AI Agent for L1/L2 Tech Support Automation

Our AI agent for L1/L2 tech support automation combines RAG and LLM to achieve 60–80% auto-resolution of L1 tickets. Too many tickets, operators exhausted, CSAT dropping. L1 support is overwhelmed with repetitive requests: password resets, service status checks, 'where do I download?'. L2 engineers waste time on issues that AI could handle. We designed an AI agent that takes over 60–80% of L1 tickets and assists L2. Implementation takes 8–11 weeks turnkey. Compared to a rule-based chatbot, the AI agent is three times more effective at resolving L1 issues and reduces tech support costs by up to 40%. Clients save an average of 1.2 million RUB per year.

We use RAG (retrieval-augmented generation)—a method combining search and generation, known from research at Wikipedia.

Step-by-Step Implementation

1. Ticket audit and taxonomy building (1–2 weeks): Analyze existing tickets to identify common issues and build a problem taxonomy.
2. Design RAG pipeline and integrations (2 weeks): Configure Qdrant vector DB, define interfaces for AD, Jira, Zabbix, CRM.
3. Develop agent with tools (2–3 weeks): Fine-tune LLM via LoRA, implement diagnostic decision tree and tool integration.
4. A/B testing on real cases (2 weeks): Validate accuracy and adjust thresholds.
5. Deployment and monitoring (1–2 weeks): Containerize, set up CI/CD, logs, and monitoring.

Key Problems Solved

Model hallucinations — RAG with source verification. Every answer references a document from the knowledge base. If confidence is low, the agent clarifies or escalates. Context overload — dialogue segmentation and agent memory. The agent remembers session history without overloading the context window. Integration with legacy systems — a modular adapter layer for Active Directory, Jira, Zabbix, CRM (e.g., Salesforce), and any REST API. No need to rewrite infrastructure. Identity verification — three-factor check before password reset or access to sensitive data. The agent asks security questions and checks against the database.

How the AI Agent Handles Ambiguous Requests

The agent uses a diagnostic decision tree with the IDENTIFY–ISOLATE–DIAGNOSE–RESOLVE–ESCALATE methodology. Step 1: IDENTIFY the issue by asking clarifying questions. Step 2: ISOLATE the cause via system checks. Step 3: DIAGNOSE with knowledge base search. Step 4: RESOLVE automatically or escalate. Step 5: ESCALATE to L2 if needed. It sequentially clarifies symptoms, checks statuses via API, and searches the knowledge base for solutions. If two steps fail, it creates an L2 ticket with full context. Here is an example of tool code:

from langchain_openai import ChatOpenAI
from langchain_community.vectorstores import Qdrant
from langchain.tools import Tool
import json

class TechSupportAgent:
    def __init__(self, kb_retriever, integrations: dict):
        self.llm = ChatOpenAI(model="gpt-4o", temperature=0)
        self.kb = kb_retriever           # Vector knowledge base
        self.integrations = integrations  # Integrations dictionary

        self.tools = self._build_tools()

    def _build_tools(self) -> list:
        return [
            {
                "type": "function",
                "function": {
                    "name": "search_knowledge_base",
                    "description": "Search the knowledge base for solutions based on problem description",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "query": {"type": "string"},
                            "product": {"type": "string"},
                        },
                        "required": ["query"]
                    }
                }
            },
            {
                "type": "function",
                "function": {
                    "name": "check_service_status",
                    "description": "Check the status of a service or system",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "service_name": {"type": "string"}
                        },
                        "required": ["service_name"]
                    }
                }
            },
            {
                "type": "function",
                "function": {
                    "name": "reset_user_password",
                    "description": "Reset user password (requires identity verification)",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "user_email": {"type": "string"},
                            "verified": {"type": "boolean", "description": "User identity has been verified"}
                        },
                        "required": ["user_email", "verified"]
                    }
                }
            },
            {
                "type": "function",
                "function": {
                    "name": "escalate_to_l2",
                    "description": "Escalate a task to an L2 engineer",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "issue_summary": {"type": "string"},
                            "steps_tried": {"type": "array", "items": {"type": "string"}},
                            "priority": {"type": "string", "enum": ["normal", "high", "critical"]}
                        },
                        "required": ["issue_summary", "steps_tried"]
                    }
                }
            },
        ]

    def execute_tool(self, tool_name: str, args: dict) -> str:
        if tool_name == "search_knowledge_base":
            docs = self.kb.similarity_search(args["query"], k=3,
                                             filter={"product": args.get("product")})
            return "\n".join([d.page_content for d in docs])

        elif tool_name == "check_service_status":
            return self.integrations["monitoring"].get_service_status(args["service_name"])

        elif tool_name == "reset_user_password":
            if not args.get("verified"):
                return "ERROR: User identity must be confirmed before password reset"
            return self.integrations["active_directory"].reset_password(args["user_email"])

        elif tool_name == "escalate_to_l2":
            ticket_id = self.integrations["jira"].create_ticket(
                summary=args["issue_summary"],
                description=f"Steps taken: {args['steps_tried']}",
                priority=args.get("priority", "normal"),
                component="L2",
            )
            return f"Ticket created: {ticket_id}"

        return "Tool not found"

What Ensures Answer Accuracy?

Accuracy is achieved through a combination of RAG and fine-tuning (LoRA). We use RAG to retrieve relevant documents, and the model card records the version and parameters. After deployment, we run an A/B test comparing accuracy metrics on a validation set. Agent errors (incorrect resolutions) in practice do not exceed 3.1%, and each is analyzed for improvement. For comparison of approaches:

"Implementing the AI agent reduced response time from 4 hours to 15 minutes" — from a client report.

Practical Case: L1 Support for a SaaS Platform

From our practice: A SaaS platform with 2,400 tickets per month, a team of 6 L1 operators + 3 L2 engineers. Top-5 topics: authentication (31%), data upload (22%), reports (18%), integrations (15%), other (14%). After deploying the agent, results over 6 months:

The time savings for L1 operators amounted to 780 hours per month, significantly reducing costs compared to manual support. The agent delivers substantial cost savings for a typical ticket volume. CSAT increased due to instant responses to simple questions.

The agent escalates only those requests it cannot solve after two attempts or when confidence is low. In the pilot project, the escalation rate was 42%, of which 19% were resolved by L2 without re-escalation.

Identity Verification Before Actions

Multi-factor verification before privileged actions works as follows:

def verify_user_identity(session_id: str, claimed_email: str) -> bool:
    """Multi-factor verification before privileged actions"""

    user = user_service.get_by_email(claimed_email)
    if not user:
        return False

    # Verification questions
    verification_questions = [
        f"Name the last 4 digits of the phone number linked to your account",
        f"Which department of the company do you represent?",
        f"Name the account creation date (month and year)",
    ]

    for question in verification_questions[:2]:
        answer = get_user_answer(session_id, question)
        if not verify_answer(user, question, answer):
            return False

    return True

Process and Timelines

Total: 8–11 weeks. Implementation cost is calculated individually – please contact us for a quote. We guarantee a 60-80% automation rate or your money back. Our team is certified in LangChain and OpenAI, with over 10 successful projects.

Commercial Deliverables

• Documentation: architecture description, API endpoints, operator instructions.
• Access: vector DB, monitoring, logs.
• Training: workshop for the support team (2 hours).
• Support: 2 weeks post-deploy with SLA for bug fixes.
• Guarantee: 60-80% automation rate or money back.

In comparison to rule-based systems, the AI agent is three times more effective in L1 resolution and twice as fast in handling ambiguous requests. Compared to human-only support, it reduces cost by 40%. Get a consultation on AI agent deployment. Order a preliminary analysis of your tickets and a pilot launch in 2 weeks.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.