AI configuration: models per task
How to set which provider and model an AI task (for example e-mail classification, company recognition, account coding) uses, with which fallback, threshold and instructions.
Under AI config every AI task (for example e-mail classification, company recognition, account coding, workflow steps) has its own configuration: a fast primary model and a fallback model, the confidence threshold and the instruction given to the model. You can also check here whether the configured models are really available.
What the page shows
- The Model routing block shows which provider and model each capability goes to. The API keys are not here but in the system settings.
- The Models resting block lists the models for which the provider reported an exhausted quota; until the Resting until time the system routes the steps elsewhere.
Adding a configuration
- Open Master datas (Google / AI / WIP) → AI config and click the blue + button at the top right (Add).
- Enter Name, Scope key (lower-case letters, digits and underscore only; unique) and Description.
- Set Confidence threshold (between 0 and 1; stricter, for example 0.95, means fewer automatic decisions), the AI active (runs during sync too) switch and Body max characters.
- On the Tier 1 (fast) tab set Provider, Model, Temperature (0.1–0.3 recommended for classification) and Max tokens; on the Tier 2 (fallback) tab the same for the fallback model.
- Write the System prompt (the categories and the answer format) and the User prompt template; Extra context (company knowledge) is always added to the request. The Auto-inject companies and Auto-inject tasks switches also add the companies and open tasks of the company group.
- Set Status and click Save.
Testing the connection and available models
- The row's Test connection icon (plug with a tick) sends one tiny real call for each provider–model pair of both tiers and shows the result per pair. The call is charged.
- The Available models button shows per provider which model names are really available with your account.
- The broom button Clear cache clears the system cache; a summary message confirms it.
Messages
- "The name cannot be empty." — "The threshold must be between 0 and 1."
- "The scope may only contain lowercase letters, digits and underscores." — if the key is taken, the system names the configuration that uses it.
- "The system prompt cannot be empty.", "Enter the Tier1 model name.", "Enter the Tier2 model name." — for workflow step configurations the system prompt and the fallback model may be left out.
Example
In the configuration that classifies Sample Ltd.'s incoming e-mails a fast Gemini model runs on tier 1 and a stronger fallback on tier 2. The threshold is 0.95, so only very confident classifications are accepted automatically. Test connection reports success for both pairs, so the accountant switches AI active on.