AI Agent Development for Business Process Automation

AI Agent Development for Business Process Automation

Imagine: an accountant spends 3.5 hours a day reconciling invoices, and an HR manager spends 2 hours on onboarding each new hire. These processes consume time, but they can be automated. We are a team of AI/ML engineers with 10 years of experience. We develop intelligent AI agents based on large language models (LLMs) that take over routine work. Unlike classic RPA, our agents understand unstructured content: PDFs, photos, emails. They adapt to data variations without reprogramming. We have 30+ automation projects under our belt — from invoice processing to employee onboarding.

The first thing we do on a project is analyze your business process. We identify bottlenecks, collect typical cases, and design the future agent's architecture.

Why Is an AI Agent More Effective Than RPA?

RPA works with rigid scripts: any deviation in the structure of an email or PDF breaks the process. An AI agent powered by LLM (GPT-4o, Claude 3.5) handles variability. It extracts data from arbitrary text, uses Large language model (LLM) for classification, and makes decisions based on rules. The table below shows key differences.

How We Build an AI Agent: Step-by-Step Plan

Our approach to agent development consists of four stages:

1. Analysis and design (1–2 weeks): study the business process, collect data, draw AS-IS and TO-BE diagrams.
2. Core development (2–3 weeks): implement a state graph in LangGraph, connect the LLM (GPT-4o or Claude 3.5) and a vector database (ChromaDB/pgvector).
3. System integration (1–2 weeks): configure adapters for 1C, Jira, Bitrix24, email. The table below shows typical integrations.
4. Testing and HITL (1–2 weeks): verify on real data, configure human decision points.

Typical Business Processes for an AI Agent

Handling incoming requests: the agent parses email/form and classifies the input. It validates data, routes to the appropriate executor, and creates a task in the tracker. Other processes — onboarding, invoice processing, reporting — follow the same principle.

Agent Architecture for Request Processing

from langgraph.graph import StateGraph, END
from langchain_openai import ChatOpenAI
from typing import TypedDict, Optional

class ApplicationState(TypedDict):
    raw_input: str              # Incoming request text
    applicant_name: str
    application_type: str       # request type
    extracted_data: dict        # extracted data
    validation_result: dict     # validation result
    routing_decision: str       # where to route
    task_id: Optional[str]      # created task ID
    notification_sent: bool

llm = ChatOpenAI(model="gpt-4o", temperature=0)

def classify_and_extract(state: ApplicationState) -> ApplicationState:
    """Classify request and extract data"""
    response = llm.invoke(f"""Analyze the incoming request and extract structured data.

Request:
{state['raw_input']}

Return JSON:
{{
  "application_type": "vacation|expense|equipment|access|other",
  "applicant_name": "...",
  "department": "...",
  "details": {{}},  // specific fields by type
  "urgency": "normal|urgent|critical",
  "missing_info": []  // what is missing
}}""")

    import json
    data = json.loads(response.content)
    return {
        **state,
        "application_type": data["application_type"],
        "applicant_name": data.get("applicant_name", ""),
        "extracted_data": data,
    }

def validate_application(state: ApplicationState) -> ApplicationState:
    """Check completeness and policy compliance"""
    app_type = state["application_type"]
    extracted = state["extracted_data"]

    validation = {"valid": True, "errors": [], "warnings": []}

    if app_type == "vacation":
        # Check vacation balance
        days = extracted["details"].get("days", 0)
        balance = hr_api.get_vacation_balance(state["applicant_name"])
        if days > balance:
            validation["valid"] = False
            validation["errors"].append(f"Insufficient vacation days: requested {days}, available {balance}")

    elif app_type == "expense":
        amount = extracted["details"].get("amount", 0)
        if amount > 50000:  # Self-approval limit
            validation["warnings"].append("Requires manager approval")

    return {**state, "validation_result": validation}

def route_application(state: ApplicationState) -> ApplicationState:
    """Determine processing route"""
    app_type = state["application_type"]
    validation = state["validation_result"]
    urgency = state["extracted_data"].get("urgency", "normal")

    if not validation["valid"]:
        routing = "reject_with_explanation"
    elif app_type == "vacation":
        routing = "hr_manager"
    elif app_type == "expense" and state["extracted_data"]["details"].get("amount", 0) > 50000:
        routing = "director_approval"
    elif app_type == "access":
        routing = "it_department"
    else:
        routing = "auto_approve"

    return {**state, "routing_decision": routing}

def execute_routing(state: ApplicationState) -> ApplicationState:
    """Execute routing"""
    routing = state["routing_decision"]

    if routing == "auto_approve":
        task_id = jira_api.create_task(
            title=f"Auto-approved: {state['application_type']} from {state['applicant_name']}",
            status="Done",
            assignee="system",
        )
    elif routing in ["hr_manager", "director_approval", "it_department"]:
        assignee_map = {
            "hr_manager": "[email protected]",
            "director_approval": "[email protected]",
            "it_department": "[email protected]",
        }
        task_id = jira_api.create_task(
            title=f"Request for {state['application_type']} from {state['applicant_name']}",
            assignee=assignee_map[routing],
            description=json.dumps(state["extracted_data"], ensure_ascii=False),
            priority="High" if state["extracted_data"].get("urgency") == "urgent" else "Normal",
        )
    else:
        task_id = None

    notification_service.send(
        to=state["applicant_name"],
        message=f"Your request has been accepted. Routing: {routing}. ID: {task_id}"
    )

    return {**state, "task_id": task_id, "notification_sent": True}

# Build the process graph
from langgraph.checkpoint.memory import MemorySaver

graph = StateGraph(ApplicationState)
graph.add_node("classify_and_extract", classify_and_extract)
graph.add_node("validate", validate_application)
graph.add_node("route", route_application)
graph.add_node("execute", execute_routing)

graph.set_entry_point("classify_and_extract")
graph.add_edge("classify_and_extract", "validate")
graph.add_edge("validate", "route")
graph.add_edge("route", "execute")
graph.add_edge("execute", END)

application_agent = graph.compile()

Practical Case: Incoming Invoice Processing

Task: 180+ invoices for payment per month. Before automation, the chief accountant spent 3.5 hours per day. From our client's practice: we implemented the agent in 6 weeks.

Agent pipeline:

1. Extract text from PDF (pdfplumber / LlamaParse)
2. LLM extracts: supplier, TIN, amount, VAT, date, number, contract
3. Cross-check with contract register (vector search)
4. Verify in 1C: contract balance, budget line item
5. If OK – create a payment order in 1C
6. If discrepancy – assign a task to the accountant with an explanation

Metrics after 3 months:

• Automatically processed without intervention: 73%
• Accuracy of details extraction: 96%
• Errors (incorrect contract link): 1.2%
• Time savings: 2.5 hours/day

How Does Integration with Corporate Systems Happen?

We connect the agent to your systems via REST API, ODBC, or file exchange. Typical integrations: 1C, Jira, Bitrix24, mail servers. For each system, we build an adapter that translates data into a format the agent understands. A vector database (ChromaDB, pgvector) stores reference documents for RAG retrieval.

Human-in-the-Loop: When the Agent Requests Confirmation

def requires_human_approval(state: ApplicationState) -> bool:
    """Determines if human intervention is needed"""
    return (
        not state["validation_result"]["valid"] or
        state["extracted_data"].get("amount", 0) > 100000 or
        state["application_type"] == "termination" or
        state["extracted_data"].get("urgency") == "critical"
    )

# In LangGraph: interrupt_before for HITL
agent = graph.compile(
    interrupt_before=["execute"],  # Interrupt before execution
    checkpointer=MemorySaver(),
)

What Is Included in the Work

• Business process analysis (AS-IS and TO-BE diagrams)
• Agent architecture design (LangGraph, LLM, vector DB)
• Development and integration with your systems
• Testing with human-in-the-loop
• Documentation (model card, instructions)
• Employee training (2 hours)
• 3-month warranty and support

Indicative Timelines

• Analysis and design: 1–2 weeks
• Development with integrations: 3–5 weeks
• Testing and HITL setup: 1–2 weeks
• Total: 5–9 weeks depending on complexity

Order AI agent implementation for your business — contact us and we will propose architecture and timelines. Get a consultation and cost estimate.

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.