How to Create AI Agents for Business Workflows with the Right Prompts

Introduction: Why AI agents are essential for modern business workflows

Today’s businesses must respond to market demands in milliseconds, manage diverse data, and reduce manual workloads. AI agents, powered by Large Language Models (LLMs) and prompt engineering, offer a way to automate these processes. They act as intelligent assistants that can understand, decide, and act within existing business tools.

Quick answer:By using well-structured prompts and leveraging foundation models like TabPFN-3.5 or frameworks like Paper2Agent, you can create agents that integrate into your systems, perform complex tasks, and improve operational efficiency.

Why choose an AI agent for your business workflow

• Interpret natural languageto understand user intent.
• Coordinate multiple actionsacross APIs, databases, and user interfaces.
• Learn from feedbackand adapt to changing processes.
• Operate 24/7without fatigue, ensuring consistency.

Current trends influencing enterprise agents

• Microsoft AI CEO criticises Anthropic
• TabPFN-3.5offers powerful predictive capabilities without additional training, making it ideal for agents handling structured data.
• Paper2Agentdemonstrates how to transform academic research into operational tools, speeding up the development of specialised agents.

How to build an AI agent for business workflows

Follow this practical, prompt-based approach to create an agent that can, for example, approve expenses, generate reports, or manage support tickets.

Step 1: Define the agent’s purpose and scope

Clarify the agent’s objective, the data it can access, and the limits of its decisions.

# Example: agent purpose
Agent: Expense approver
Purpose: Review reimbursement requests, verify eligibility, and approve or reject.
Data: Credit card statement extract (JSON), company policies (document).
Limit: Maximum reimbursement of €5,000 per transaction.

Step 2: Design the main prompt

Use a structured instruction that includes context, actions, and output format.

# Main agent prompt
Instruction: You are an expert expense approver. You will receive a reimbursement request in JSON format and company policies.
1. Calculate the total amount and check that it does not exceed €5,000.
2. Verify that the category is among the approved ones (travel, supplies, equipment).
3. If anything is out of line, respond with "Rejected: [reason]".
4. If approved, respond with "Approved: [amount]€".

Format:
{
  "status": "Approved/Rejected",
  "reason": "optional string",
  "amount": optional number
}

Input data:
{{input_data}}

Policy:
{{policy_text}}

Step 3: Integration with MCP (Model Context Protocol)

MCP allows an agent to expose tools that other systems can call. Use the Paper2Agent examples to connect the agent to an ERP database or an email service.

# Example MCP tool for reading a ledger
{
  "name": "read_ledger",
  "description": "Retrieve recent transactions from the accounting system.",
  "parameters": {
    "type": "object",
    "properties": {
      "start_date": { "type": "string", "format": "date" },
      "end_date": { "type": "string", "format": "date" }
    }
  }
}

Practical example: expense approval agent in action

Imagine a workflow where an employee uploads a screenshot of a receipt. The agent:

1. Uses MCP to extract the JSON from the ledger.
2. Applies the main prompt to verify the amount and category.
3. Responses: "Approved: €1,200" or "Rejected: Category not allowed".
4. Writes the result to the ERP system and notifies the user of the outcome.

Result: 70% reduction in processing time and consistent decision-making in line with policies.

Best practices and tips for successful implementation

• Security first:Clearly define data boundaries and implement alignment controls, as suggested by the Microsoft AI CEO.
• Use optimised foundation models:TabPFN-3.5 is ideal for tabular data processing; combine it with LLMs for text reasoning.
• Document prompts:Version main prompts in control repositories for audit trails.
• Iterative testing:Create a sandbox where the agent can operate freely, learning from false positives/negatives.
• Monitor performance:Track latency, accuracy, and impact on the surrounding system.

Conclusion: Transform your workflows with intelligent AI agents

In 2026, AI agents are no longer a technological luxury; they are a cornerstone for any business looking to scale operations. By designing clear prompts, leveraging cutting-edge foundation models, and following the safety controls emerging from the Microsoft-Anthropic debate, you can create software assistants that automate tasks, improve compliance, and free up teams for higher-value activities.

Key action:Start with a single use case (e.g., expense approval), create a structured main prompt, and connect the agent to an existing MCP tool. Measure ROI and scale the experiment to other workflows.