What is an AI agent? A plain-English guide

You will hear the term 'AI agent' a great deal at the moment, often used loosely. This guide explains what it means in plain English, what an agent is made of, and how to think about where it belongs in your organisation.

From answering to acting

Most people's first experience of modern AI is a chat assistant. You ask a question and it gives an answer. It is helpful, but it only responds; it does not do anything on your behalf.

An AI agent goes a step further. You give it a goal, and it works out the steps needed to reach that goal, carries them out using the tools it has been given, checks the results, and adjusts. Put simply, a chat assistant answers; an agent acts.

Imagine asking a capable new colleague to 'prepare the monthly supplier review'. They would know to pull last month's invoices, compare them against contracts, flag anything unusual and draft a summary. An agent is designed to work in a similar way, within limits you set.

The four parts of an agent

Under the surface, most agents share four ingredients.

• A goal. This is the outcome you want, such as 'reconcile these two reports' or 'draft a response to this customer complaint'. A clear, narrow goal produces far better results than a vague one.
• Tools. These are the things the agent is allowed to use: searching a document store, reading a database, running a calculation, filling in a form. An agent can only act through the tools it has been given.
• Memory. This is the information the agent keeps as it works, such as what it has already checked or what a previous case looked like. Some memory lasts for one task; some is kept so it can learn your preferences over time.
• Autonomy. This is how much the agent may do before a person steps in. It is a choice you make, not a fixed feature of the technology.

Levels of autonomy

Autonomy is the part leaders should think about most carefully, because it is where risk lives. It helps to picture a ladder.

• Suggest. The agent prepares information or a draft, and a person does everything else.
• Prepare and propose. The agent does the full piece of work and presents a recommendation with its reasoning. A person approves, edits or rejects it.
• Act within limits. The agent completes routine, low-risk actions on its own, such as tagging a document, and passes anything outside agreed limits to a person.
• Act and report. The agent acts and a person reviews a log afterwards. This suits only the most routine, easily reversed tasks.

Most valuable enterprise work sits on the second rung. The agent takes on the time-consuming preparation, and a person makes the call.

A chat assistant answers. An agent acts. The question is how far you let it act before a person decides.

Where agents help

Agents are at their best on work that is repetitive, rule-informed and spread across several systems. Good examples include gathering documents for a review, checking records against a policy, drafting standard correspondence, or triaging a queue so that the most urgent items reach a person first.

They are also good at the tedious parts of skilled work. An analyst may spend much of their day collecting information before they can apply their expertise. An agent can do the collecting so the analyst can do the thinking.

Where a person must approve

Some decisions should always come back to a person, whatever the technology can do. As a rule of thumb, keep a human approval step wherever an action:

• moves money, changes a customer's account or commits the organisation to something;
• carries regulatory, legal or safety consequences;
• cannot easily be undone;
• involves a situation the agent has not seen before.

A well-built agent makes approval easy. It shows what it did, which sources it used and why it reached its recommendation, so the reviewer can decide quickly and with confidence.

How we build agents at Sort AI

We build agents from first principles around a specific workflow, using your real data, in your own cloud or on-premise environment. Each agent gets only the tools it needs, and anything that matters goes to a person for approval. We prove it on one queue, judged by your experts, before it goes any further.