What is an AI agent, and what is agentic AI?
An AI agent is software given a goal and tools that works through the steps itself. Agentic AI is AI that works this way. What both mean for an owner.
·
6 minute read
An AI agent is software, built on an AI language model, that is given a goal, a set of tools and permission to work through the steps itself. Agentic AI is the name for AI that, given a goal, plans the steps and carries them out with tools, with little prompting from a person between one step and the next.
An agent reads what arrives, decides what to do next, acts in the systems it has been given, checks the result and carries on until the job is done or it needs a person, where a chat assistant answers the question in front of it and stops.
How do official sources define an AI agent and agentic AI?
The UK government's glossary for civil servants defines AI agents: "Small computer programmes that can perform tasks on your behalf." The government's AI Playbook describes agentic AI as autonomous AI systems that can make decisions and perform actions with minimal human intervention, and says what makes that possible: an agent understands its environment, identifies the tools and functions at its disposal and uses them to take actions towards its objective. The Information Commission's Office, in its report on agentic AI, describes it as combining the capabilities of generative AI with additional tools and new ways of interacting with the world.
Three parts of that matter to an owner.
A goal, not a question. "Get this repair booked with an approved contractor" is a goal. The agent works out the steps.
Tools. An agent can do things because it has been connected to things: a mailbox it can read, a record it can update, a calendar it can book. Without tools it can only write text.
A loop. It acts, looks at what happened and decides the next step. That is what lets it cope when the reply is not the one it expected.
What is the difference between generative AI and agentic AI?
Generative AI produces a draft for you to use. Agentic AI uses the draft itself: it sends, files, books or updates, inside whatever limits it has been given. Asked to reply to a customer whose emailed booking leaves out the delivery slot, a generative tool writes a good email; an agentic system finds the account and puts the job on hold with its queries listed, for a planner to release.
What does an AI agent look like in a working day?
Take a letting agency. A tenant emails at seven in the evening to say the kitchen tap will not stop dripping. Nobody asked the agent anything. It reads the email, finds the tenancy and the property in the agency's software, checks which contractor the landlord has approved and whether a job of this kind is likely to fall inside the limit in the management agreement, logs the job, asks the tenant for the two details that are missing and drafts the instruction to the contractor for a property manager to send. In the morning the property manager finds a job that is logged, complete and waiting for one decision. That work, set out as a role, is maintenance request coordination.
The same shape, with a person approving whatever matters, fits most written admin. Our article on whether AI can do back-office admin sets out where the line sits.
How much does an AI agent do on its own?
That depends on a setting for each kind of action.
Suggest. The system proposes and a person does the action.
Act on approval. The system prepares the action and it happens when a person confirms. This is human in the loop.
Act within limits. Routine, low-risk actions go ahead under a standing instruction, inside fixed guardrails, and are recorded.
Never. Some actions are not available to the system at all.
A well-designed agent uses all four at once: it files an attachment without asking, waits for approval before writing to a customer about a dispute, and cannot issue a refund whatever it is told. The Playbook's phrase for this is meaningful human control at the right stages.
What is an AI agent confused with?
A chatbot. A chatbot waits to be asked and replies in words. An agent is given work and acts in your systems. Our comparison of an AI worker and a chatbot sets out the difference.
Automation. A rule that says "when a form arrives, create a record" does the same thing every time and stops when the form is unusual. An agent reads the unusual one and decides what to do with it.
An AI assistant inside an application. The helper in your email or document software works for one person, on what that person has open, when that person asks. An agent works on a queue of its own. Whether ChatGPT, Claude or Copilot is enough for your admin is a separate question, taken in ChatGPT or Copilot against an AI agent.
An AI employee. An agent that has been given one job in a business and a person who manages it, as our explainer on what an AI employee is sets out.
Why does the definition matter to an owner?
Because an agent acts, the questions that matter change. With a chat assistant you ask whether its answers are right. With an agent you ask what it can reach, what it is allowed to do there, and who sees its work before it leaves the building.
The Information Commission's Office says something close in its report on agentic AI: organisations remain responsible for the data protection compliance of the agentic AI they deploy, and choices about the data and tools a system can access, and the controls around it, really matter. Its chapter on risks adds that agentic systems operate, by design, with limited human involvement.
Before an agent is given a job, an owner should be able to answer five things:
What is the job, written down, and where does it end?
Which systems can it read, which can it change, and by what route? Those limits are its guardrails.
What can it not do, and is that enforced by the system or only by an instruction to the model?
Which steps wait for a person's approval?
Where is the record of what it did? That is its audit trail.
What stays with a person?
An agent is suited to the repeatable part of a job: reading, looking up, matching, chasing, drafting, updating. Decisions that carry professional judgement, a legal consequence or a client relationship stay with a person, and so does the telephone. In the letting agency, whether to spend the landlord's money above the agreed limit, and what to say to a tenant who is upset, is the property manager's.
An agent that nobody manages is a liability with a login. An AI worker is an agent given a single job, a specification and a manager, and our explainer on AI staff describes the model. What one costs is on our homepage.
In your business the agent would have one defined job, a specification you approve and one of your own people checking its work. If there is a queue in your business that an agent could work, talk to our agent.
Carlos Oliveira
Founder of The AR Dept
Losing billable hours to this process every week?
Check if your team’s routine admin can be handed to an AI worker.
