Let's talk
Back to the syllabus

Episode 1 of 8 · 5 min read

What an agent is and what it is not

An agent takes in information, decides the next step within limits and acts in your systems. That is how you tell it from a chatbot or a rules robot.

Context

Almost everything is called an agent these days. The assistant on a website, the robot that copies invoices, the program that sends emails on its own. When a vendor says they sell agents, find out what is behind the label, because each tool solves a different problem.

Three things get mixed up. A chatbot converses. Classic automation, known as RPA (robotic process automation), repeats fixed steps. An AI agent does a job with some judgment. All three are useful, each for something different.

The core idea

A chatbot, in its basic form, answers. You ask it something, it replies, and that is the end of its part. It can explain or draft very well, but what happens with that answer is up to you: you copy, paste, decide and send.

Classic automation follows a script. If an email arrives with the word "invoice", it files it in a certain folder and copies the amount into a certain cell. It is fast and never tires. But it depends on everything arriving as expected. If the vendor changes the format of their PDF, the script stops or, worse, keeps running on badly read data.

An agent works in a cycle with three parts:

  1. It takes in information. It reads what arrives: an email, a file, a message, a query to your accounting system. The format need not be perfect, because the agent understands natural language.
  2. It decides the next step. Using what it read and the instructions you gave it, it chooses what comes next. It asks for a missing piece of data, checks another system, or stops and tells you. The limits are yours.
  3. It acts in your systems. It creates a record, leaves a draft ready, reconciles a line item. Then it checks the result and repeats until it finishes or must ask for help.

The difference is in the second part. The script does not decide, it only follows. The basic chatbot decides what to say and does not act in your systems. The agent decides and acts, and that is why it needs permissions, limits and supervision. Some products mix both, so ask what they do with real permissions.

How it is done

Choosing well starts with the work, not the technology. Ask the task three questions:

  • Do the inputs always arrive the same way, or do they change from case to case. A report with the same format every week calls for a script. Emails written by different people call for something that understands language.
  • Does the work end in an answer or in an action. If all you need is a text or an explanation, a chatbot is enough. If the result has to be recorded in a system, you are talking about something else.
  • What happens if it gets it wrong. Anything that makes decisions can fail. Define what the agent may do on its own and what needs your approval before it runs.

From there comes a one-line rule: if the work changes from case to case and ends with an action in your systems, you need an agent, and if not, something simpler will do.

And "simpler" is good news. Moving the fixed payments of the 15th into a control sheet is work for a script. Drafting an internal announcement is work for a chatbot. An agent there adds cost and complexity and gains nothing.

When you do need one, you treat it like any new colleague. You give it access only to what it needs, you explain how the work is done, and you review what it delivers, especially at the start. Agents make mistakes. A misread number or a misunderstood instruction will happen, so the work gets reviewed.

An illustrative case

Imagine a distributor of building materials. This is not a real client, just a scenario to see the cycle. Its purchasing manager, whom we will call Ana, gets supplier order confirmations every morning. Some arrive as PDFs, others in the body of an email, and one comes as a photo of a printed sheet.

A script will not work here, because the formats keep changing. A chatbot will not either, because the problem is not drafting, it is recording. An agent fits like this:

  1. It receives the email and reads the confirmation, however it comes.
  2. It looks up the purchase order in the system and compares product, quantity and price. If everything matches, it decides to record the confirmation.
  3. If the price differs, it decides to touch nothing. It leaves Ana a summary with the difference and the original email.

With numbers invented for the example, say 40 confirmations arrive each day and 35 match without trouble. The other 5 reach Ana already compared, with the figure that does not add up flagged.

Look at what changes in Ana's job. Before, she spent the morning typing and comparing. Now she reviews exceptions and uses that time for what only she can do: calling the supplier, negotiating a price, deciding whether it makes sense to switch brands. She also checks a sample of what the agent recorded on its own, because sometimes it will misread a number, and she needs to find out early. The agent multiplies how much Ana can cover. The decisions remain hers.

Takeaways

  • A basic chatbot answers, classic automation follows a fixed script, and an agent takes in information, decides the next step within limits and acts in your systems.
  • The one-line rule: if the work changes from case to case and ends with an action in your systems, you need an agent. If not, something simpler will do.
  • Agents make mistakes. You set limits for them, you review their work, and the important decisions stay in a person's hands.
  • Before buying anything, describe a real task from your office and answer the three questions in this text. At Omni Agent that review is part of Discovery, the first phase of any project, and sometimes the conclusion is that a script is enough.