- An agent doesn’t just answer. It decides on steps and uses tools to complete a task.
- The best first agents handle repetitive, well-defined work with clear rules.
- Good tools and guardrails matter more than the choice of model.
- Keep a person in the loop for anything sensitive, costly or customer-facing.
What an AI agent actually is
A chatbot answers questions. An AI agent goes further: given a goal, it works out the steps, uses tools such as your CRM, a database or an email system, checks the results and keeps going until the task is done or it needs help.
Take a simple request: “Follow up with everyone who asked for a quote last week.” A chatbot can tell you how to do it. An agent can find those enquiries, draft a personalised reply for each, and queue them for someone to approve.
That ability to act is what makes agents useful. It’s also why they need to be designed carefully.
Where agents help, and where they don’t
Agents work best on tasks that are repetitive, follow reasonably clear rules and touch several systems. Good candidates include:
- Sorting and summarising incoming emails or support tickets.
- Pulling key fields from documents and entering them into a system.
- Preparing first drafts of routine replies, reports or updates.
- Looking up information across tools to answer an internal question.
They’re a weaker fit for one-off strategic decisions, work where a single mistake is very costly, or processes nobody can clearly describe. If the steps can’t be written down, an agent will struggle to follow them reliably.
If you can’t describe the task clearly to a new colleague, you’re not ready to hand it to an agent.
The four building blocks
Every useful agent comes down to four parts. Getting each one right matters more than which model sits underneath.
Designing good tools
Tools are how an agent interacts with your business, so they deserve the most care. A few principles go a long way:
- Small and specific. “Find open quotes for a customer” is safer and more reliable than general database access.
- Least privilege. Give read access where reading is enough. Add write access only where it’s needed.
- Clear inputs and outputs. Well-described tools are used correctly far more often.
- Safe by default. Actions that are hard to undo, like sending, paying or deleting, should go through a review step.
Keeping people in the loop
The goal isn’t to remove people from the process. It’s to remove the repetitive parts so they can focus on judgement. In practice, that means deciding up front which actions an agent can take on its own and which need approval.
A simple rule of thumb: let the agent prepare, and let a person approve anything that leaves the building, costs money or is hard to undo.
Over time, as you see how the agent performs, you can widen what it does on its own, one step at a time.
Measuring whether it works
Before an agent goes live, agree how you’ll judge it. Useful measures include:
- Quality: how often its output is accepted without changes.
- Time: how long the task takes with the agent compared with before.
- Escalations: how often it correctly hands a case to a person.
- Cost: the running cost per task, including model usage.
Test against a set of real past examples first, so you can see where it does well and where it struggles before anyone relies on it.
A practical way to start
- Pick one repetitive task that clearly costs your team time.
- Write down the steps, the rules and what a good result looks like.
- Build a small agent with only the tools that task needs.
- Test it on real examples, with a person approving every action.
- Measure the results, then decide whether to expand it.
Starting small keeps the risk low and the learning fast. Once one agent is working well, the same patterns make the next one much quicker to build.
If you’d like help identifying a good first use case, our AI & Business Automation team can help.