AI agents & business automation

A chatbot answers. An agent acts. The difference is a loop: the model receives a goal, chooses a tool, sees the result, and decides what to do next, until the goal is met or a limit is hit. That loop is what lets AI move from drafting text to actually checking an order, raising a ticket or updating a sheet.

The four parts of an agent

1. Model — decides the next step.
2. Tools — functions it may call: a database query, an API call, a calculator, a file reader.
3. Loop — run tool, feed the result back, repeat with a step limit.
4. Memory — the running conversation, plus anything stored between sessions.

Tools are just functions with a description

You describe each function so the model knows when to call it, then execute the call yourself in normal Python:

def check_order(order_id: str) -> dict:
    # replace with a real database lookup
    return {"order_id": order_id, "status": "in transit", "eta": "2 days"}

def raise_ticket(order_id: str, issue: str) -> dict:
    return {"ticket_id": "TKT-3391", "order_id": order_id, "issue": issue}

TOOLS = {"check_order": check_order, "raise_ticket": raise_ticket}

TOOL_SPECS = [
    {
        "name": "check_order",
        "description": "Get delivery status and ETA for one order id.",
        "input_schema": {
            "type": "object",
            "properties": {"order_id": {"type": "string"}},
            "required": ["order_id"],
        },
    },
]

def run_tool(name: str, args: dict) -> dict:
    if name not in TOOLS:
        return {"error": f"unknown tool {name}"}
    try:
        return TOOLS[name](**args)
    except Exception as exc:               # never let a tool crash the loop
        return {"error": str(exc)}

The agent loop then repeats: send the conversation to the model, and if it asks for a tool, run it with run_tool, append the result as a new message, and send again — capped at, say, five steps so a confused agent cannot spin forever.

Which processes are worth automating?

A useful rule for a small business: automate the retrieval and drafting, keep the sending and paying with a human. A dispatch clerk approving 60 drafted replies an hour is a far better system than an agent emailing customers unsupervised.

Guardrails that matter

• Least privilege: read-only tools by default; any write, refund or send needs explicit approval.
• Step and cost limits: cap loop iterations and tokens per task.
• Validate tool inputs: an order id should match your format before it touches the database.
• Log everything: prompt, tool calls, results and final output, for debugging and audit.
• Treat fetched text as data, not instructions — a supplier PDF must never be able to tell your agent what to do.

Key takeaways

• An agent is a model plus tools plus a bounded loop, not a smarter chatbot.
• Reliability comes from narrow tools and hard limits, not from a longer prompt.
• Measure it like a process: tasks completed without human edit, cost per task, minutes saved.