The distinction that actually matters #
Put three things side by side and the differences become obvious.
| Arrangement | How it decides what to do | What can go wrong |
|---|---|---|
| Ordinary automation | Follows a path somebody wrote in advance | It breaks, loudly and in the same place every time |
| A model answering a question | Produces text. Does nothing else | The text is wrong, and a person notices or does not |
| An agent | Picks its own steps from the tools it has been given | It does something nobody predicted, correctly following a bad reading of the situation |
That last row is the whole story. An agent's failure mode is not a crash. It is a confident, plausible, wrong action, taken quickly. This is why the important specification for an agent is never the model. It is the list of tools it holds and the boundary around them.
Tools, in this context, means permissions #
When a supplier says an agent has tools, what they mean is that it has been granted the ability to do specific things: read a calendar, search a document store, create a record, send an email. Each of those is a permission, and permissions are the part of the design a business owner can understand and should insist on seeing.
A sensible way to read any proposal is to ignore the description of the intelligence entirely and read only the list of what it can touch. If that list includes your customer records, your mailbox and your payment system, then the question is not how clever it is. The question is what happens on the day it is confidently wrong. Our page on what should always wait for a person covers where that boundary belongs.
Why the outputs are guesses, and why that matters more for agents #
The Information Commissioner's Office advises treating AI outputs as statistically informed guesses rather than facts, and notes that a system does not have to be one hundred per cent statistically accurate to comply with data protection law.
The gap between those two situations is the entire risk difference, and it is why a business is usually better served by an arrangement where the software does the fetching and the drafting and a person does the sending. That is less impressive to demonstrate and considerably easier to live with.
When an agent is genuinely the right shape #
There is a real case for them, and it is narrower than the word's popularity suggests.
An agent earns its place when
- The sequence of steps genuinely varies
- Nobody can write down every branch in advance
- The cost of a wrong step is low and recoverable
Research and triage tasks often fit. Anything that touches money or a customer relationship usually does not.
A fixed path is better when the steps are known
- It costs less to build
- It can be tested against real examples
- It fails in predictable places
- Anybody can be shown what it does
Better in every respect that matters to a business.
The test for whether your work is the fixed kind is in how to tell which jobs are worth automating.
Three questions to ask about any agent #
-
What is the complete list of things it can do?
Not what it is for, what it is capable of. A short list is a good sign.
-
Which of those happen without a person seeing them first?
The honest answer is a specific list, not "it is supervised".
-
What does it leave behind?
If you cannot look back a week later and see what it did and why, nobody can investigate the one time it went wrong.
Those questions belong in any supplier conversation, alongside the data protection ones in the questions to ask an AI supplier.
The regulatory ground is still moving #
Worth knowing before you build anything ambitious: the ICO's own guidance on AI and data protection was last updated on 15 March 2023 and now carries a notice at the top saying that, because of changes made by the Data (Use and Access) Act, the guidance is under review and may be subject to change. That was the position when we checked it on 19 September 2026.
The practical reading is not that the rules are unknowable. It is that a design which keeps a person in front of consequential actions will survive whatever the updated guidance says, and a design that removed the person may not.
Where to start instead #
Before deciding whether you want agents, it is worth knowing how much repeated work you actually have. The cost of admin time calculator counts the hours and puts a real employer cost against them. If the total is small, the answer to every question on this page is that none of it applies to you yet.