AI agent
A RAG-grounded support agent that answers from your knowledge base.
The AI support agent answers customer conversations using a large language model that is grounded only on your knowledge base. It's enabled per channel and runs on Chatasist's managed AI infrastructure — there's nothing to configure about model hosting or API keys.
The models behind the agent were re-evaluated in August 2026. After testing candidates against real production support conversations, the agent moved to models that reliably check your knowledge base before answering instead of guessing. Knowledge-base embeddings continue to run on their own dedicated model. This is a managed-infrastructure change — there's nothing for you to reconfigure.
Enable and tune it#
On any channel, turn on AI answers and pick a tone:
| Tone | Use it for |
|---|---|
| Friendly | Warm, approachable consumer brands. |
| Professional (default) | Most B2B and formal support. |
| Playful | Casual, energetic products. |
| Empathetic | Sensitive or high-emotion support. |
Optionally enable Learn from agent replies so the agent can also draw on how your human team has answered similar questions in the past.
How grounding works#
When a customer message arrives on an AI-enabled channel, the agent runs a similarity search over your knowledge base and answers only from the most relevant chunks plus the recent conversation history. If the answer isn't in your content, it says so plainly and offers to open a support ticket instead of inventing one.
The agent is written to answer like your best human operator: it replies to the actual question in its first message, keeps replies short and concrete, asks for one missing detail at a time, and never re-asks for something the customer already said. It offers a ticket after a real attempt to help — not as an opening move.
The agent knows today's date. When your content states a deadline or validity period — for example a campaign "valid until 12 June" — the agent compares it to the current date and won't present an offer as active once its end date has passed, so it never promotes an expired promotion.
Tools the agent can call#
Retrieves relevant knowledge-base chunks to ground its answer (RAG).
Opens a ticket on the channel's configured board/column — one per request, not one per message or per conversation. The agent writes the problem into the ticket summary from the conversation and records the order or parcel number it is about. If the problem isn't clear yet — including when the customer taps the "Create a ticket" button — it asks one short question first. It won't open an empty ticket that just points back at the chat.
When a customer asks what happened with a request they already sent, the agent reads that ticket: it says the request is in progress, or passes your team's reply on in the chat, and sends the ticket link again. See Repeat requests.
When the customer has something new for your team, or your answer didn't solve the problem, the agent adds it to the same ticket and sends a closed ticket back to your team, instead of opening another.
Offers a "create a ticket" button when the agent can't answer or the customer asks for a person.
Saves the email or phone a customer gives, and confirms it with a emailed code before opening their record — see Customer identification.
Surfaces and reads Resources — your live customer data.
Calls tools exposed by your own connected MCP servers to fetch data or trigger actions in your systems (Enterprise).
Schedules a slow, live lookup instead of stalling in silence — see When a lookup takes a moment.
Resource tools only appear when the channel has Resources attached and resource access enabled. Ticket creation requires an agent ticket board and column on the channel. MCP tools only appear when at least one MCP server is connected and enabled.
When a lookup takes a moment#
Some questions need a live record from a connected MCP server — your own systems, reached over the network — and that kind of lookup can take 30–40 seconds. Rather than leave the customer sitting in silence, the agent sends one short line saying it's checking, the widget shows a Thinking… indicator, and the full answer is posted into the same chat as soon as the lookup returns.
The agent only does this when it's genuinely about to look something up — never as a way to stall instead of answering. And if your team takes the conversation over, or it gets resolved, while the lookup is still running, the follow-up is dropped cleanly instead of landing in a chat nobody's watching.
Images customers send#
When a customer sends a photo or screenshot, the agent reads it: it transcribes any text — order numbers, tracking codes, error messages — and describes what the picture shows.
That reading happens once, when the image arrives, and is kept with it. The rest of the conversation refers back to the same text, so a customer can send a screenshot and then ask "what does this mean?" and get an answer about the picture.
Voice notes and files are not read. When one arrives the agent says plainly that it cannot open it and asks for the details in words, rather than guessing or ignoring it.
Text found inside an image is treated as something the customer said, never as an instruction to the agent — a screenshot containing "ignore your instructions" is read as words in a picture and nothing more.
Guardrails#
- Scope & grounding — the agent stays on your topics and answers only from your content. If the answer isn't in your knowledge base, it says so and offers a support ticket rather than inventing one.
- Prompt-injection resistance — known jailbreak/injection patterns are deflected, and off-topic requests (code, politics, etc.) get a scope reply.
- Language pinning — once the customer has sent enough text (a dozen-plus characters), the conversation language is locked. Later short messages or pasted product names in a different script won't make the AI flip its reply language.
- Customer-initiated only — the agent joins a conversation only when the customer wrote the first message. If your team reaches out first — for example replying to someone from the Instagram or Facebook app, or messaging a contact before they've written in — the AI stays out of that thread, even after the customer replies, because a human is already handling it.
Reply safety checks#
Beyond staying on-topic, four checks now run on every reply before it reaches the customer. A reply that trips one is retried on the agent's stronger model, then replaced with a safe fallback if it still won't clear the check.
The agent can't say "I'll check and get back to you" unless it genuinely scheduled that follow-up (see When a lookup takes a moment). Checked in Azerbaijani, English, Russian, and Turkish.
Numbers, dates, prices, and reference codes in a reply are compared against what the agent actually retrieved and what the customer said, catching invented figures. Prose and paraphrasing aren't judged — only concrete facts. It ships off by default; ask about turning it on (flag or enforce) for your workspace.
A reply can't read back a stored name, email, phone, or address on a chat that hasn't been verified, and can't claim a chat is verified when it isn't — see Customer identification.
Leaked internal text, raw markup, and looping repetition are stripped before a reply ever reaches the customer.
Apologies for a late reply. The agent says sorry for the wait only when the customer really has been waiting — more than five minutes since the first message of theirs that nobody answered — and then once, not at the start of every message. A stock "sorry for the wait" in front of a reply that arrived within seconds is removed before it sends. An apology about something specific, such as a late parcel, is the agent answering the customer and is left as written.
The agent also can't claim to have transferred a chat to an operator — it has no such ability. A reply that makes that claim is caught and replaced before it sends; your team can still take over any conversation from the inbox.
Automation behaviors#
| Behavior | What happens |
|---|---|
| Closure detection | Detects when the customer signals they're done and sends a "anything else?" resolution probe. |
| Resolution follow-up | If the customer confirms no further help is needed, the chat is resolved and a CSAT survey is requested. A new question reopens the support flow. |
| Survey response | A bare 1–5 reply (with an optional short comment) to a pending survey is recorded as the rating deterministically — no model involved. Anything else is answered normally and the survey stays pending. |
| No-reply auto-close | If a conversation goes quiet after a reply and the customer doesn't respond, it's closed automatically and a CSAT survey is requested after a short wait. No reminder message is sent first. |
| Ticket-intent summarizer | Generates a clean ticket title and summary from the conversation. |
| Ticket offer | When the agent can't answer from your knowledge base — or the customer clearly asks for a person (e.g. "connect me to an operator") — it offers to open a support ticket so the request reaches your team. On the website widget this shows a Create-ticket button; on Instagram, Facebook Messenger, WhatsApp and Telegram (where there are no tappable buttons) the agent makes the offer in plain text. Your team can also take over any conversation directly from the inbox at any time. On a holiday, the agent also mentions it's a non-working day and that follow-up may be slower. |
Meta AI persona#
On Facebook Messenger channels you can give the agent its own Messenger persona so its replies render under a distinct name and avatar.
Usage, quota & cost#
Each plan includes a monthly AI-resolutions allowance. A resolution is counted when the AI agent resolves a conversation unassisted (no human agent touched it). When the quota is used up, the agent soft-stops answering new conversations — customers can still open tickets, and your team replies from the inbox as usual. There's no overage charge, and the counter resets each billing period. Workspaces with BYOK (bring your own AI key) bypass the quota entirely. See Plan limits & enforcement.
Chatasist also tracks token usage and an estimated cost per workspace and per conversation (prompt, completion, and embedding tokens) for reporting.
Give the agent accurate content to answer from.