Chatbot and the angry customer: the first three messages

A complaint is a bot's hardest test

«My order hasn't arrived for three days, nobody answers, this is a disgrace!» — when a message like this comes in, the chatbot's first reply decides a lot. A poor reply — «Hi! How can I help?» or a template about order status — makes the customer angrier, because it shows the bot did not listen. A good reply calms things down, collects what is needed and gets the conversation to someone with authority quickly.

This article shows step by step what a chatbot should do when an unhappy customer writes: how to recognise risky intents, what to say in the first three messages, what to collect, what the bot can resolve and what it must never resolve, escalation and later analysis. The full catalogue of escalation cases is a separate article; here we go deep into one case — dissatisfaction.

Risky intents: recognising dissatisfaction

Dissatisfaction does not always come with an exclamation mark. The bot should recognise four kinds.

1. 1Open angerCapitals, exclamation marks, «disgrace», «outrageous», insults. Easy to spot.
2. 2A repeat contact«I'm writing again», «third time», «I told you yesterday». The anger may be hidden, but patience has run out.
3. 3Cold dissatisfaction«I'll never buy from you again», «give me my money back». Polite, but the customer has already decided to leave.
4. 4A threat«I'll take you to court», «I'll post about this», «I'll go to consumer protection». This needs a person immediately.

Vexvon's analytics report shows the share of positive, neutral and negative messages month by month, which helps track the overall trend. But in a given conversation, the decision comes from the bot's rules.

The first three messages

1. 1First — show it was heard«Waiting three days without an answer is genuinely frustrating — we're sorry.» Short, specific, referring to what the customer said. A template apology — «we apologise for any inconvenience» — does not work.
2. 2Second — ask for what is needed«Please send your order number so we can check exactly what happened.» One question, no explanation.
3. 3Third — say what happens next«I've passed this to the specialist on duty; they'll write to you within 15 minutes.» A concrete time, a concrete next step.

Calming language: what works and what doesn't

1. 1WorksReferring to the customer's words («three days»), a specific apology, short sentences, the time of the next step.
2. 2Doesn't workLong explanations, repeating «we understand», emoji, quoting company policy, firing questions at the customer.
3. 3RiskyPromises: «it'll arrive today», «you'll be refunded». The bot cannot promise an outcome it does not know.

A complaint in a public comment

A complaint under an Instagram or Facebook post is seen by everyone, and the reply is for everyone too. Two rules apply: the public reply is short and polite, and the details move to a private conversation.

1. 1Public reply«We're sorry about this. We've sent you a private message so we can check your order right away.» No argument, no explanation, none of the customer's personal details.
2. 2Private messageThe order number, the problem and the solution are discussed only in DM. In Vexvon, on Instagram and Facebook comments the sensitive reply goes to a private message and a short reply to the comment; if the private message fails, the public reply stays as a fallback.
3. 3Don't deleteDeleting or hiding a complaint usually makes things worse. An answered complaint shows other customers that the team responds.

Information to collect

The bot's main value is that the agent does not have to ask the customer again. With a complaint, the bot asks only for what the resolution needs:

• The order, contract or ticket number
• What happened — in the customer's own words, a sentence or two
• What the customer wants: delivery, an exchange, a refund, or just an explanation
• A suitable time and channel to get in touch

If the customer already gave these details in earlier messages, the bot does not ask again — which is why conversation history matters.

Authority to resolve: what the bot can do

1. 1The bot canShow the order status (if live data is connected), explain the rule, log the request, hand over to a person with a concrete time.
2. 2The bot cannotDecide on compensation, a discount, a refund, an exception, or who is at fault. Those are decisions for an authorised employee.
3. 3The bot must notArgue with the customer, blame the courier or another party, or say the complaint is unfounded.

Escalation: when and to whom

• Threats, legal words, insults — immediately, without queueing
• A repeat contact — immediately, with the history of the earlier contact
• A refund or compensation demand — to an employee with the authority
• Cold dissatisfaction («never again») — to whoever owns customer relations
• The customer types «agent» or «person» — always immediately

In Vexvon, when a customer writes «I want an agent» the team is notified, and on swearing or abuse the bot goes silent on its own. Once an agent joins, the bot stops in that conversation. The full list of triggers is in chatbot escalation rules.

A sample conversation

1. 1Customer, 14:02«This is the third time I'm writing! The order still hasn't come, give me my money back!»
2. 2Bot, 14:02«We're sorry you've had to write three times — that shouldn't have happened. Send your order number and I'll pass this straight to the person responsible.»
3. 3Customer, 14:03«48213.»
4. 4Bot, 14:03«Noted: order 48213, delay, refund request. It's gone to the manager on duty, who will write to you within 15 minutes.»
5. 5The manager's screenRepeat contact, dissatisfaction type «anger», order 48213, request «refund», the bot promised 15 minutes. The manager writes at 14:09 and starts from the solution.

What the agent sees: the context packet

At handover the agent's screen should show the following, so they can start without asking questions:

• The kind of dissatisfaction: anger, repeat, cold, threat
• The order number and its status
• The outcome the customer wants
• What the bot has already said and what time it promised
• Earlier contacts, if any

Analysis afterwards

Every complaint conversation is information about both the customer and the process. Once a week, read the complaint conversations and answer three questions:

• What is the root cause — delivery, product quality, a slow reply, wrong information?
• What did the bot do well, and where did it make things worse?
• Which process needs fixing so this kind of complaint becomes rarer?

The share and trend of negative messages are in the analytics report. Calling a customer after a negative review is a separate process, covered in the negative review recovery call.

Limits

A chatbot does not resolve dissatisfaction — it keeps it from getting worse and gets it to the right person quickly. A bot does not always read irony, sarcasm or cultural nuance correctly, so handing over is safer when in doubt. With complaints about health, safety or financial loss, the bot only collects details and hands over at once.

Conclusion: the bot calms, a person resolves

When an unhappy customer writes, the chatbot's job fits into three messages: show it was heard, ask for what is needed, and say what happens next with a concrete time. Everything that needs a decision — compensation, exceptions, refunds — goes to an authorised person with full context. A bot built this way does not inflame a complaint; it speeds up its resolution.

The rest of the support section is in this category. To write your complaint scenarios together, get in touch.

Frequently asked questions

1. 1Should the bot answer an unhappy customer, or go straight to a person?The bot shows in the first message that it heard and collects the details; with threats, abuse or a repeat contact it hands over at once.
2. 2Should the bot apologise?Yes, briefly and specifically — referring to what the customer said. But not with a sentence that admits fault or promises an outcome.
3. 3Can the bot offer compensation?No. Compensation is an authorised employee's decision. The bot only logs the request.