Frontline support automation for a fintech company: reducing operator workload by up to 70% with an AI agent

−70%

L1 Costs

savings

seconds

Response Time

instead of minutes

75%

Auto-Resolution

no operator needed

Request Flow

Customer writes

Freshdesk

Creates ticket

AI Agent

Responds (75%)

Operator

Complex cases

AI handles routine requests, leaving operators to focus on expert cases.

Challenge

The Linear Growth Trap

More users — more tickets. A "Groundhog Day" of identical questions consuming team resources.

Growing customer base increased the load on frontline support.
Up to 60–70% of requests were repetitive and handled manually, reducing team efficiency and distracting operators from truly complex cases.
Response quality and speed depended on individual employees, while the existing knowledge base was used inefficiently due to lack of convenient access.
Support load was unpredictable, and scaling the service required linear growth of the team and operational costs.

Why Not a Chatbot

Why a Simple Chatbot Doesn't Work in Fintech

Cost of Errors

An inaccurate answer about fees or limits means losses for the client and legal risks for the company.

Hallucinations

GPT is too "creative" where dry facts and exact numbers from documentation are needed.

Security

Sending customer data to public neural networks is unacceptable. Fintech requires isolation and control.

Solution

The Solution

We implemented an AI agent for frontline support automation, integrated into the company's existing infrastructure. Customers continue reaching out via Telegram, and the support team keeps working in Freshdesk — no process changes required.

The AI agent processes requests registered in Freshdesk, analyzes their content, and resolves a significant portion without operator involvement. Only cases requiring individual analysis and expert intervention reach the support team.

The agent manages ticket statuses, clearly understands its responsibility boundaries, and correctly escalates requests when human involvement is needed. When transferring a ticket to an operator, the agent preserves all collected context, allowing work to continue without information loss.

The agent works like an ultra-experienced librarian: understands the customer's question, instantly finds the right information in the company's knowledge base, and forms an accurate answer strictly following instructions.

Technically, this is a Retrieval-Augmented Generation (RAG) approach — the agent uses only verified company knowledge, not "fantasizing" based on general internet data.

This enables:

consistent quality and relevance of responses,
knowledge updates without model retraining,
maintaining control and compliance with fintech domain requirements.

The solution architecture was designed with a security-first principle: sensitive user data is not passed to the AI model context, and access to information and settings is strictly controlled.

"
An interesting UX insight: users don't realize they're communicating with an algorithm. At the client's request, we don't emphasize the use of AI to preserve the familiar user experience and trust in the support channel.

UX & Management

During the pilot launch, it became clear that users perceive the dialogue with the agent as communication with a support operator.

A separate admin panel is used for system management. The support team can update and maintain the knowledge base, as well as manage the AI agent's instructions and behavior.

This allows quickly setting temporary response scenarios — for example, during incidents or emergencies — and adjusting agent logic without code changes or model retraining.

Architecture

Centralized Core & Scaling

At the core is a centralized hub that connects Telegram support with Freshdesk and enables flexible request flow management within a unified process.

The architecture supports request segmentation by user categories, request types, and service levels.

For specific segments — such as VIP clients or critical scenarios — dedicated AI agents with their own processing and escalation rules can be deployed.

This approach enables scaling support without infrastructure complexity, reducing operator load, and maintaining system manageability as users and scenarios grow.

Results

The Result

The AI agent independently resolves about 75% of requests: users get their questions answered without operator involvement and are satisfied with support quality.

In cases requiring escalation, the agent collects all necessary information and transfers the request to an operator with full context, reducing resolution time and eliminating repeat questions.

Additionally:

Reduced L1 support workload and operational costs

First response time dropped from minutes to seconds

System ready for traffic growth without hiring new operators

Tech Stack

Integrations

Telegram Freshdesk RAG AI-agent Knowledge Base Deepseek API

Scalability without compromising response quality.

Reliable request routing.

Response quality control and transparent metrics.

Let's Work Together

Available for freelance projects, long-term contracts, and technical consulting. Currently based in Vietnam (UTC+7).

Write an Email LinkedIn