AI Assistant in Work Chat: Technical Implementation

# AI Assistant in the Work Chat: Technical Implementation and Practical Use Cases for the IT Team

A company automating clients' business processes faced a paradox: its own processes remained manual. The CEO spent hours answering repetitive questions, formulating tasks, and supporting managers. The solution came from within—implementing an AI assistant directly in the work Telegram chat. In three months, it became a full-fledged team member, handling requests, creating tasks in GitHub, and diagnosing bugs. This article covers the technical details of the architecture, real-world examples, and an honest assessment of limitations.

AI Assistant Architecture: Context Matters More Than the Model

The system isn't built on fine-tuning or custom models, but on smart context delivery via API to Claude. The key component is the Context Builder, which gathers not just the latest message, but the entire conversation thread. This lets the AI understand what's being discussed, even if the question is vague ("does this work?").

Access to internal systems is organized through the Knowledge Base:

• Codebase: Full access to current source code (PHP + Vue), allowing it to find function implementations and check validation.
• CRM database: SQL queries to client data—plans, activity, interaction history.
• GitHub: Reading existing issues and creating new ones via API.
• Internal documents: Technical specs and guides.

The architecture looks like this:

Telegram Bot API
    |
Message Handler (Python)
    | filtering: mention / reply / keywords
    |
Context Builder
    | last N messages + thread topic
    |
Knowledge Base
    +-- Project code (indexed)
    +-- Client DB (SQL queries)
    +-- GitHub Issues
    +-- Internal documents
    |
LLM (Claude) + System Prompt
    |
Response -> Telegram / GitHub API

90% of the effectiveness comes from precise context and a well-thought-out system prompt, not model complexity.

Practical Use Cases: From Support to Task Assignment

Real-Time Bug Diagnosis from Code

A manager gets a call: the client can't save a template with a link. Instead of waiting for a developer response, they post in the chat:

"Guys, the client is asking—can you create a template with a link? It's not saving for her."

The AI assistant responds in 30 seconds:

• Finds the TemplateController@store controller.
• Analyzes validation—the strip_tags() call strips HTML.
• Concludes: plain-text URLs are fine, HTML isn't.

Response:

"Checked the code. TemplateController@store takes the template text through strip_tags()—HTML links get stripped. But plain-text URLs (no <a href>) save fine. Ask the client to paste the link as plain text, without formatting."

This saves up to 30 person-hours a week previously spent on routine support queries.

Automatic GitHub Task Creation

When the CEO writes "Rom, write up a task for Dima," the AI:

• Analyzes the entire conversation thread.
• Finds the existing photo hashing code.
• Formulates specs with technical details: Hamming distance, matching thresholds, combined scoring.
• Creates an issue with correct labels and assigns the performer.

Result—a structured task the developer can jump right into, no need for tech lead clarifications.

Where AI Excels and Where It Falls Short

Strengths

• Routine question answers: Plans, client activity, object counts in the database.
• Task formulation: Turning casual requests into technical specs.
• Initial diagnosis: Narrowing bug search areas, pointing to relevant code.
• Off-hours support: Instant responses at night or on weekends.
• Client response generation: Draft emails based on conversation context.

Limitations and Risks

• Complex inter-system bugs: AI sees each service's code but doesn't always grasp runtime interactions.
• Business decisions: Data gathering is possible, but decisions remain human.
• Emotional support: Conflict situations with clients need a live manager.
• Long contexts: In discussions stretched over days, AI can lose the thread.
• Hallucinations: Confident fabrications (e.g., about a nonexistent Excel import feature) caused three incidents in three months.

To mitigate risks, the system prompt includes a rule: if uncertain, explicitly state it and recommend checking with developers.

Metrics and Strategic Implications

Over three months:

• ~2500 messages from AI in the chat
• ~120 tasks created in GitHub
• 60% of managers' questions resolved without CEO involvement
• Average response time: 15–40 seconds
• 3 critical incidents from hallucinations

Key takeaway—value isn't just in cost savings, but in speed. Clients get near-instant answers, reducing churn.

Key Points

• 80% of AI effectiveness depends on context quality, not model choice.
• Trust in AI grows nonlinearly: from formal queries to addressing it by name.
• One error reaching a client can undermine trust in the whole company.
• AI doesn't replace people—it transforms roles: CEO shifts to strategy, managers to hands-on support.
• CRM evolves from a human-managed tool to an autonomous system where humans observe and correct.

Next steps—AI integration for call analysis (STT → LLM → reports), personalized email generation, and automatic client profile creation.

— Editorial Team