Support is the department where quality is measured in seconds. A customer who couldn't check their order status at 11:40 p.m. is, by morning, writing reviews not about the order but about the company. And the question itself is almost always simple: where's my package, how do I change my plan, why was I charged. In our reviews we see the same picture again and again: up to 60% of inquiries are routine, the answers were written long ago, but they sit in a knowledge base that takes an operator two or three minutes to reach – and the customer never reaches it at all.
Let's look at how the platform takes over those 60%, what's left for people, and why being bilingual here is not an option but a basic requirement.
What's really happening in the inquiry queue
When we break a month of inquiries down by topic, the distribution usually looks like this: a third are statuses and balances, another quarter are routine instructions ("how do I connect," "how do I cancel," "where do I get a document"), around ten percent are payment questions that require a check against the accounting system. And only the remaining 25–30% are cases that genuinely need a person: a nonstandard situation, a conflict, a complex complaint.
The problem is that the operator handles all of this the same way. One queue. One pace. An average of 6–8 minutes per inquiry, for both an order status and the review of a disputed payment. The customer with a complex problem waits behind ten customers with simple ones.
| Type of inquiry | Share of the queue | Who closes it |
|---|---|---|
| Order status, balance, deadlines | ~35% | Agent, answer in seconds |
| Routine instructions and terms | ~25% | Agent, from the knowledge base |
| Payments with a system reconciliation | ~10% | Agent + human confirmation |
| Complaints, nonstandard cases | ~30% | Operator, with prepared context |
How the platform answers: not a script, but a knowledge base
The support agent on our platform doesn't work off a tree of buttons. It answers from the company's knowledge base: policies, plan terms, instructions, the history of similar inquiries. Before answering, it verifies the data at the source: the agent reads an order status from the accounting system rather than guessing it. We describe in detail how the connection to systems works on the integrations page.
A short scenario from our reviews. An online store, nighttime, a customer writes on WhatsApp: "Where's my order, I paid three days ago." The agent finds the order by phone number, sees that the package is at the sorting hub in Almaty, replies with the delivery date, and offers to notify the customer at handoff. Forty seconds. An operator would have joined this conversation at nine in the morning, seven hours later.
A second scenario, over the phone. A subscriber calls to clarify a charge. The agent answers by voice, pulls up the payment history, explains that it's an annual subscription to an add-on service, and offers to turn it off. If the subscriber starts to argue, the agent doesn't dig in but transfers to an operator and hands over the full context: who's calling, what the dispute is about, what's already been checked. The operator doesn't start the conversation with "please state your name."
Kazakh and Russian, with no language switch
In Kazakhstan a customer writes however is convenient for them: the question in Kazakh, the clarification in Russian, the contract number in digits mixed in with a bit of Kazakh. The platform's agent has native-level command of Kazakh and Russian and holds the language of the conversation on its own: the customer switched to Kazakh, and the agent switched with them, in the same message. For companies with customers in the regions this isn't decoration but a direct percentage of retention: in our reviews we regularly hear that up to a third of inquiries in the western and southern regions come in Kazakh, and the queue for Kazakh-speaking operators is twice as long.
Night shifts: what "we work from 9 to 6" actually costs
A night shift of two operators is on the order of 700,000–900,000 tenge a month with taxes, and there's still a queue after midnight. The alternative that most companies choose is simply not to answer at night. Then in the morning the department starts the day by digging out from under the pile: 80–120 unhandled inquiries, some customers have already written again, some have moved to another channel.
The agent removes this dilemma: it closes routine questions at any time of day, and it turns complex ones into tickets with a priority and assembled context. The morning shift gets not a pile but a sorted list: here's a complaint, here's a return, here's a customer at risk of leaving, start with them.
SLA stops being a lottery
The most honest support metric isn't "average response speed" but the share of inquiries closed on time per the SLA. Average speed looks nice in a report and says nothing about the customer who waited four hours.
Real SLA = share of inquiries closed within the target time × share of inquiries resolved on first contact. "We've forwarded your question" is not a resolution, it's deferring the wait.
With an agent on the first line the structure changes: routine inquiries get an answer in seconds at any time, and the SLA is measured only on the cases where a person is working. In our reviews we see time to first response drop from 15–40 minutes to under one minute across 60% of the queue, and the load on operators falls so much that they spend on a complex case not 8 minutes but as long as it takes.
Quality: every conversation is scored, not a 5% sample
Classic quality control has a supervisor listen to 3–5% of calls and fill out a checklist. The other 95% stay out of view. The platform reviews every conversation, both the agent's and the operator's: was the policy followed, was the question resolved, what's the customer's tone at the end of the conversation. The manager sees not a sample but the full picture: which topics the agent handles on its own, where escalations happen most often, which operator has a rising share of conflict-ending conversations.
This also settles the "what if the agent says too much" question. Every answer it gives is recorded, tied to a source in the knowledge base, and available for review. On the topics you've flagged as sensitive – money, cancellations, personal data – the agent doesn't improvise: it either answers strictly by the approved wording or hands off to a person.
How it gets launched
Getting started doesn't require a "perfect knowledge base." What you already have is enough: policies, an FAQ page, an export of closed inquiries from the past 2–3 months. We set up a prototype on your scenario in about a week, free of charge: you ask the agent the real questions your customers ask and look at the answers. Reaching production with channels and accounting systems connected takes from 8 weeks. The platform runs on a subscription starting at 12 million tenge a year; the full description of the support scenario is on the customer service solution page, and how customer data stays within the company's perimeter is covered in the security section.
Frequently asked questions
Will the agent replace operators?
No. The agent takes over routine inquiries, the part of the work that burns operators out and drives them to quit. What's left for people is complaints, nonstandard cases, and customers with whom it matters to speak in person. In practice teams don't shrink, they stop being torn between the queue and quality.
What happens if the agent doesn't know the answer?
It doesn't make things up. If the answer isn't in the knowledge base and the connected systems, the agent says so honestly, asks for details, and creates an inquiry for an operator with the full context of the conversation. The handoff threshold is configurable: you can make the agent more cautious on sensitive topics.
How long will it take to fill out the knowledge base?
Less than it seems. The agent builds the base from documents that already exist and the history of inquiries: preparation usually takes one to two weeks in parallel with setting up the prototype. Gaps show up fast: for questions without answers, the platform itself compiles a list of topics that need to be added.