ClaudIA reads the entire interaction and automatically fills in the custom attributes for the conversation and contact. Instead of the agent typing the reason for contact, category, and similar details at the end of the interaction, the AI extracts this information from the conversation itself.
Where to find it
Settings → AI → Tags & Attributes — screen restricted to Administrator profile.
If the item does not appear in the menu, enabling this feature on your account is handled by the Cloud Humans team — contact us to request it.
Tags: automatic tag application is under development. Currently, the screen fills custom attributes; tags will be added later.
The three-decision process
For each field, you define what, how strictly, and — for the entire account — when.
1. What: the instruction
Enabling a field requires writing an instruction: what ClaudIA should fill in that field and how to decide the value based on the interaction. Without an instruction, the field cannot be enabled.
Two aids in the instruction modal:
- Generate with AI — ClaudIA proposes an instruction based on the field's name, type, and description. Review before saving.
- Analyze — evaluates the quality of the instruction and returns Excellent, Good, or Needs Improvement. If you edit after analysis, the screen warns that the result is outdated.
A good instruction indicates where to look and how to choose. Compare:
- ❌ "Reason for contact."
- âś… "Fill with the main reason for contact, choosing the most specific category based on what the customer reported. If the customer mentioned multiple topics, use the one that triggered the opening of the interaction."
2. How strictly: minimum confidence
Each field has a minimum confidence to fill, adjustable in 5% increments. The default is 70%.
The AI returns, alongside the value, how confident it is in that reading. Below the minimum, the value is not saved — but it is recorded with the reason for auditing. Increasing the number makes the AI more cautious (fewer fields filled, higher accuracy); decreasing does the opposite.
3. When: triggers
Triggers apply to all enabled fields, not per field:
| Trigger | When it fires | Default |
|---|---|---|
| Resolution | when the conversation is resolved | on |
| Transfer to human | when ClaudIA hands over the conversation to a person | off |
| Manual | when an agent triggers on demand | off |
With the Manual trigger enabled, a Fill with AI link appears at the top of the conversation's attributes section. The filling runs in the background; the result appears in the conversation automatically, without needing to reload.
Limit of enabled fields
Your account has a limit on the number of fields enabled simultaneously, shown at the top of the screen (Enabled fields 3 of 5). The default is 5 fields. When reaching the limit, you must disable one to enable another. If your operation requires more, the limit can be adjusted by the Cloud Humans team.
What happens in the conversation
The most important rule: AI only fills empty fields. Values entered by an agent are never overwritten — even if the person types while AI is thinking, manual input wins.
Each execution leaves a card in the conversation timeline, summarizing (“ClaudIA filled 3 fields”), the trigger that fired, and who triggered it (System for resolution and transfer, or the agent's name for manual). Opening the card, field by field:
- Filled or Not filled
- Confidence obtained and required (
needs 70%) - Reason — AI's reasoning for that value
When a field is not filled, the card explains why:
| Result | What happened |
|---|---|
| Low confidence | AI found a value, but below the minimum |
| Already filled | the field had a value; AI does not overwrite |
| Abstention | conversation lacked evidence to decide |
| Hierarchy conflict | response contradicts the level above in the list hierarchy |
| Invalid | value does not match the field's type or options |
In other words: even when AI does not record anything, you see what it suggested and the reason — enough data to calibrate the instruction or confidence minimum.
Chain fields (nested lists)
Fields linked by value — "appears when Reason = Charge" — form a chain. Enabling the root field covers the entire chain: AI responds to all levels in a single reading, and the most specific answer fills the higher levels. Disabling the root silences the entire chain.
Best practices
- Start small. Enable two or three high-value fields, monitor the cards for a few days, then expand.
- Calibrate based on evidence, not intuition. Many “Low confidence” suggestions are correct? Lower the minimum. Wrong recorded values? Raise the minimum or make the instruction more specific.
- Instruction per field, not per process. Each instruction should refer to only one field, with explicit decision criteria.
- Leverage field structure. In lists, mention options or selection criteria in the instruction — AI decides better when you tell it how to choose.
- Review after changing the field. New options in a list require reviewing the instruction and re-analyzing quality.
Quick questions
Will AI delete or change something my team entered? No. Only empty fields are filled.
I enabled the field but nothing happened. Check three things: is a trigger linked; did the conversation pass through this trigger after enabling (executions are not retroactive); and was the field empty. The conversation card shows what happened at each execution.
Can I fill a field no one sees in the conversation? Yes. AI filling is independent of Attribute Permissions — a hidden field for everyone still gets filled.
Does this work for conversations on any channel? It works for native CloudChat interactions. Conversations mirrored from external systems do not go through filling.
Where can I see the history? In the card of each execution, within the conversation. All decisions are recorded — both those saved and those discarded.