An AI that answers from your data.
The reply lands while they are still deciding, in their own language.
An AI reply agent reads an incoming customer message and writes an answer using the business's own records. It is given the knowledge base, product list, opening hours and the customer's past messages, and it is instructed to answer only from those. What it used is stored with the reply.
- Answers assembled from your knowledge base, orders and hours
- Writes back in the language the customer wrote in
- The sources used are recorded against every reply
- One switch turns it off for a workspace or a single thread
How it works.
- 01
Give it something to read
Add your policies, product information and common answers to the knowledge base, and record orders and opening hours. The agent has no source material until you do.
- 02
Let a message arrive
When a customer writes, the message is matched to a customer record and the agent is given the answers it found, the customer's recent orders, the catalog, business hours and the thread so far.
- 03
Read the draft
The draft arrives with the topic it was filed under, how sure the agent is, and whether it wants a person to look. The sources it used are shown beside it.
- 04
Decide who sends
With the agent switched on, routine replies go out and every one is logged. Switch it off for a workspace or a thread and your team answers instead.
Grounded in what you connected
The agent is not asked to know your business. It is handed the material at the moment a message arrives, and told to answer from that and nothing else.
- The passages from your knowledge base that match the question
- The customer's recent orders, and a rule against inventing a status
- Your product or menu list, with real names and prices
- Opening hours, so it does not promise something out of hours
- What you already know about this customer, and what they asked before
Sources kept with the answer
Every reply keeps the list of what it answered from: the knowledge you wrote, the order, the help article, your opening hours, what you know about this customer.
- The list is written with the draft, so it is there for every reply
- An answer with no source behind it is marked as such
- Anything the agent cannot answer cleanly goes to a person
Speaks the customer's language
The agent mirrors the language and script of the latest customer message instead of translating the conversation into one house language.
- Language and script matched to the customer's last message
- No switch of language unless the customer switches first
- Voice notes transcribed first, then answered as text
Off is a real setting
There is a switch for the whole workspace and a switch for a single conversation. Turn either off and the agent stops drafting there.
- Workspace switch that stops every automatic reply
- Per conversation pause, optionally with a time limit that lifts itself
- Blocked contacts are skipped entirely
- Every automatic send is written to the audit record
Promises it cannot make
The agent is forbidden from offering a refund, a credit or a discount your own answers do not support, and from telling a customer where an order is when it does not have the order.
- No refund, credit or discount it was not given permission for
- A customer cannot talk it out of its instructions
- Refund, complaint and billing questions recognised as those
Hold sensitive replies for review
A refund question waits for a person even with the agent switched on. So does anything the agent is not sure enough about.
- Topics you nominate always wait for approval
- Anything below the confidence you set waits too
- Held drafts sit in a queue to approve, edit or reject
What it does not do.
With the agent switched on, routine replies send automatically. The workspace switch and the per conversation switch are how you stop it.
It answers from what you connect. An empty knowledge base and no order records means very little to answer with.
It does not take payments, issue refunds or change an order.
Common questions
Where do the answers come from?
- From your knowledge base, your order records, your product list, your opening hours and the conversation itself. The agent is instructed to answer from those sources.
Will it make up an order status?
- It answers from the order itself and never invents a status. If the order is not in the data it was handed, it cannot report one.
How do I stop it replying to a particular customer?
- Turn the agent off for that conversation. You can set a time limit so it resumes by itself, or leave it off until you turn it back on.
Can I stop it for the whole business?
- Yes. There is a single workspace switch. With it off, no automatic reply is drafted or sent on any channel.
What languages does it handle?
- It replies in the language and script the customer used in their latest message. There is no fixed language list and no per contact language setting.
Does it answer voice notes?
- Inbound voice messages are transcribed and the text is used to write the reply. Where the audio is too poor to transcribe, the message is shown as an audio placeholder instead.
Can I see why it said what it said?
- Every reply keeps the sources it used, so you can trace an answer back to the knowledge passage, the order or the help article behind it.
What if the AI cannot produce an answer?
- The message is left for a person and you are told. It will not keep trying and it will not send something it could not stand behind.
Does it learn from our corrections?
- Your team can rate a suggestion up or down and the rating is stored. Answers improve by editing the knowledge base, which the agent reads on the next message.
Is my data used to train a public model?
- No. Your material is read to write one reply, then it is done. It never goes into a shared training set.
Every question in one place, answered once.