AI configuration
Iterochat's AI answers your customers from your knowledge base. This page covers the controls that shape how it does that: its tone, the questions it surfaces, whether it ever hands off to a human, and how you measure the value it adds.
You configure all of this from the dashboard. Changes take effect on the next customer message, with no redeploy.
Custom AI instructions
Open the AI Instructions tab to tell the AI how to talk to your customers: tone, personality, and brand voice. You write instructions as titled blocks. Each block has two parts:
- Title: a short label so you can recognize each rule at a glance.
- Content: the instruction itself, up to 4,000 characters per block.
For example, a block titled "Friendly tone" might say: "Always greet customers warmly. Use a conversational but professional tone. Refer to the company as 'we' and 'our team'."
You can create as many instruction blocks as you need, reorder them, and toggle individual blocks on or off without deleting them:
- Disabled blocks stay in the list, greyed out, so you can keep a draft around and switch it on later.
- Enabled blocks are applied live to every new conversation.
Custom instructions only steer tone, persona, wording, and style. They cannot override:
- the AI's language rules,
- its answer rules (it only answers from your knowledge base, and it won't invent facts), or
- its escalation triggers.
If an instruction conflicts with any of those, the conflicting part is ignored. Safety and escalation rules always stay in place.
AI model
The AI Model card in Settings lets you choose which model powers your assistant. Models differ in how they balance answer quality, speed, and how quickly they use your plan's included usage.
- Each option shows a short description so you can pick the right fit, and a sensible default is marked Recommended.
- More capable models tend to give higher-quality answers but use your balance faster. Lighter models are faster and more economical.
- Some models are available only on paid plans. Those appear locked until you upgrade.
Only owners and admins can change the model. The change takes effect on the next customer message.
AI identity
The AI identity card in Settings → AI & automation sets the name and face customers see on AI replies. This is separate from the widget header title.
- Name is the label next to each AI message (default “AI Assistant”).
- Icon is a built-in face (sparkle, bot, and similar).
- Photo is optional. Upload a JPEG, PNG, or WebP (up to 2 MB); it is cropped to a square. A photo replaces the icon until you remove it.
Only owners and admins can change the identity. The inbox, playground, and widget apply the current name and face at display time, including on older conversations. Each change is written to the Settings → Compliance audit log.
Test in the Playground
Before a change reaches real customers, you can try it in the Playground. Open the Playground tab (available to owners and admins) and chat with your assistant exactly as a customer would.
The Playground runs your live setup: your current knowledge base, your AI instructions, and the model you selected. So what you see is what a customer would get. For each message you can check:
- The reply itself, in the language you wrote in.
- Whether the AI would escalate to a human instead of answering, and why.
Use it to sanity-check a new instruction, confirm a freshly added document changes the answer, or see how the AI handles a tricky question, all without touching a real conversation.
Playground chats use the real AI, so they draw on your plan's included usage just like a live customer chat. They never appear in your conversation queue or metrics.
AI-only mode
By default the AI answers what it can and escalates the rest to your human agents. AI-only mode changes that. While it's active, the AI handles every conversation on its own and never escalates. Instead of offering to connect the customer to a person, it answers from your knowledge base. When it doesn't have the answer, it says so honestly and suggests the customer rephrase or reach you through another channel.
You control AI-only mode from the AI-Only Mode card in Settings, two ways:
- Manual toggle: flip the switch to turn it on or off right now. The card shows the current state as "Active, manually on" or "Inactive".
- Schedule: set the days and time windows when AI-only mode should turn on automatically (for example, outside your support hours). For each day you enable, set a start and end time, or leave both blank to mean all day. Pick the timezone the schedule runs in, then save. When a scheduled window is active, the card reads "Active, scheduled".
The manual toggle and the schedule work together. The card always reflects whether AI-only mode is currently active, whichever turned it on.
Use AI-only mode when no agents are available to take a handoff: overnight, on weekends, or while you're still piloting. With it on, your customers always get an answer or a clear "I don't have that" rather than waiting in a queue no one is watching.
Suggested questions
The Suggested Questions card sets the tappable chips a customer sees when they first open the chat widget. A click sends that question as the customer's first message. This gives people an easy way to start and steers them toward what the AI answers well.
You can set up to four questions, each up to 120 characters. From the card you can:
- Add, edit, and remove questions.
- Hit Restore defaults to go back to the built-in starter set.
- Remove all of them and save to turn the chips off entirely. The widget then opens with no suggestions.
Make your chips mirror your most common real questions ("Where's my order?", "How do I reset my password?") and make sure your knowledge base actually answers them. A chip that leads to a confident answer is the best first impression the widget can make.
Support cost
The Support Cost & ROI card lets you put a number on what the AI saves you. Enter your hourly support cost and pick a currency (USD, BDT, EUR, GBP, or INR).
Iterochat then estimates money saved on the Metrics tab. The estimate is based roughly on the number of AI-handled chats multiplied by about five minutes of agent time each, at the rate you set.
Leave the rate blank to hide the money figure. The rest of your metrics still work, you just won't see a currency value. This setting is for your own reporting; it never affects how the AI behaves or what customers see.
Files customers send
Customers can attach files to a chat. The AI reads some of them directly and uses what they contain when it answers:
- Images (JPEG, PNG, GIF, WebP): the AI can see the image, so a customer can send a screenshot of an error or a photo of a product and ask about it.
- PDFs: the AI reads the document and can answer questions about what is inside.
- Plain-text files (
.txt): the AI reads the text. Very long files are read up to a limit, so the most important content should come first.
Other file types, including video, audio, and office documents such as .docx and .xlsx, are still delivered to your team, but the AI does not read their contents. It notes that a file was attached and, when it cannot answer from it, hands off to a human as usual.
Files are only read as part of answering the customer. The AI still answers from your knowledge base and the conversation, and it escalates when it should.
When the AI escalates
Outside of AI-only mode, the AI escalates a conversation to your agents the moment any of the triggers below apply.
The escalation triggers are:
- Knowledge gap: the answer isn't clearly in your knowledge base. The AI escalates rather than guess.
- Human request: the customer explicitly asks for a person ("agent", "real person", "human", "support").
- Sensitive intent: refunds, complaints, locked or compromised accounts, payment disputes, or money that hasn't arrived. These always go to a human.
- Frustration: the customer is clearly frustrated after a failed attempt ("you don't understand", "this is useless", or the same question repeated).
When it escalates, the AI sends the customer a short, polite message that it's connecting them to a team member. It also writes a handoff summary for the agent: two to four sentences in English covering what the customer wants, the key facts already gathered, and what was tried.
Owners and admins can send escalation alerts to external systems with outbound webhooks. Webhooks fire on every escalation (escalation.created) and optionally again if nobody claims the chat within a time limit you set (escalation.unclaimed).
Your agent reads that summary at the top of the conversation and picks up exactly where the AI left off. See Handling conversations.