Introduction: Simplifying business processes with AI agents
Today’s businesses are constantly seeking faster ways to reduce manual errors, accelerate execution times, and free up teams from repetitive tasks. AI agents are the solution: small, autonomous programs that interpret data, make decisions, and act in a coordinated manner within a business workflow.
This article explains how to design, select, and deploy intelligent agents in 2026, leveraging the latest open-source and proprietary innovations.
How AI agents work in business workflows
An AI agent consists of three main layers:
- Perception:Acquires data from emails, CRM systems, ERP platforms, IoT sensors, or external APIs.
- Reasoning:Utilizes large language models (LLMs), such as the new Qwen3.8-Flash-Next models, or specialized networks to analyze context and generate an action plan.
- Action:Executes tasks through built-in tools, including API calls, scripts, and low-code automations.
The result is an autonomous cycle that can be triggered by an event, a schedule, or human intervention.
Why choose an agent over simple automation?
- Flexibility:Agents can adapt to varied inputs thanks to prompt-based reasoning.
- Scalability:Multiple agents can collaborate to handle complex workloads.
- Natural interaction:Users communicate with agents in plain language, reducing reliance on rigid scripts.
Which agent is right for you? A comparison between open-source and proprietary solutions
The agent landscape in 2026 offers distinct options:
- Open-source agents:Models like Qwen3.8-Flash-Next (125B MoE with 6B active parameters) provide multimodal capabilities and a lightweight footprint, ideal for on-device inference. The community offers libraries such as LangChain and AutoGen for rapid agent development.
- Proprietary agents:Platforms like Microsoft Power Platform, Google Vertex AI, and Amazon Web Services’ AI services provide pre-trained agents with drag-and-drop UIs, suitable for non-technical teams.
Choose based on control, budget, and workflow complexity.
Creating an agent with prompt engineering: a practical example
Below is a minimal example of an agent that extracts and validates invoices from incoming emails. The code uses Python, LangChain, and an open-source LLM compatible with the API.
from langchain.agents import create_react_agent
from langchain.tools import Tool
from langchain.llms import OpenAI
import re
# 1. Define the invoice extraction tool
def extract_invoice(text: str) -> dict:
# Simple regex to demonstrate the concept
match = re.search(r'Factura n° (\d+) - (\d+,\d+) €', text)
if match:
return {"number": match.group(1), "amount": match.group(2)}
return {}
invoice_tool = Tool(
name="extract_invoice",
func=extract_invoice,
description="Extracts invoice number and amount from an email text."
)
# 2. Initialize the LLM (replace with Qwen3.8-Flash-Next if available via API)
llm = OpenAI(model="gpt-4o-mini", temperature=0)
# 3. Create the agent with a system prompt guide
agent = create_react_agent(
llm=llm,
tools=[invoice_tool],
prompt="""
You are an expert accounting assistant. When you receive an email text, first call the tool to extract invoice data.
If the tool returns a valid result, confirm the extraction. Otherwise, ask the user for missing details.
"""
)
# 4. Run the agent on a sample email
sample_email = "Hi, attached is Factura n° 1234 - 567,89 € for consulting services."
result = agent.run(sample_email)
print(result)
"""Quick tips for prompt engineering
- Be specific:Instruct the agent to return data in a desired format, e.g., "Return the data only as JSON."
- Use context:Include information about customer history or business rules in the prompt.
- Test with edge cases:Feed in anomalous data to verify the agent's robustness.
Integrating agents into existing systems
Most businesses already have a technology stack. Agents can be integrated through:
- APIs and webhooks:Trigger the agent when a new record is created in Salesforce, NetSuite, or a SQL database.
- Low-code platforms:Power Automate, Airtable, or Zapier offer pre-built AI agent components.
- Serverless middleware:Use AWS Lambda or Azure Functions to run the agent on real-time events.
An integrated architecture reduces duplicate data development and ensures a consistent workflow.
Real-world cases: from Gatik to internal processes
Gatik uses AI agents to coordinate autonomous trucks, dynamically adjusting routes based on traffic, weather conditions, and load capacity. The same principle applies to inventory management: an agent monitors stock levels, predicts demand spikes using LLM-based forecasts, and automatically triggers purchase orders.
Companies can replicate this model for:
- Employee onboarding:Agents that collect documents, validate them, and initiate HR processes.
- IT support:Automatic detection of issues, initiation of resolution workflows, and closure once resolved.
- Billing and accounting:As illustrated in the code example above.
How to measure the ROI of AI agents
Track concrete metrics:
- Execution time:Reduction in average processing time (e.g., from 5 days to 30 minutes).
- Error taxonomy:Count of errors corrected by the agent versus those handled manually.
- Cost per action:Compare the cost of agent infrastructure with savings on labor costs.
A simple dashboard with these KPIs will help you demonstrate value to stakeholders.
Next trends: benchmarking with Pipette and beyond
Liquid AI has open-sourced Pipette, a benchmark set that evaluates on-device models in terms of speed, quantization, and resource usage. For teams building agents, Pipette provides a standardized way to compare the performance of new models like Qwen3.8-Flash-Next on real hardware (smartphones, edge). Adopt Pipette in your tests to ensure agents maintain acceptable performance at scale.
Looking ahead, the emergence of multimodal AI agents (vision + language + reasoning) will further transform workflows involving documents, images, and voice-first interactions.
Conclusion: From idea to intelligent automation
Creating AI agents for business workflows is no longer a futuristic dream. With mature open-source models, consolidated prompt engineering libraries, and low-code tools, any company can build autonomous agents that perceive, reason, and act in a coordinated manner.
Start with a low-risk use case (like invoice extraction), measure improvements, and scale to more complex processes. The competitive advantage comes from the ability to transform business data and rules into fast, accurate, automated actions.