The word “agent” is now stuck on almost every AI product, from a website chat box to a full back-office system. That makes it hard for a business owner to know what they are being sold. This article gives a working definition you can use in a meeting, explains what separates an agent from ordinary software, and sets out when you are better off without one.
A working definition
An AI agent is software that is given a goal, looks at the situation, decides on a next step, takes that step using tools it has been allowed to use, checks the result and repeats until the goal is met or it has to stop. The key words are decides and tools.
A normal program follows a path someone wrote in advance. If the invoice has a PO number, match it; if not, send it to the accounts clerk. Every branch is fixed. An agent is given the goal (“get this invoice matched or explain why it cannot be”) and a set of tools (search the purchase orders, read the supplier’s email, look up the delivery note). It chooses which tool to use and in what order, based on what it finds.
The language model inside the agent is the part that reads, reasons and chooses. The tools are the part that acts. Without tools, a model can only talk. Without a model, the tools are just a normal program.
What makes something an agent, and what does not
Three things together make a system an agent:
- A goal, not a script. It is told what outcome to reach, not every step to take.
- Access to tools. It can read from or write to real systems: a stock sheet, a mailbox, an order database, a calendar.
- A loop. It looks at the result of each step and decides what to do next, rather than answering once and stopping.
A chat box on your website that answers questions from a list of FAQs is not an agent. It answers once and waits. A tool that rewrites a product description when you paste one in is not an agent either. It is a useful single task. Neither is worse for that. They are simply different things, and they cost less to build and run.
A system that receives a WhatsApp order, checks stock, works out the courier charge for the customer’s district, drafts a confirmation, and waits for a staff member to approve it before sending, is an agent. It moved through several steps and chose its path based on what it found.
When you do not need an agent
Agents are more expensive to build, test and run than fixed software, and they can make mistakes a fixed program never would. So the honest question is whether the extra flexibility pays for itself. Three signs it does not:
The steps never change
If you can write the whole process on one page as “if this, then that” and it covers nearly every case, write it as ordinary software or a spreadsheet rule. Sending a payment reminder three days before a due date is a scheduled job, not a reasoning problem.
The volume is tiny
If a task happens five times a week and takes ten minutes each time, that is under an hour a week. Building, testing and watching an agent for that job will cost more than the time it saves for a long while. Keep the person doing it.
Every mistake is expensive
If one wrong action means a lost customer, a legal problem or money gone, an agent should not be acting alone there. It may still help by preparing the work for a person to check. But if the checking takes as long as doing the work, the agent adds cost without saving time.
When an agent does earn its place
Agents work best where the input is messy and varied but the goal is clear. Customer messages that arrive in English, Sinhala and Tamil, sometimes typed in Roman letters, sometimes as a photo of a handwritten list. Supplier emails that each use a different format. Documents that need to be checked against several sources before anyone can act on them.
A worked example helps. Say three staff members each spend ten hours a week reading incoming orders, checking stock, and replying to confirm. That is thirty hours a week. If an agent does the reading, checking and drafting, and each person spends two hours approving drafts, the team gets back around twenty-four hours a week. Whether that is worth it depends on what the agent costs to build and run, which is a sum you can do before you start. Our automation payback calculator walks through it.
What an agent does not do well: work that depends on relationships, taste or judgement about people. It can draft the reply to an unhappy long-standing customer, but a person who knows that customer should decide what goes out.
A simple test before you start
Before you commission any agent, write down three things: the goal in one sentence, the systems it would need to touch, and the one action it must never take without a person approving. If you cannot fill in the first line clearly, you are not ready. If the second line is empty, you need a chatbot or a writing tool, not an agent. If the third line is hard to agree on, that is the conversation to have with your team first.
Write those three lines this week for the task that annoys your staff most. It will tell you more about whether you need an agent than any demo will.