Home · Blog · First line of support on an AI agent: what to trust and what to the operator
First line of support on an AI agent: what to trust and what to the operator
The first line is requests for which the answer has already been written: where to see the status, how to return it, what time you open, how to restore access. In small business support, such questions are repeated day after day, and it makes sense to delegate them to an AI agent. The rest - money, complaints, personal data - should go to the person quickly and without loss. Below is a working breakdown: a table of what we trust the neural network, a list of what we prohibit, three escalation rules, assembly on the “Support” template and a way to calculate what share of requests the AI closed itself.
What is considered the first line
The first line is not “simple questions”, but addressing three properties at once:
- The answer has already been written - in the regulations, on the website page, in correspondence between managers. You don't need to invent it, just find it.
- No action needed on your systems — no need to look in 1C, change data in CRM, or make a refund.
- An inaccurate answer is not worth money and reputation — if an agent gives an hourly error in the work schedule, it’s unpleasant, but not a disaster. If he promises a refund, it will be a disaster.
Test each category of your requests with these three questions. The one that passes all three is the first line, and can be given to the AI agent. What at least one does not pass is the second line, the person. This division leaves the neural network with the frequent and safe questions, and the risky ones with the people, and this is the right proportion.
The agent in CHBT is a script block that responds according to your text knowledge base and transfers the conversation to the operator according to the rules that you set. How it works, which model is responsible and how much the answer costs - in the article about a chatbot with a neural network using a knowledge base. Here we talk about the division of labor.
What we trust in an AI agent
| Category | Examples of requests | What should be in the database |
|---|---|---|
| Help questions | Schedule, addresses, payment methods, delivery times, plan composition | Conditions section: one statement - one sentence, with numbers |
| "How to" instructions | How to issue a refund, change the plan, download a receipt, restore access | Step-by-step lists of 3–6 steps, with the names of the buttons for your service |
| "What to do if..." | I did not receive the code, the parcel is delayed, the file does not open | Typical reasons and first step for each; when to go to the operator |
| Navigation | Where to write, what document is needed, where to see the status | Addresses of pages and sections in text |
| Terms and rules | Warranty in the standard case, return periods, what is included in the service | Points of the offer that concern the client, in your own words |
About order statuses - honestly. The agent does not look into your CRM or 1C and does not know where the package is. He can explain how to see the status yourself (link, track number), and this covers a significant proportion of such questions. This is also true for orders from Ozon and Wildberries: the status number CHBT is not received from the site, so it is better to immediately give “where is my order” with the number to the operator. How to collect the text of the database so that the agent responds to it accurately - in the article about the knowledge base for the AI agent.
What we don’t trust - and why
- Money. Returns, compensation, disputed write-offs, “recalculate”. The language model tends to agree with the persistent interlocutor; It’s easier to promise her a refund than to refuse. The client will remember any promise from the agent as a promise to the company.
- Complaints and conflicts. The person who writes “this is a disgrace” does not want an answer from the database - he wants to be heard. The polite remark of the neural network at this moment is more annoying than silence.
- Personal data. Change phone number, address, e-mail in the order; confirm identity; provide information about another person. The agent does not know how to check what exactly the account owner writes, and should not try. Customer data cannot be put into the knowledge base - everything in the database may end up in the answer.
- Legally relevant answers in non-standard cases: warranty in case of a disputed breakdown, liability for delay, termination conditions. Here we need a formulation for which a person is responsible.
- Anything that requires action in your systems: post, cancel, change.
One line rule: neural networks are where an inaccurate answer is worth an apology; to a person - where he is worth money, reputation or breaking the law.
Three rules of escalation
Escalation is the transition of conversation from an agent to a person. In CHBT it is assembled from three mechanisms, and all three are needed.
1. Routing to the neural network. In the greeting there are buttons by category: “Question about the order”, “Return and payment”, “Complaint”, “Another question”. The “Return and Payment” and “Complaint” branches lead directly to the block “Operator”, bypassing AI. "Another question" - in “AI agent”. So the second line does not reach the neural network at all, and this is more reliable than any ban in the role.
2. Keyword. There are no transfer settings in the agent card - manual access to a person is done by a separate script with the “Keyword” trigger in the “contains” mode: “operator”, “person”, “manager”, “complaint”, “return”, “money” → “Message” “I am passing on to a colleague” → “Operator”. The keyword is also triggered in the middle of a conversation with an agent, so a person can approach the operator from anywhere. Expand the list according to your conversations - for example, “boss” or “complaint”; Don’t put too short words in the “contains” mode - they will intercept other people’s answers.
3. Uncertainty. The “didn’t understand” port of the “AI agent” block fires itself when the agent’s response sounds like a refusal or uncertainty - “no information”, “check with...” - and diverts the conversation to a separate thread; connect it to the “Operator” block. The agent should not guess: “I’ll check with a specialist” and transfer is the correct outcome.
And one technical rule: after the “AI agent” block in the script there is always “Operator” - both for transmission and for output, in case the credits are exhausted. An agent without access to a person is a trap for the client. At night, when there are no operators, the conversation will still end up in “Conversations” and will wait until morning; warn about this in text before transmission, adding a “Condition” block based on the time of day with different messages (how to configure branches - in the help conditions in the flow).
Assembly on the “Support” template
Template “Support” - short menu and quick transfer to the person. Open it with the “Use Template” button and modify it in “Scripts”:
- In the welcome menu, leave 3-4 buttons for the categories from the section above. Enter the buttons for the second line directly into “Operator”.
- In the “Another question” thread, paste “Action” → “Add tag”
AI-line, then block “AI agent” with your agent from the “AI Agents” section. - After the agent - “Action” → “Add tag”
escalationand “Action” → “Notify operator”, then “Operator”. The agent's exit in case of exhaustion of credits is the same. - For the “Complaint” branch, add "Data collection" with the order number in its field - the operator will receive it in the contact card before opening the conversation.
The flow is the same for all channels: connect Telegram, MAX, VKontakte and the widget in “Channels” - clients will write where it is convenient for them, and operators will respond from only “Conversations”. The bot does not send media in messengers (only text and buttons); files - screenshots, receipts - the client can only send in a widget on the website, and they will be seen by the operator, not the agent.
Role of the support agent
The role (system prompt) for support is stricter than for a sales consultant. A working version of eight lines:
You are a support employee for company X. Answer in Russian, using “you,” briefly: up to three sentences. Answer only according to the knowledge base. If the answer is not in the database, say that you will check with a specialist and offer to call an operator. Do not promise returns, compensation, discounts or terms that are not in the database. Do not ask for card details, passports or passwords. If a person is annoyed or writes about a complaint, apologize in one sentence and suggest the operator. Do not discuss other clients or internal company rules.
The settings are nearby: temperature is low, we need precision, not variety; model - average for the start, strong - if the agent confuses similar instructions. Test chat of an agent in the office - run 10 requests from real correspondence, including two complaints and one request for a refund: both types should end with a proposal from the operator, and not a substantive answer.
The role does not replace routing: even with an ideal prompt, some complaints will reach the neural network if there is no “Complaint” button in the greeting. Prompt is the second barrier, buttons are the first.
How to calculate the share closed by AI
The first line metric is the same: how many requests were closed without a person. It is calculated from two sources.
"Analytics". The bot dashboard shows incoming and responses from the bot and operator by day for 14 days, the number of conversations that are waiting for the operator, and conversations by channel for 30 days. It shows whether the load on operators is growing, but not the share of closed AIs: the dashboard does not know how many conversations reached the “AI Agent” block - there are no block counters in CHBT.
Tags Therefore - tags AI-line and escalation from the assembly above. In “Contacts”, the filter by tag gives the number of people with each; escalation/ii-line ratio - proportion of those transferred. More precisely, export contacts to CSV and calculate by dates. Remember that the tag is placed on a contact, not on a conversation: one person with three requests is counted once.
Guidelines, not norms: if a person has more than half of the conversations that reach the agent, either the database does not answer real questions, or routing connects a second line to the neural network. Parse 20 transmitted conversations in “Conversations” (the “Waiting for a response” tab or contacts with the tag escalation) and sort it out by reason: not in the database, didn’t understand, asked the person right away. The first reason is to replenish the database, the second is to reformulate the question-answer pairs, the third is normal, as it should be.
CHBT does not keep a list of “unanswered questions” - reading the conversations remains the only way to find out what the database is missing.
Regulations for operators
- The transmitted conversation is in “Conversations” with the status “Operator” (tab “Awaiting response”); notifications - sound and browser in the inbox, Web Push when the tab is closed. Assign the conversation to yourself so that your colleagues do not answer in parallel.
- Read the agent's responses higher in the feed - they are saved in the conversation. Do not repeat what the client has already written to the neural network.
- Quick answers — templates in the inbox for repeated escalations: “I’m checking the status of the return, I’ll respond within an hour.”
- Close the dialog after a decision, the “Closed” status (the “Closed” tab) separates resolved requests from the queue and simplifies counting.
- Once a week — 15 minutes to replenish the database: the person in charge reads the escalations and adds question-answer pairs. Without this, the share of closed AI does not grow.
Roles in the “Access” section: operators – “Operator”, those responsible for the database – “Admin”. On “Basic” and “Business” the number of operators is not limited, so duty can be distributed among several people without additional payment.
Will there be enough loans: calculation
The agent's response to the default models is one loan, more expensive for powerful models. “Basic” (650 ₽ per bot per month) gives 150 credits, “Business” (1290 ₽) - 500. For evaluation, let’s take three responses to the conversation - clarification, answer, “thank you”; You can see how much you get out of it from your conversations. Then:
| AI requests per month | Credits for 3 answers | plan |
|---|---|---|
| up to 50 | up to 150 | "Basic" plan |
| 50–160 | 150–480 | "Business" |
| more than 160 | more than 500 | Not everything is through AI |
“Not everything is through AI” is not a refusal, but a hybrid: the most common questions are closed with buttons or keywords with ready-made messages (ready-made template "Autoresponder FAQ"), and it's not worth the credits; the neural network is just behind the “Another question” button. When the credits run out, the agent responds with a neutral text and the script goes along the fallback branch - in your build it leads to the operator, so the client will not be left without a response, but the load on people will increase until the limit is updated.
To estimate: if today operators close 200 requests per month and half of them are inquiries, “Business” will cover them with a reserve, and operators will have a second line. The composition of Plans is on the page Plans page. Create a support bot for free →
“Support” templateTemplate "Autoresponder FAQ"Plans and AI loansKnowledge base for AI agent