See which customer conversations are going wrong, and what your content is missing
Most conversations with an AI assistant go fine, and the few that do not are the ones a business needs to see. SLAtech AI reads every finished conversation and marks how it went: the customer's mood, the risk that they will escalate or leave, and the moments that explain it - each one linked to the messages it came from. Next to it, the questions your content could not answer, grouped by meaning, so you can see which missing answer costs you the most.
Why a conversation that goes wrong needs a signal
A customer who is unhappy with a chat rarely says so to anyone. They ask twice, they ask for a person, they mention a refund, and then they leave. Reading every conversation by hand does not scale past a few a day. Signals do the reading and put the handful that need a person at the top.
What is marked in every conversation
The customer's mood - positive, neutral or negative - and whether it got better or worse as the conversation went on. The escalation risk - low, medium or high - with the reason in one line. Flags for the moments that matter: asked for a person, said they would go elsewhere, complained, asked the same question twice or more, got no answer, asked about a refund. Whether their question was left unanswered. Each signal links to the messages it was drawn from, so you can check it in a second. A conversation of one or two messages is marked "not enough data" rather than guessed at, and a conversation that continues is marked again.
In the inbox
Every conversation in the inbox shows its mood and its risk. You can filter and sort by risk, see the reason by hovering over the badge, and jump straight to the messages behind it.
The signals page
One page for the whole account: the share of negative conversations day by day, the most common reasons for escalation, and the high-risk conversations of any period you choose. Below them, the knowledge gaps: questions the assistant found nothing for or answered poorly, grouped by meaning rather than by exact words - the topic, how many times it was asked, examples in the customers' own words, and a button that adds the answer to the knowledge base.
Alerts and your own rules
When a conversation is marked high risk, the account owner gets an email and a message in the channels already connected, such as Telegram. In the workflow builder, a conversation signal is a trigger: you set the conditions on risk and flags, and choose what happens - a webhook, a call to your own system. Every week the owner also gets a summary of the signals in the account's language.
What it does not touch
Signals are worked out in the background after the conversation, so they never slow down an answer, and they do not change what the assistant says. Personal identifiers are masked before a conversation is analysed. Signals are part of the conversation's data: they are included when a customer asks for their data and deleted when they ask for it to be erased.
Frequently asked questions
How do I find out which chatbot conversations went badly?
SLAtech AI marks every conversation with the customer's mood, an escalation risk with its reason, and flags such as "asked for a person" or "asked about a refund". In the inbox you sort by risk and open the messages behind it.
What does escalation risk mean?
How likely the conversation is to end with a complaint, a demand for a person or a lost customer: low, medium or high, always with the reason in one line and links to the messages it is based on.
Can I be alerted when a customer is about to leave?
Yes. A high-risk conversation sends the owner an email and a message in the connected channels. In the workflow builder you can set your own conditions and send it to a webhook or your own system.
How are knowledge gaps found?
From the questions the assistant found nothing for or answered poorly. Similar questions are grouped by meaning, so the same gap asked in different words is counted once, with examples and a button to add the answer.
Does the analysis slow down the assistant's answers?
No. It runs in the background after the conversation and does not change what the assistant says.
Does it work for conversations in Hebrew and Russian?
Yes. Conversations in Hebrew, Russian and English are analysed alike, and the inbox, the signals page and the weekly summary are in the account's language.
Get in touch
Leave your details and we will reply. You do not need a sales call to get an answer.
Other questions
-
How to add an AI chatbot to a Shopify store
Read the answer -
How to add an AI chatbot to a Squarespace site
Read the answer -
How to add an AI chatbot to a Tilda site
Read the answer -
How to add an AI chatbot to a Webflow site
Read the answer -
How to add an AI chatbot to any website
Read the answer -
How to add an AI chatbot to a Wix site
Read the answer -
How to add an AI chatbot to a WordPress site
Read the answer -
AI for a small business: the three places it actually pays
Read the answer -
A Chatbase alternative for websites in Hebrew
Read the answer -
Customer service automation: what to automate, and what to leave alone
Read the answer -
An AI chatbot that sends its leads to Fireberry
Read the answer -
An AI chatbot that adds its leads as clients in Green Invoice
Read the answer -
An AI chatbot for your website that answers in Hebrew, right to left
Read the answer -
An AI chatbot that sends its leads to the iCount CRM
Read the answer -
An AI chatbot for a medical clinic website in Israel
Read the answer -
An AI chatbot for a website in Hebrew, Russian and English
Read the answer -
RAG vs fine-tuning: which one your use case actually needs
Read the answer -
An AI chatbot for your website that answers your customers in Russian
Read the answer -
Find the questions your website does not answer
Read the answer -
A virtual assistant for your website, trained on your own content
Read the answer