Home / Case studies / TermoLand
TermoLandThermal spa network · Russia
Case study CRM-native AI manager

“Anna” — the AI manager that became part of the call-center team

TermoLand stopped losing after-hours enquiries and reduced pressure on operators. Anna now handles 17,000 messages a month inside the CRM, follows leads through the funnel, and escalates only when human judgment is needed.

Discuss a similar system
Operational context

The enquiry volume had outgrown a fully manual call center

Every peak and shift change created another opportunity to lose a high-intent customer.

Before
100% of conversations depended on live operators, while turnover consumed training capacity.
After-hours enquiries waited until morning, when some customers had already booked elsewhere.
Peak response time stretched to several hours.
Context was lost between shifts and communication channels.
After
Anna processes 17,000 messages a month inside the same CRM used by operators.
Customers receive a response within two minutes, 24/7.
Answer accuracy reaches 99%, with uncertain cases routed to people.
AI runs follow-up on 80% of deferred deals; operators focus on closing.
The solution

We embedded an AI colleague, not a website widget

Anna works in TermoLand’s CRM on the same conversation timeline as the human team. That single design decision makes hand-offs seamless and keeps AI accountable to the operating process.

CRM-native operation

AI and human messages live in one system. When Anna escalates, the operator sees the full customer history.

Multi-model architecture

Smaller models classify and route requests; larger models handle customer-facing answers, balancing speed, cost, and quality.

Funnel follow-up

Anna does more than answer questions: she reminds, motivates, and moves the lead toward a purchase decision.

Honest escalation

When confidence is low, AI brings in a manager instead of inventing an answer and putting the brand at risk.

Quality-control loop

Operators intercept questionable answers using explicit rules, and real cases continuously improve the system.

Natural-language service

Customers write naturally. There are no rigid menus and no “press 1 to continue” interaction model.

Client perspective

What changed for the operating team

“We used to drown in call-center turnover and missed enquiries — and lose customers. Now Anna, our AI bot, handles it quickly, accurately, and without fatigue.”

Alexander Polkovnikov · CPO, TermoLand
Measured results

Six months of production operation

Figures taken from TermoLand’s CRM after the AI manager went live.

17K
Messages per month
handled by AI
99%
Answer accuracy
errors below 1%
2 min
Response time, 24/7
previously: hours at peak
80%
Deferred deals
followed up by AI
AI manager Anna follows up with a customer

AI-led follow-up and purchase motivation in the live workflow.

Human operator and AI response comparison

Response comparison during peak call-center hours.

Implementation lessons

What makes customer-facing AI reliable in production

01

Design for errors, not perfection

At volume, every model will miss something. Reliability comes from interception rules and a human quality loop, not a mythical perfect prompt.

02

Use different models for different jobs

Classification and routing do not need the same model as customer-facing reasoning. Matching model cost to task protects both economics and quality.

03

Make uncertainty visible

A clean hand-off is cheaper than a confident but incorrect answer. Escalation is a feature, not a failure.

04

Put AI inside the team’s system of record

When the agent works inside the CRM, governance, context, and ownership stay intact.

Roadmap

Anna 2.0 — from service agent to revenue operator

The next stage gives Anna the ability to initiate payment processes directly in the CRM. That moves the agent from answering and follow-up to completing the commercial workflow, while the call center retains control over non-standard and sensitive cases.

Could your call center benefit from a 24/7 AI colleague?

We will map the workflow, identify safe automation boundaries, and design the agent around your CRM and operating rules.

Discuss your use case