AI agent
An AI agent is a system that pursues a goal over several steps — planning, calling tools, checking intermediate results — rather than returning a single answer to a single prompt. The interesting question for a business is not how autonomous it is but whether it can write: a system that reads your documents and drafts something is a different risk proposition from one that places an order.
Read access or write access?
This is the decision that determines everything downstream, and it is worth making explicitly rather than inheriting from a product demo. The majority of value in enterprise use comes from read-only work: searching, comparing, summarising, drafting for review.
Write access multiplies value occasionally and risk always. Every action that changes state needs a rollback story, an audit trail and someone who owns the failure. We enable it when a case genuinely requires it, and narrowly.
Why do agent projects fail more often?
Because errors compound along the chain. Five steps at 95 per cent each leave a run correct 77 per cent of the time; ten steps drop below 60. Each individual step looks reliable, which makes the cause hard to see from the outside.
The remedy is rarely a better model. It is a shorter chain with a checkpoint in the middle. Two steps with a human approval between them beat eight without — in output quality and in whether anyone keeps using it.
What makes a process suitable?
Repeatability and reversibility. A task that runs the same way every day and whose intermediate result someone signs off is a good candidate. A task with many exceptions and no way to undo is not — there a single error costs more than the automation saves.
The most useful question in a vendor conversation is therefore not "what can the agent do" but "what happens when step three of five was wrong". An answer that is not immediate means you are being shown a demonstration rather than an application.
What it is not to be confused with
Workflow automation
Classical automation follows fixed rules: on condition A, do B. Agentic means the system chooses the path. For well-described processes the rule wins — it is cheaper, faster and auditable. Agents earn their keep only where the path is not known in advance.
Chatbot
A chatbot answers a question. An agent pursues a goal across several moves and reaches for tools on the way. The risk difference is material: a wrong answer annoys, a wrong action has to be unwound.
Frequently asked
What is an AI agent in simple terms?+
A system that is given a goal rather than a prompt, breaks it into steps itself, uses tools to carry them out, and evaluates what came back. The everyday example is a system that reads an incoming email, finds the relevant contract, checks a deadline and drafts a reply — four steps from one instruction.
Do agents need access to our systems?+
Read access to the relevant documents, yes. Write access is a separate decision and should be treated as one. Starting read-only and extending later is almost always the right sequence, and it is much easier than withdrawing permissions after an incident.
Does an agent fall under the EU AI Act?+
Not because of how it is built. Classification follows the use case: an agent that pre-sorts invoices is minimal risk; one that screens job applicants is an Annex III high-risk case. If it interacts directly with people, the Article 50 transparency duties apply on top.
How many steps is too many?+
Fewer than most products suggest. Past three or four unsupervised steps the compound error rate makes results unpredictable and, worse, unpredictable in ways nobody notices. If a process genuinely needs ten steps, it needs checkpoints, not confidence.
The scoping question
Name one process that runs the same way every week, and say what happens if it goes wrong halfway. Ten minutes on that establishes whether an agent is the right instrument or ordinary automation would do.
Request a test