Do you need help & advice with AI & Automation?
Everything called AI for the last two years has essentially answered questions. You type, it replies, you decide what to do. An agent is different. You give it a goal and it goes and does things.
That is a genuinely useful shift and it is also where the quiet problems start, because it is much harder to notice something that acted than something that answered.
What is the difference between AI and an AI agent?
One produces text. The other takes actions in your systems.
| What it does | What can go wrong | |
|---|---|---|
| Generative AI ChatGPT, Copilot, Claude | Produces something for you to look at. Draft, summary, answer | It is wrong and you use it anyway. You can see the output before it matters |
| An AI agent | A prebuilt tool that does one specific job. Sorts email, extracts invoice data, answers a common query | It does the job wrong, repeatedly, and nobody notices for a fortnight |
| Agentic AI | Given a goal, works out its own steps, and carries them out across systems | It does something you did not anticipate, and the record of why is hard to reconstruct |
The practical distinction is not how clever it is. It is whether a person sees the work before it takes effect.
Should my business use AI agents?
Yes, for narrow specific jobs with a human checkpoint. No, for anything open-ended.
The way most people are experimenting with this at the moment is to hand over something vague and see what comes back. That produces noise. The version that works is the opposite: give it one well-defined task, give it a framework to work inside, and have it come back and ask when it gets stuck.
That last part is the one to insist on. The industry calls it human in the loop, which makes it sound optional. Treat it as the condition of use.
| Reasonable to try | Not yet |
|---|---|
| Sorting or routing incoming email | Anything that sends something to a client unseen |
| Pulling data off invoices into a system | Anything that commits money |
| Drafting a standard reply for someone to send | Anything with access to everything, “to be helpful” |
| Flagging documents that need attention | Anything a supplier has built into their system on your behalf |
The unanswered question. If your agent does something costly or unlawful, where does the liability sit? Nobody has settled that yet, and until somebody has, the practical answer is that it sits with you. That is a good reason not to be the business that finds out.
How risky is this really?
Riskier than the tooling suggests, because the controls have not been built yet.
McKinsey found that 80% of organisations have already seen an AI agent behave in a way they did not expect. That number is high because the field is young, not because the technology is bad.
For a sense of how young: one of the more talked-about agent platforms has been renamed twice, ships with essentially no security controls, and the agents running on it have produced their own religion and a Reddit clone that reads as pure noise. Interesting to watch. Not something to connect to your accounts system.
The wider risk is not dramatic. It is that agents fill the internet, and your business, with plausible noise that nobody asked for.
What has to be in place first?
Two things, and neither is technical.
| What | Why it comes first | |
|---|---|---|
| 1 | An AI policy that says what is approved and what data can go where | Otherwise people experiment with agent platforms unsupervised, and those platforms have no security to speak of |
| 2 | Somebody whose job this is, by name | An agent is a member of staff with system access. Every other member of staff has a manager |
Without those two, an agent is an unsupervised employee with the keys to everything and nobody to report to. With them, it is a tool doing a defined job.
The uncomfortable bit is the pace. People are poor at judging exponential change: it looks flat, then it does not. Sitting it out and catching up later is not the safe option it feels like.
[VIDEO_PLACEHOLDER]
So, what now?
Do not start with an agent. Start by writing down who owns AI in your business and what the rules are. Then pick one narrow job, insist on a human checkpoint, and see whether it actually saves anybody time.
If you would rather somebody kept track of this on your behalf, including which agents your suppliers have quietly built into their systems, that is what an AI Manager does. The basics are in getting started with AI in your business.
Frequently asked questions
What is an AI agent?
A prebuilt AI tool that does one specific job rather than answering questions.
A chatbot handling common queries, software that sorts your email, a tool that lifts data off invoices. Focused, predictable, and limited to the task it was built for.
What is agentic AI?
AI that plans its own steps to reach a goal you set, rather than doing one defined task.
Give it an objective and it decides how to get there, which can include using your systems and making decisions without checking back. More capable, and considerably more supervision required.
What does human in the loop mean?
A person checks or approves the work before it takes effect.
It sounds like a nice-to-have. For anything that touches a client, a price or a system of record, treat it as the condition of using an agent at all.
Who is liable if an AI agent makes a costly mistake?
Unsettled, which in practice means you are.
The law has not caught up, and vendor terms generally do not accept responsibility for what an agent decides. Assume the risk is yours until somebody demonstrates otherwise.
Do we need agents to get value out of AI?
No. Most businesses have not exhausted the value of ordinary AI yet.
Drafting, summarising and searching your own documents saves real time and carries far less risk. Agents are the next step, not the first one.
Last reviewed: 17 August 2026. Agent platforms change fast, so this page is reviewed quarterly.
Related
- What should you automate first
The jobs that eat time and add nothing.
