How to Create AI Agents for Business Workflows with Granite 4.2

Introduction: Why AI agents are essential for business workflows today

Modern businesses face constant pressure to be faster, more accurate, and more responsive. AI agents offer a practical solution to this challenge, combining large language models with specific purposes to automate end-to-end tasks. In a world where giants like Amazon are expanding delivery services with drones (Prime Air) and open-source models are evolving towards native reasoning, designing AI agents for business workflows is no longer a luxury but a strategic necessity.

Why choose an AI agent over a simple script?

• Dynamic decision-making:Unlike static rules, an AI agent can evaluate diverse contexts, adapt its behavior, and handle exceptions in real time.
• Multichannel collaboration:AI agents can read emails, Slack messages, ERP data, and external systems, unifying a company’s digital touchpoints.
• Continuous scalability:With models like IBM Granite 4.2, the same agent can be updated or scaled without rewriting entire workflows.

The result is automation that preserves human control, improves compliance, and reduces operational burden.

How to choose the right model: IBM Granite 4.2 and Liquid AI Pipette

Granite 4.2: Native reasoning and open-source sizes

IBM has released Granite 4.2 in three sizes (3B, 8B, 30B), all under the Apache 2.0 license. Each model features a "think-and-act" capability, making it ideal for agents that need to analyze complex logic, generate explanatory steps, and interact with enterprise systems without costly input tokens.

Liquid AI’s Pipette: Testing agents on-device

Liquid AI engineers have open-sourced Pipette, a benchmarking suite that evaluates models across device, quantization, runtime, and hardware. Using Pipette, teams can measure agent performance on mobile devices before deploying them at scale, ensuring a smooth user experience even with limited connectivity.

Designing effective prompts for enterprise agents

A well-structured prompt is the engine that transforms an LLM into a useful agent. Follow these steps:

• Define the purpose:What is the agent’s ultimate goal? (e.g., invoice approval, resource monitoring)
• Specify context:Include policies, data hierarchies, and system constraints.
• Set output format:Use JSON or structured markdown for reliable parsing.
• Add a feedback loop:Allow the agent to request clarification or escalate issues.

Example prompt: Invoice approval

You are an invoice approval agent for a large company. Your task is to examine each incoming invoice, compare it to approved spending limits, and decide: APPROVE, REJECT, or ESCALATE.

Provide the decision and a brief justification.

Here are the invoice details:

{invoice_data}

Spending limits:

{spending_limits}

Policy:

• Invoices above $5,000 require managerial approval.
• Duplicate invoices must be flagged as fraudulent.
• If the vendor is not in the whitelist, ESCALATE.

Output in the following JSON format:

{json_format}

Code example: LangChain agent for a purchase workflow

from langchain.agents import create_react_agent
from langchain.tools import Tool
from langchain.llms import OpenAI
import json

# Define a custom tool for querying the ERP system
def query_erp(query: str) -> str:
    # In a real environment, this would connect to a database or API
    return "{\"approved\": true, \"limit\": 10000}"

erp_tool = Tool(
    name="query_erp",
    func=query_erp,
    description="Retrieves approved spending limits for a given department"
)

# Initialize the model (replace with Granite 4.2 via a compatible API)
llm = OpenAI(model_name="granite-4-2-30b", temperature=0)

# ReAct prompt (Reasoning + Acting)
prompt = """
You are a purchase approval agent. Use the available tools to gather the necessary data, then decide APPROVE, REJECT, or ESCALATE based on the following rules:

- If the cost exceeds $5,000, managerial approval is required (ESCALATE).
- If the vendor is not in the whitelist, ESCALATE.
- If the vendor is in the whitelist and the cost is less than or equal to $5,000, APPROVE.
- Otherwise, REJECT.

Invoice data: {invoice}
Vendor information: {vendor}
Cost: {amount}

Provide the decision and justification in JSON.
"""

agent = create_react_agent(llm, [erp_tool], prompt)

# Example execution
invoice = {"id": "INV-001", "vendor": "Acme Supplies", "amount": 4200}
vendor_info = {"Acme Supplies": true}  # whitelist
result = agent.run(json.dumps(invoice))
print(result)

Real-world use cases: From Prime Air logistics to internal resource management

Amazon is expanding Prime Air to nearly 500 cities by the end of 2026. Delivery drones operate as autonomous agents, managing tracking, obstacle avoidance, and route optimization in real time. The same planning and reasoning technologies behind these systems can be adapted for:

• Resource monitoring:Agents that monitor IT equipment status and report failures.
• HR workflow management:Bots that review applications, verify certifications, and propose interviews.
• IT support:Agents that filter support tickets, gather information, and automate first-level solutions.

These examples show that design principles are not limited to one industry but are cross-cutting.

Checklist for getting started with AI agents in business workflows

• â..."" Define a clear, measurable business objective.
• â..."" Choose a model with native reasoning (e.g., Granite 4.2) that fits your stack.
• â..."" Create a structured prompt that includes policies, context, and output format.
• â..."" Implement reliable tools (database connections, APIs, file system)
• â..."" Use Pipette to benchmark performance on mobile devices before deployment.
• â..."" Implement a human feedback loop for supervision and correction.
• â..."" Monitor compliance, uptime, and user experience after launch.

Conclusion: Transform your processes with intelligent agents

AI agents are moving from experimental novelty to essential components of modern business workflows. By leveraging cutting-edge open-source models like IBM Granite 4.2, benchmarking tools like Liquid AI Pipette, and well-designed prompts, companies can automate complex processes, preserve control, and scale intelligent operations responsibly. Start with a single use case, measure with Pipette, and watch your teams focus on higher-value activities.

IMPORTANT: We are in 2026. All temporal references must be current. Do not mention past years such as 2024, 2023, etc. The content must be fresh and relevant to the present. CURRENT CONTEXT (August 2026): Here are some recent trends and news stories to inspire you: - Meta AI Introduces MetaRoCE: A Clean-Sheet RDMA Transport Built for AI-Scale Ethernet: Training and serving frontier models is now a networking problem as much as a compute problem. Collective operations like all-reduce and all-to-all sy... [2026-08-25] - XPENG IRON humanoid robot draws record physical AI funding: XPENG’s physical AI unit has secured over $900 million at a $6.3 billion valuation to scale its IRON humanoid robot platform. The Chinese electric v... [2026-08-24] - MIT AI forecasts extreme weather without historical data: MIT engineers have built an AI tool that forecasts extreme weather without training on historical disaster data. Kai Chang, a mechanical engineering g... [2026-08-25] Use this current information as inspiration to create an original and relevant 2026 prompt.