AI features in Iris AI
Applies to: Iris AI. Last reviewed: November 2026.
AI runs across Iris AI: it reads each incoming conversation, helps route it, drafts replies and helps agents write. Every AI feature is off until your administrator turns it on, so you choose what your team uses. An agent always reviews and sends replies; Iris AI does not reply to customers on its own.
Turn features on and off
Go to Admin > AI > Features (Care tab). Each feature has an on/off switch and Settings, where you can choose the AI model it uses and, for some features, give it instructions. Admin > AI > AI Connections lists the AI providers in use and their fallback order.
Admin > AI > Features
What the AI can do on each conversation
| Area | Feature | What it does |
|---|---|---|
| Understand | Summary | Condenses the conversation into a few bullet points. |
| Sentiment and emotion | Positive, neutral or negative, plus the main emotion (for example frustration). | |
| Language detection | Identifies the language of each message. | |
| Entity detection | Picks out brands, products, competitors, places, names and IDs, even where you have no keyword rule for them. | |
| Author signals | Flags likely bots, spam and influencers. | |
| Relevance scoring | Scores how relevant a conversation is to your brand (0 to 100), with the reason. | |
| Route | AI tag suggestions | Suggests existing tags, or new ones, with an explanation. Agents apply or dismiss them. |
| AI priority rules | Sets priority from rules you write in plain language. | |
| AI work queue rules | Routes cases to queues from rules you write in plain language. | |
| Reply | Draft responses | Prepares a draft reply for each of your AI templates, ready when the agent opens the case. |
| Generate draft | Lets agents create a draft on demand, choosing the template, language and extra instructions. Off by default. | |
| Proofread | Fixes spelling and grammar in the agent's reply, in over 100 languages. | |
| Translate | Translates incoming messages into the agent's language, and replies into the customer's language. |
Agents can correct any AI result (for example change the sentiment or relevance), and rate it with thumbs up or down. The AI performance dashboards show how features are used and rated (Iris AI Analytics, dashboards and the Supervisor console).
AI templates: how drafts follow your brand voice
Draft responses are guided by AI templates, under Admin > AI > Templates. A template is not a canned answer. It tells the AI how to reply: the tone, the role it plays, what to include and what to avoid. The content of each draft comes from the conversation itself, the member's history with you, earlier similar conversations, and your community's knowledge base.
Example instructions for a community moderator template:
- Reply in the same language as the member's latest message.
- You are a moderator replying publicly. Be friendly, expert and concise; use plain English, no marketing language.
- Thank the member for the detail they shared. Point to an existing knowledge-base article when one answers the question.
- Keep the reply under 200 words. No emojis.
Create one template per type of reply (for example Community answer, Escalate to support, Thank you and kudos). Each opened case then shows one draft per template.
Relevance scoring
In the relevance scoring Settings, describe in plain language what matters to your brand and what doesn't. For example: "Questions, problems and how-to posts about our products are relevant. Reports that our service or portal is down are highly relevant." and "Job adverts and unrelated chat are not relevant." Add your own brands and sub-brands, and your competitors. Less relevant items are not removed; they are simply ranked lower.
Models, providers and your data
- Iris AI works with several leading AI providers and can fall back to another provider automatically if one is unavailable.
- Your content is not used to train AI models.
- You can use your organization's own AI provider keys, so requests run through your own provider account. Ask your Customer Success Manager to set this up.
- Your Customer Success Manager can confirm the providers, hosting region and data-processing terms that apply to your instance.
Recommended starting set
For a community team, a good first week is: Summary, Sentiment and emotion, Language detection, Relevance scoring (with your own description), AI tag suggestions, and Draft responses with two or three templates. Review the drafts and ratings with your team after a week, then refine the templates.