Insight series · Part 3 · 14 September 2026 · Part 1 · Part 2 · Part 4

What is an AI agent — and when is it really an agent?

“AI agent” has become one of those phrases that seems to mean almost anything. This article sharpens the distinction Aadi cares about: chatbot, workflow, agent and digital worker — and why autonomy is a dial, not a switch.

A chatbot is called an agent. A workflow with an LLM in the middle is called an agent. Software that sends an email is called an agent. And genuinely adaptive systems that plan, use tools and make bounded decisions as they work are also called agents.

They are not all the same thing. So what actually makes an AI system agentic?

An AI agent is given an objective and some freedom to decide how to achieve it. That last part matters.

Start with a chatbot

Suppose you ask:

“What is our cancellation policy?”

The AI retrieves relevant information and answers. You asked a question. It answered. Useful? Absolutely. But that alone doesn't make it an agent.

Question → AI → Answer

The AI is helping you think or find information. It isn't really doing the job for you.

Now give AI a workflow

Suppose the task becomes:

“Process this cancellation request.”

Your company has a defined process:

Identify customer → Check order status → Retrieve cancellation terms → Calculate refund → Prepare confirmation → Manager approval → Update CRM → Send confirmation

Software can automate this workflow. AI might perform some individual steps. But if the path has been predetermined in software — always do A, then B, then C — it is more accurately described as an AI-enabled workflow.

That isn't a criticism. For many business processes, a predictable workflow is exactly what you want.

So when does it become an agent?

Now imagine something unexpected happens. The customer's order cannot be found using the email address.

Instead of simply failing, the AI decides: “I'll search using the postcode and surname.” It finds two possible customers, then decides: “I can't safely determine which customer this is. I need clarification.”

Or it discovers installation has already started, determines the normal cancellation workflow no longer applies, retrieves the appropriate contractual terms and routes the case for human review.

The important difference is that every step wasn't specified beforehand. The system was given an objective, relevant context, available tools, and rules and boundaries — then made bounded decisions about how to proceed. That is much closer to an AI agent.

OpenAI describes agents as systems that independently accomplish tasks on a user's behalf, using an LLM to manage workflow execution and tools to gather information or take actions — see their practical guide to building agents.

Anthropic makes a similar distinction: in a workflow, the route through the task is predefined; in an agent, the model dynamically directs its own process and tool usage — see building effective agents.

Workflow vs agent

Workflow

“Follow these steps to achieve this outcome.” The process determines the route.

Agent

“Achieve this outcome, using these permitted capabilities and staying within these boundaries.” The AI has some discretion over the route.

That doesn't mean unlimited freedom. For business use, the interesting question isn't “How autonomous can we make the AI?” It is: “How much autonomy does this particular job actually require?”

Autonomy is a dial, not a switch

It is tempting to divide AI into autonomous or not autonomous. Reality is more useful if you think of autonomy as a dial.

Imagine an AI dealing with customer refunds.

• Level 1 — Answer: “What is our refund policy?” It finds information and explains it.
• Level 2 — Advise: “Based on policy and this customer's circumstances, what refund appears appropriate?” It analyses and recommends.
• Level 3 — Prepare: “Calculate the refund and prepare the customer communication.” It does the work but doesn't execute the consequential action.
• Level 4 — Act with approval: It prepares everything and asks: “Refund of £425 prepared. Approve?” A human approves.
• Level 5 — Act autonomously within limits: Refunds under £100 that meet defined criteria may process automatically; a £2,000 refund gets escalated.

The same underlying AI could operate with very different levels of autonomy depending on the job, risk and permissions. Research on deployed agents treats autonomy as something shaped by system design, human oversight and deployment choices — see Anthropic on measuring agent autonomy in practice.

Agentic doesn't mean autonomous everything

A useful business agent does not need permission to do whatever it wants. It needs enough freedom to complete its job.

Think about a new employee. You wouldn't say: “You're intelligent, so you now have authority over every bank account, customer record and company system.” Their authority depends on their role.

The question isn't simply “What can the model do?” It is: “What is this digital worker authorised to do?”

Tools are what let an agent affect the world

An LLM on its own primarily processes and generates information. Give an agent tools and it might search customers, read CRM records, check calendars, retrieve invoices, calculate prices, prepare documents, create tasks or send approved email.

As in part 2, APIs and technologies such as MCP connect AI to those capabilities. The agent can decide which permitted tool it needs as it works through the job. Tool use is one of the defining characteristics of practical agentic systems.

Agents work in a loop

A chatbot interaction often looks like: ask → answer.

An agent is more likely to look like:

Understand → Plan → Act → Observe → Decide next → Act again → Check whether the objective is complete

That loop might happen several times. Example objective: book a suitable meeting with Sarah next week. The agent checks the calendar, finds no slots, checks constraints, discovers Wednesday afternoon can move, checks Sarah's availability, finds two possibilities — then either chooses within rules or asks for approval.

This ability to plan, act, observe and adjust is a useful way to think about agentic behaviour. Anthropic describes agents operating in a self-directed loop until the task is complete or human input is required — see building effective agents.

What about multi-agent systems?

Instead of one agent doing everything, several specialised agents might collaborate — customer agent, technical agent, pricing agent, proposal agent, quality agent.

That can be useful. But more agents do not automatically mean better AI. Every additional agent introduces more communication, decisions, errors, complexity and potentially cost.

Sometimes five agents are appropriate. Sometimes one agent with five tools is better. Sometimes you don't need an agent at all — a conventional workflow may be safer, cheaper and easier to understand.

The goal should never be “How many agents can we deploy?” It should be: “What is the simplest reliable way of getting this job done?”

So what is a digital worker?

This is where we make a deliberate distinction in Aadi.

An agent describes an AI system and how it behaves. A digital worker describes the role that system performs in the business.

You don't hire “a human brain with email access.” You hire an accounts assistant, a customer service adviser or a sales coordinator. The job gives meaning to the capability. The same principle applies to AI.

A digital worker has:

• A job — what is it responsible for?
• An objective — what outcome is it trying to achieve?
• Knowledge — what information may it use?
• Tools — which systems may it interact with?
• Permissions — what may it read, change or create?
• Boundaries — what must it never do?
• Workflow — what process should normally be followed?
• Autonomy — which decisions may it make itself?
• Human oversight — when must somebody review or approve?
• Evidence — what should be recorded?
• Definition of done — how do we know the job is complete?

That is why we prefer to talk about digital workers rather than simply agents. The technology matters. The business job matters more.

Chatbot, assistant, workflow, agent or digital worker?

• Chatbot — talks.
• AI assistant — helps.
• AI-enabled workflow — follows a defined process.
• AI agent — pursues an objective and makes bounded decisions about how to achieve it.
• Digital worker — agentic capability organised, governed and measured around a defined business job.

These aren't rigid scientific categories. The industry does not have one universally agreed definition of “agent.” The distinction is useful because it forces a more important business conversation.

Start with the job, not the agent

Imagine someone says: “We need an AI agent.” The next question should probably be: “Why?”

What process isn't working? What job needs doing? What outcome matters? What information and systems are involved? What needs judgement? What can safely be automated? Where must a human remain accountable?

Only then decide whether you need a simple LLM call, RAG, a deterministic workflow, an agent, multiple agents, or some combination. Sometimes the right answer is surprisingly simple.

The bigger picture

This brings the first three parts of our Plain-English AI series together:

• LLM — intelligence (part 1)
• RAG — relevant knowledge
• APIs and MCP — connections and standardised capabilities (part 2)
• Tools — things AI can use
• Workflows — predictable processes
• Agents — bounded discretion towards an objective
• Guardrails and humans — boundaries and approval
• Digital worker — the job packaging

Process → Job → Digital worker → Outcome

Don't start by asking “Where can we use an AI agent?” Start with: “What job needs to get done?” Then give AI exactly the intelligence, knowledge, tools and authority required — and no more.

Next: Part 4 — what does sovereign AI actually mean? (US, EU, UK, China and the SME sovereignty test.)