare two tools.
Can you use your own OpenAI or Gemini key in Odoo?
In Odoo 19, a personal key was actually mandatory on Odoo.sh and on your own server, and optional on Odoo Online. The Odoo 20 release notes state that all AI features now need credits, and the documentation page on keys is being rebuilt. Check on your own version and hosting before building a budget on it.
Is AI billed per outcome a better deal?
It ties the price to what you get: Intercom charges $0.99 per outcome of its Fin agent, Zendesk €1.50 to €2 per automated resolution on its French site. It is easy to read, but expensive at high volume, and everything depends on what the vendor counts as an outcome. Ask to see its written definition, then compare with an answer drafted by AI and approved by your team.
How do you stop a credit bill from spiralling?
Set the caps before switching on a single feature: one cap per tool, alerts, and someone who receives them. At HubSpot, limits are not on by default, and alerts go out at 75, 85 and 90% of the allowance. At Microsoft, agents stop at 125% of prepaid capacity. Then review consumption every month, action by action.
Sources
- Odoo, Odoo 20 release notes (new tab)odoo.com
- HubSpot, Understand HubSpot credits and billing (new tab)knowledge.hubspot.com
- Microsoft, Copilot Studio, credit management (new tab)learn.microsoft.com
- HubSpot, Product and Services Catalog (new tab)legal.hubspot.com
- Salesforce, Agentforce pricing (new tab)salesforce.com
- Intercom, Fin pricing (new tab)intercom.com
- Zendesk, pricing (new tab)zendesk.com
- Microsoft, Copilot Studio Licensing Guide, April 2026 (new tab)cdn-dynmedia-1.microsoft.com
- Odoo, "Odoo AI" service and its credits (new tab)iap.odoo.com
- Odoo, pricing page (new tab)odoo.com
- N8n, n8n Cloud pricing page (new tab)n8n.io
- N8n, assistant credits documentation (new tab)docs.n8n.io
- HubSpot, Create smart properties (new tab)knowledge.hubspot.com
- N8n, top up assistant credits (new tab)docs.n8n.io
- OpenAI, API pricing (new tab)developers.openai.com
- Mistral AI, API pricing (new tab)mistral.ai
- Google, Gemini API pricing (new tab)ai.google.dev
- Anthropic, API pricing (new tab)platform.claude.com
- Twenty, cloud version credits (new tab)docs.twenty.com
- Twenty, release notes 2.40.0 of 15 September 2026 (new tab)twenty.com
- Twenty, AI FAQ (new tab)docs.twenty.com
- Odoo, 19.0 documentation on AI API keys (new tab)odoo.com
- Zendesk, pricing in France (new tab)zendesk.fr
Microsoft, Copilot Studio Licensing Guide, April 2026
Read on cdn-dynmedia-1.microsoft.com(new tab)Jim · founder of Brain Plus AI
Also in business ai
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.
Understand44 min
AI comparison: models, releases and scores since 2022
The major AI models released since ChatGPT, publisher by publisher: what each release changed, the score the publisher published that day, and how to read these figures without getting it wrong. A comparison kept up to date.
Understand7 min
AI news for leaders, October 2026: what changes
New models from Mistral and Google, a well-known software company switching provider, websites closing the door to agents, a French study on hiring young people and a cyber law postponed again. The facts of October 2026, and what each one changes for a business.