AI agents: Customer service

What it does: Answers from the company's approved answers with a source link and hands off anything not covered.

What a person decides: New answers for the knowledge base.

For whom: Customer.

How to measure: Share of questions handled without staff.

What it does: Transcribes voice messages in Kazakh or Russian and drafts a reply.

What a person decides: Sending.

For whom: Staff member.

How to measure: Response time to voice messages.

What it does: Finds promises in conversations and sends a reminder before the deadline.

What a person decides: The final decision on the outcome.

For whom: Staff member, manager.

How to measure: Promises kept on time.

Example result: “Action: Marat should call Yerzhan back by 5 pm — in request #2041 he wrote ‘I'll call back today before five with the exact price,’ and 30 minutes are left. If the price isn't ready, message the customer now with a new time. Impact: the customer isn't left waiting and doesn't walk away upset. Risk: Marat may be out on a job — then the on-duty front desk calls with the same information. Fallback: if Yerzhan already got an answer in another channel, clear the reminder.”

What it does: Recognizes a formal claim, sets the legal deadline and drafts a reply.

What a person decides: The reply and compensation.

For whom: Staff member, manager.

How to measure: Claims answered on time.

What it does: Tells the customer the order status from the orders export.

What a person decides: The final decision on the outcome.

For whom: Customer.

How to measure: “where's my order” calls.

Example result: “Your order #1042 is packed and already with the courier — we'll deliver tomorrow between 10:00 and 14:00, and the courier will call an hour ahead. If another time suits you better, just let me know and I'll pass it to delivery.”

What it does: Checks return eligibility against the rules and prepares a decision.

What a person decides: Approving the return.

For whom: Staff member.

How to measure: Time to a return decision.

Example result: “Action: process a return for Aidana — bought 9 days ago (receipt #5531, September 29); the photos show no signs of wear, tags attached; store policy allows returns within 14 days. Impact: she gets a decision today, not after a week of messages. Risk: she paid with Kaspi installments — the refund takes longer through the bank, so let her know. Fallback: if signs of wear turn up on inspection, offer an exchange instead.”

What it does: Tracks warranties, checks eligibility, and updates the customer on repair status.

What a person decides: The warranty decision.

For whom: Customer, staff member.

How to measure: Warranty response time.

Example result: “Your phone's warranty is valid until March 12, so the repair is free for you. We're replacing the charging port now — the phone will be ready Thursday after 3 pm, and we'll message you as soon as it's ready for pickup. Please bring your warranty card.”

What it does: Answers from the knowledge base, collects contact details and the request, and sends staff a list in the morning.

What a person decides: The final decision on the outcome.

For whom: Customer.

How to measure: Leads saved overnight.

Example result: “Good evening! We're closed for the day, but your question won't get lost — tomorrow from 9:00 the front desk will reply to you first. What's your name, and what can we help with? If it's about prices or our address, I can answer right now.”

What it does: Translates replies and texts into Kazakh and back.

What a person decides: Publishing.

For whom: Staff member.

How to measure: Share of replies in the customer's language.

Example result: “Сәлеметсіз бе! Жазылуыңызды ертеңге, сағат 15:00-ге ауыстырдық. Уақыт ыңғайсыз болса, жазыңыз — басқа уақыт ұсынамыз. (Hello! We've moved your appointment to tomorrow at 15:00. If the time doesn't work, let us know and we'll offer another.)”

What it does: Sums up the customer's history and the question in three lines.

What a person decides: The final decision on the outcome.

For whom: Staff member.

How to measure: Time to get up to speed on a conversation.

Example result: “Situation: Assel has been a customer for 2 years; yesterday she complained about the wait (request #318), today she asks to move her appointment from Thursday to Friday. Key point: after yesterday's complaint she mustn't be kept waiting again. Options: 1) Friday morning when there's no line — eases the frustration, but there may be no opening; 2) another day — risks her feeling unheard. Recommendation: Friday morning plus a short apology for yesterday. Today: reply before lunch.”

What it does: Suggests a polite version of the reply before sending.

What a person decides: Which version to send.

For whom: Staff member.

How to measure: Complaints about tone.

Example result: “We understand waiting is frustrating — we're sorry it happened. We can book you for tomorrow at 10:00 with no wait: does that time work? If another day is better, let us know and we'll find one.”

What it does: Finds frequent unanswered questions and proposes new answers for the knowledge base.

What a person decides: Approving the answer.

For whom: Manager.

How to measure: Share of questions without a ready answer.

Example result: “Action: add a ready answer about installments — asked 14 times this week. Text: ‘Yes, Kaspi installments for 3 or 6 months with no markup, arranged at checkout.’ Impact: one answer for everyone. Risk: the terms come from a front-desk reply — confirm with the accountant. Fallback: remove the answer.”

What it does: Spots threats, VIP customers and conflicts and alerts the manager right away.

What a person decides: Next steps.

For whom: Manager.

How to measure: Manager response time.

Example result: “Category: threat of a lawsuit, regulator or public exposure. Why: in request #512 the customer writes ‘I'll sue and post a video if you don't refund me by Friday’; the request is flagged as a key customer, and it's his third message today. Route to: the manager now; reply to the customer after the manager decides.”

What it does: Messages to a new customer after purchase: what happens next, first steps, and where to send questions.

What a person decides: Special terms for the customer.

For whom: Customer.

How to measure: New customer questions; first-month refunds.

What it does: Takeaways from customer feedback: what people mention most and what to improve.

What a person decides: Changes to operations.

For whom: Manager.

How to measure: Repeat complaints.

What it does: A message to customers with holiday hours and a greeting — based on the schedule in company records.

What a person decides: Sending and the schedule.

For whom: Customer.

How to measure: Customer questions about holiday hours.

What it does: A clear message to everyone affected: what changed, from what date, and what the customer should do.

What a person decides: Sending.

For whom: Customer.

How to measure: Customer calls after changes.

Example result: “Good afternoon! From November 3 to 14, the entrance to our office is through the courtyard — the main entrance is closed for facade repairs. Parking at the main entrance is closed on those days; the nearest is at the shopping center across the street. Our hours stay the same. If you can’t find the entrance, give us a call and we’ll meet you.”

What it does: Sorts inbound offers: spam, worth a reply, or pass to the manager — with a short draft reply.

What a person decides: Reply to the partner and any agreements.

For whom: Manager.

How to measure: Manager time on inbound offers.

Example result: “Category: useful offer — reply. Why: Nurly business center is offering a discount on its event hall, and we have an open house planned for November. Next step: ask for terms and open dates in November. The other two emails are spam: website promotion and a ‘guaranteed loan.’”