The support agent that knows when to stay quiet
It reads every incoming ticket against your help articles. When it finds a real answer, it replies on its own. When it doesn't, it says nothing and leaves the ticket for a person.
One reply per customer message · Steps aside for any human reply · Silent below your threshold
Ticket #4821
Where do I download my invoice?
Best match in your docs
Your threshold is 70% — this cleared it, so the reply was sent.
SUPPORT ASSISTANT · 1 min ago
Your invoices are under My Account → Billing, where each one has a Download PDF link. Older invoices stay available indefinitely.
A customer writes in
A new ticket arrives, or a customer replies to one that is already open.
It searches your docs
The AI looks for the article that best answers what was actually asked.
It scores the match
That best match gets a relevance score, which is compared to your threshold.
It replies — or waits
Above the bar, the answer posts as a ticket reply. Below it, the ticket waits for a person.
The threshold
One number decides whether it speaks
Before replying, the AI scores how well its best matching article actually answers the question. You set the bar it has to clear.
Your Billing article answers this directly, so the AI handled it in about a minute.
Pricing pages mention plans but not custom deals — a judgement call, so it left it for you.
Nothing in your docs matches. Exactly the kind of ticket a person should read first.
The black line is the threshold — 70% in this example. Set it high and the AI only handles questions your docs answer squarely. Set it lower and it attempts more, but with looser matches.
The guardrails
An AI replying to your customers without supervision is only acceptable if it can't run away with the conversation. These are the limits.
One reply per message
Each customer message gets at most one AI attempt. There is no loop and no follow-up nagging.
Humans win, always
A teammate replying cancels the scheduled AI response before it is ever sent.
Respects the threshold
Below your bar it stays silent rather than sending a vague non-answer.
Stops on closed tickets
Resolved, closed and cancelled tickets are off-limits, so old threads never reopen themselves.
Only your content
Answers come from your synced articles, not from general internet knowledge.
Everything is logged
Each attempt records the relevance score and why it stopped, so you can tune it with evidence.
Timing
Instant, or after a decent pause
Set a delay in minutes or hours before the AI answers. A short pause gives your team the chance to take anything they want to handle personally — and if someone replies first, the AI cancels itself.
Context
What it reads before answering
It doesn't answer the last message in isolation — it sees the whole ticket, so a reply mid-thread makes sense in context.
- Ticket subject
- Full reply history
- Customer name and email
- Urgency level
- Product tags
- Order details
The details
Everything the auto-responder does, in one table.
- Trigger
- A new ticket, or a customer reply on an open ticket. One scheduled attempt each.
- Cancelled by
- A teammate reply, or the ticket moving to resolved, closed or cancelled.
- Threshold
- A relevance score from 0 to 100%. Below your setting, the AI stays silent.
- Delay
- Configurable in minutes or hours before the reply is posted.
- Posted as
- A normal ticket reply, under an author name you choose.
- Status change
- Optionally moves the ticket to "in progress" when it answers.
- Context sent
- Subject, customer name and email, order details, product tags, urgency, and the full transcript.
- Logging
- Every attempt logged with relevance score, cost and the reason it declined to answer.
- Source
- Your synced help articles — the same index the chatbot uses.
Common questions
It schedules at most one reply per customer message. A new ticket earns one attempt, and each customer follow-up earns one more — it never replies to itself, and it never sends twice for the same message.
The moment a teammate replies, the scheduled AI response is cancelled. The same happens if the ticket is resolved, closed or cancelled before the AI gets there.
The reply is posted under an author name you choose, so you decide how to present it. Most teams name it something clearly assistant-like rather than pretending it is a person.
Then the score comes in under your threshold and it stays silent — no "I could not find anything" noise in the ticket, just a ticket waiting for a human as normal.
Yes. Every attempt is logged with the relevance score and, when it declined to answer, the reason it stopped. That log is how you tune the threshold.
It respects ticket status, so anything resolved, closed or cancelled is off-limits. For finer control, the hooks let you opt individual tickets out in code.
Start it at a high threshold
Let it handle only the questions your docs answer squarely, then loosen it once you trust what it sends.