Tech Explained: Agentic AI

You've probably used AI to draft an email, summarise a document, or answer a question. That's useful. But it's a bit like having a very knowledgeable colleague who can only ever sit at their desk and answer questions. They can't actually do anything for you.

That's where agentic AI comes in.

Instead of just responding, agentic AI can plan, decide, and act. Autonomously, across multiple steps and using real tools. It doesn't wait to be asked. It works toward a goal.

The change is not that the model has become cleverer. It is the same kind of model underneath. What changes is that it is given an objective instead of a question, access to real tools instead of a text box, and permission to keep going until the work is done.

This is the shift that's making a lot of business leaders sit up and pay attention.

What's the difference?

Before looking at how agents work, it helps to place them next to the two kinds of automation you already have in the business. Ordinary software follows rules you wrote in advance. Conversational AI understands language but cannot touch anything. An agent sits in a third category: it understands the request and it acts on it.

The same task, invoice handling, makes the three levels easy to tell apart.

Ordinary software is reliable but brittle. It does exactly what it was told and nothing else, so the moment an invoice arrives in an unexpected format someone has to step in. Conversational AI handles that ambiguity well, because it understands language, but it stops at the answer. Every action still has to be carried out by you or your team.

Read the last two rows together and you have the whole shift. With conversational AI you are involved at every step, because the tool produces one response and then stops. With an agent you set the goal and review the outcome, and everything in between happens without you.

The building blocks of an agent

Although it might look like magic is involved, an AI agent is actually just a combination of four things working together.

None of the four is remarkable on its own. A model without tools can only describe what ought to happen. Tools without a goal have nothing to aim at. A goal without memory means the agent cannot tell what it has already tried, so it repeats itself. It is the combination that turns a language model into something that finishes work rather than commenting on it.

How an agent actually thinks

Here's what happens under the hood when you give an agent a task. The mechanism is simpler than it sounds, and knowing it makes the behaviour a lot easier to predict.

It begins by thinking: what is my goal, and what is the best next step towards it? Then it acts, using a tool, making a decision or taking an action in the real world. Then it observes: what came back, and did it work?

That observation feeds straight into the next round of thinking, so the agent decides what to do next based on what it has just learned rather than on a plan fixed in advance. The cycle repeats until the goal is complete. This is also the point where an agent differs from ordinary software, which runs the same sequence regardless of what happens along the way.

This loop, sometimes called a ReAct loop (Reason plus Act), is what gives agents the ability to handle complex, multi-step tasks that no single AI prompt could manage alone.

A real-world example

Let's make this concrete.

Notice what that one sentence set in motion. The agent searches the web for relevant industry events, then filters what it finds by date, location and relevance. It opens your company calendar to check for conflicts, and only once it has a clean shortlist does it draft the summary email. Then it stops and flags the draft for your review instead of sending it.

Three different systems, five steps, and a deliberate handover back to you at the end. What would have taken a junior team member 45 minutes takes the agent a few seconds. And your involvement was just writing one sentence and reviewing the output.

A day in the life

That example is a single task. The more interesting picture is what happens when the same principle runs on a process, every night, without anyone opening a laptop. Below is an illustration of one night on the invoice flow from the comparison at the top of this article. The figures are indicative, not measured results.

Notice where the human sits in that night. Not at every step, and not nowhere. The agent handled the volume and made the routine calls, then handed over precisely the items where a judgement was required, with the context already assembled.

That handover is a design decision rather than a happy accident. An agent is only useful in a process like this if it knows when to stop, and if the people picking up the exceptions get enough context to act in minutes instead of reconstructing what happened.

The spectrum of autonomy

Not all agents are created equal. The level of independence you give an agent is a design choice. And an important one at that, because it decides how much speed you gain and how much risk you take on.

Low autonomy, high human oversight. The agent drafts and suggests. A human approves every action before it's taken. Nothing reaches the outside world without someone signing it off. Lower risk, slower output.

Medium autonomy, human at key checkpoints. The agent works through a workflow, pausing for human approval at significant decision points rather than at every step. The sweet spot for most businesses, because it keeps the speed while preserving a place to intervene. The night above runs at this level.

High autonomy, minimal human involvement. The agent executes end to end with minimal interruption. Fastest output, highest risk. Only appropriate for well-tested, low-stakes workflows, where a mistake is cheap to spot and cheap to undo.

Where businesses are using it today

None of this is theoretical. Agents are already running in ordinary companies, usually on work that crosses several systems and that nobody enjoys doing by hand. Six areas where they are in use today:

• Communications. Drafting, sorting, and responding to emails based on priorities you define.
• Research. Gathering, analysing, and summarising competitive intelligence, market data, or customer feedback.
• Scheduling. Managing calendars, booking meetings, and coordinating across teams.
• Development. Writing, testing, and debugging code (increasingly used by engineering teams to accelerate delivery).
• Reporting. Pulling data from multiple sources, building reports, and surfacing the numbers that matter.
• Customer support. Handling routine queries, escalating complex issues, and maintaining context across conversations.

What these have in common is that the work is repetitive but not identical each time, and that it sits between systems rather than inside any single one. That is exactly the space where an agent has something to offer that ordinary software does not.

What it isn't (yet)

It's worth being clear about the current limits, because the gap between what an agent can demonstrate and what it can be trusted with day after day is still real.

The split is consistent. Agents are strongest where the work is defined in advance, repeats often and rewards consistency. People are still needed where the work calls for judgement with nuance, for reading a relationship, or for deciding what should happen in a situation nobody has met before.

Agentic AI is a powerful amplifier of human work. The businesses seeing the most value aren't replacing their teams, they're freeing them up to focus on the things that actually require a human.

Your questions, answered

Five things that come up in almost every conversation we have about this.

The one thing to remember

It's a fundamentally different kind of technology, and it deserves to be treated as one. A system that only produces text can be wrong and cost you a moment's reading. A system that takes actions can be wrong and cost you something that has already happened.

Is your business ready?

The entry point doesn't need to be dramatic. A single, well-defined workflow with clear inputs, clear outputs, and low stakes if something goes wrong, is enough to start learning.

Pick a process that happens often enough to matter, touches more than one system, and currently takes time from someone whose time is worth more elsewhere. Then agree what a good result looks like before anything gets built, because that description is what you will be reviewing against.

The goal is to build the knowledge, the processes, and the confidence to expand from there.