Samokat: AI handles 80% of customer requests
80%
Of requests automated
x2.5
Saved per request
A Russian groceries-and-home-goods delivery service. Daily volume is over 660,900 orders. Coverage spans 50 Russian cities. Per month, agents handle requests from 370,000+ unique users.
About the project
Our partnership with this groceries-and-home-goods delivery service began in 2021. The original goal was to automate customer communication, cut customer-care costs, and lift agent engagement. Within just 3 months Samokat hit its target outcome and recouped its investment by 600%.
Starting point
Before working with us, Samokat routed every message to an agent. Operators handled repetitive requests by hand, which drove up chat response time, dragged CSAT down, and burned agents out on routine.
As the customer base grew, so did agent load — agents stopped keeping up and lost motivation.
Samokat didn't want to inflate headcount: customer-care payroll was already a significant cost line.
At peak times, wait times got worse — frustrating customers and pulling CSAT down further.
With no topic routing, every agent had to first identify what the request was about — at high volume, that slowed everyone down.
Goals
Take routine off the agents and lift their engagement.
Cut customer-care operational cost.
Automate text-channel request handling.
Improve service quality: shorten wait times and route by topic.
Solution
At the start of the partnership, Samokat had no automation in their text channels — every request hit a human agent.
We analyzed Samokat's requests and business processes to surface the most common questions agents handled.
Built a topic taxonomy of user requests, accounting for every way the same question gets asked.
Implemented conversation scenarios in Lia's logic — Lia ran dialogues structurally and sequentially, answering like a manager would. Many customers didn't even realize they were chatting with a bot.
Connected Lia to the Telegram chat. Over 80% of dialogues were automated; 20% of those closed with no human involvement.
Added Lia to the Infobip omnichannel solution Samokat already used. Managers could now see the conversation history and route by topic, switching the dialogue between agents.
Trained Lia to close the highest-frequency questions on her own: where's my order, minimum order, hours, jobs, app help.
Set up agent assist: agents now received bot conversations with all the necessary context already attached, split into queues — cutting handle time. When needed, Lia transferred the conversation to a human, with the entire customer history visible to the specialist on Lia's platform.
Since Lia's launch, our specialists analyze the conversations where the bot didn't handle the request and continuously fine-tune Lia.
Request and process analysis
To train the AI, we ingested 100,000+ phrasings and conversations from the entire history of Samokat's agent interactions over the prior several months. After clustering, we identified the 25 most frequently asked questions — and started automation with those.
Building the project on top topics
The Lia team built the project in stages:
Analyzed answers for the most common topics
Lia's analysts split every phrasing into topics — intents — so the AI could later match them to specific customer requests.
Implemented request handling in Lia's logic
So Lia could handle requests the way a human would, the team studied the agent scripts and mirrored them in the bot. They built the trigger-and-response logic and wrote the bot copy from a structured spec.
Trained Lia to recognize text with errors and typos
Every topic was loaded with the variations of how customers actually ask it. Early in a project, Lia typically supports a single intent with around 20–50 phrasings — including grammatical errors and typos.
Variations covered by the "Courier got lost" intent
Across a project's lifetime, intents grow to 100+ phrasings each, raising recognition accuracy.
Chat-platform integration
For convenience, Lia is integrated into the Infobip omnichannel chat platform. With it, managers see the conversation history, route by topic, and switch the dialogue between agents.
One month after launch, Lia identified and automatically recognized roughly 64,500 messages — 59.7% of all user messages. Coverage (sessions without an agent + sessions with scripted handoff to an agent) was 51.2%.
Fine-tuning existing topics and labeling new ones
After Lia went live, our specialists analyze the conversations where the bot didn't manage to handle the request.
Based on those, the team trains the bot further.
For example: Lia originally didn't answer questions about time-limited promotions. Today she can walk a customer through the "Summer Marathon of Prizes" terms in detail — and if more questions remain, transfer to an agent.
Through ongoing tuning, Lia gets smarter; the next time the same request comes in, she can handle it on her own without a human agent.
After Lia went live, Samokat's leadership measured the unit economics and the cost of an agent-handled request by topic. ROI on automation reached 500–600%.
Lia has been saving customer-care time and accelerating customer interactions for 3.5 years now — freeing managers to work on higher-value tasks.
How requests are handled today
Stage 1. Answers to the most frequent questions are surfaced in a widget on Samokat's website.
Tapping a menu item shows the answer to that question.
Stage 2. If the customer didn't find the answer in stage 1, they can switch to the bot in Telegram or contact support by phone.
In the messenger, customers can pick a request from the menu or just type their question in plain text. Either way, the bot offers the right resolution path.
The Telegram bot's button menu collects the most common topics, simplifying the customer's path to a resolution. Lia answers the question triggered by a specific button and continues a live conversation on the chosen topic.
Lia's responses in Samokat's support bot — handling typos and parsing phone numbers
Stage 3. When needed, Lia transfers the conversation to a human agent — either at the customer's request or when she can't recognize any of the known intents in the message.
Lia hands off to an agent on customer request
That happens in 20% of cases. The other 80% of customer requests are fully or partially closed by the bots.
Simple solutions
As with many of our customers, Samokat's optimization-and-automation project was delivered turnkey by Lia's team. We owned the business-process analysis, the AI bot development, integration, and configuration from scratch.
"Highlights of working with the Lia team: smooth and fast collaboration, a usable self-serve interface, ongoing bot training, and proactive scenario suggestions from their team."
Samokat's product-team auditors, alongside an analyst who writes the bot scripts, monitor the customer-service level via the Lia bot and regularly check its quality.
After the AI bot launched, Lia's specialists keep monitoring the system and ship updates on customer request, so Lia keeps getting more effective. For instance, we update Samokat's promo info inside Lia's relevant scenario as soon as it changes — and notify customers about it.
If you don't want to fall behind market leaders, switch to AI-powered chat work now. Lia's typical implementation in mid-market and SMB takes about 3 days; the investment recoups within the first 2–3 months.
For a free consultation with our team, leave a request in the form below. Our manager will follow up on a video call at your convenience and:
identify bottlenecks in your sales operation;
walk through which processes Lia can optimize;
break down how Lia helps you grow margin;
explain pricing tiers and implementation options for your business.