AI scenarios

Configure scenarios at /ai/scenarios. Viewing requires view permission; creating and editing require management permission. Without the relevant permission, the card stays unavailable or read-only.

Scenario contract

Set stable key and name, description, category, enabled, useCase (text_improve, text_fix, json_analysis, custom), outputMode (text or json), systemPrompt, userPromptTemplate, tags, and an optional responseSchema. Select connectionId or inherit the module default.

Scenario limits are maxPromptChars, maxOutputChars, maxContextChars, requestTimeoutMs, and overflowMode (reject or truncate). Do not disable the system prompt without understanding the consequence: disableSystemPrompt=true leaves only the user template.

Test before enabling

Run the scenario from Playground with synthetic input. For JSON, ensure the schema is a valid object, required fields are meaningful, and the consuming form can handle the response. Test long input, timeout, missing connection, and partial loading. Deactivate an obsolete scenario (enabled=false) instead of reusing its key.

Form and edit conflicts

Creating or editing requires both key and name. If supplied, responseSchema must be a JSON object; enter tags as a comma-separated list. Refresh in the scenario modal reloads the latest saved scenario data and warns when the record was deleted or cannot be found. When a scenario has a connection, the form can open that connection’s card; deactivation keeps the record and sets enabled to false.

Safety

Prompts must not request passwords, tokens, unnecessary personal data, contract amounts, or another customer’s files. A scenario that changes CRM, tasks, calendar, or documents always requires human review. Its description should name the owner, allowed context, and completion criterion.

If a scenario does not load, does not run, or returns an empty response, show the reason and recheck its settings instead of treating it as success. Demonstrations should use synthetic data, no money or secrets, and labels checked in the interface language.

Related: AI Playground, AI connections, AI and Assistant settings.