Organization & reporting
Intelligence
Control interpretation, response style, AI budgets, and recorded outcome feedback.
Open this workspace pageRead the analytics
- Choose Last 7 days, Last 30 days, or Last 90 days in the Analytics period selector.
- Review LLM avoidance, Conversations, Estimated AI cost, and Tokens consumed.
- Under How requests were resolved, compare recorded sources such as rules, templates, typed decisions, and optional models.
- Under Quality and budget, review confidence, handover rate, monthly committed spend, unconfirmed provider charges, and recorded feedback.
- Remember that these are operational measurements, not a guarantee that a classification or external action was correct.
LLM avoidance measures requests without a generative model call; typed-decision calls can still have cost. A dash in a confidence statistic can mean no relevant observations. Timed-out provider calls keep estimated budget reservations until reconciled.
Set the organization controls
| Control | How to choose it |
|---|---|
| AI budget mode | Economy asks for clarification on uncertainty. Balanced allows interpretation. Premium also permits optional language enhancement when configured and allowed. |
| Response style | Fast uses templates; Balanced can shorten long approved answers; Generative adds optional wording and summaries in Premium mode. |
| Use Korenact AI typed decisions | Enables the configured typed-decision service after rules and known answers. If the service is unconfigured, rules and safe fallbacks remain. |
| Automatic confidence | Threshold for automatic interpretation; the form accepts 0.70–1.00. |
| Clarification threshold | Lower-confidence handling threshold; the form accepts 0.50–1.00. Keep thresholds consistent with the approved workflow and server validation. |
| Monthly AI budget (USD) | Blank means unlimited. Zero disables paid AI calls. This is an AI spend control, not a subscription limit. |
| Tokens per conversation | Limit for model usage attributed to a conversation. |
| Agent tokens per month | Per-agent monthly token budget; blank means unlimited. |
| Maximum output / input tokens | Bounds for individual model requests. Output accepts 64–2048; input accepts 256–16000. |
- Change only the controls approved for the organization.
- Choose Save controls.
- Wait for “Intelligence controls saved.” If an error appears, correct the indicated setting.
- Test representative requests and inspect the recorded decision and cost.
- Changes apply to the organization’s next request.
Advanced models, queues, commands, and templates
Open Models, queues, exact commands, and templates only when you are configuring an approved advanced setup. The editor uses JSON. Keep the displayed keys and structure, and ask the administrator to supply model identifiers and server-installed credential references.
| Setting | Purpose |
|---|---|
| allowed_model_routes / preferred_model / models | The model routes, choice, and per-model configuration approved for this organization. Credentials stay on the server. |
| routing_queues | Department or department:language destinations, such as billing:es. Routing & staff provides a more direct editor for conditional routes. |
| exact_commands | Map specific phrases to registered intents rather than free interpretation. |
| templates | Approved response text using supported variables such as {{answer}}, {{reference}}, {{date}}, and {{time}}. Missing values keep the standard response. |
Keep API keys, tokens, and patient-specific secrets out of this configuration. Select valid JSON, choose Save controls, and test before expanding its use. Rules still authorize actions regardless of the interpretation source.
Record an observed outcome
- Find the relevant Recent decisions entry and open its linked conversation.
- Compare the actual conversation and provider outcome with the intended request.
- Expand Decision details to review recorded operational metadata.
- Select Observed outcome: Resolved, Transferred, Appointment booked, Route corrected, Asked again, Workflow failed, or Human override.
- Enter a concise Staff note if useful, without adding unnecessary sensitive data.
- Choose Record outcome and confirm the recorded label.
Owners, admins, and supervisors can record feedback. Feedback does not edit rules, retrain a model, rerun an action, or verify a booking by itself. The decision record contains operational classifications and results, not a model’s private chain of thought.