At EY in Singapore, staff are using AI agents to analyse documents while people retain final judgement, according to CNA’s September 14 report on business adoption.
For a smaller business, the starting point could be more familiar: gathering the information needed to answer a customer enquiry. If you mainly use AI to ask questions or tidy up an email, you might wonder how it can do that work for you.
What changes when an AI becomes an “agent”?
From writing an answer to taking a step
Start with the model: software trained to interpret requests and generate responses. You may have heard it called a large language model, or LLM.
Give it a customer enquiry and it can draft a reply. On its own, that reply is text. To check your stock or send the message, the model needs software that can perform those actions.
These capabilities are usually called tools. A tool might search a product catalogue, read a calendar or create an email draft. The model requests a tool action; the connected software carries it out and returns the result.
An agent combines the model with tools and a process for using them. It can choose a step, see the result and decide what to do next.
A customer enquiry, worked through
Imagine a customer asks whether you have 20 chairs available for delivery next Friday.
An assistant set up for this job could:
- Read the enquiry and identify the product, quantity and requested date.
- Look up the relevant stock record.
- Check the delivery information available to it.
- Prepare a reply for a staff member to review.
The useful part is bringing the information together, so a member of staff can review a prepared reply instead of checking several systems and copying the details themselves.
If the delivery information is missing, it might ask for clarification. If the stock tool returns an error, it might try again. Which choices it makes depends on its instructions, available tools and the way the system is built.
A fixed automation can also handle enquiries. The difference is that its steps are predefined, while an agent has more freedom to choose its next step. A predictable task may only need the fixed workflow.
Who tells it what to do?
Someone configures the purpose, tools and conditions for starting work. You might give it a task directly, or a new enquiry might trigger a job arranged earlier.
The agent can then choose intermediate steps without a person specifying each one. That is useful, but it also creates room for mistakes. It may misunderstand the goal, trust incorrect information or take an action the person setting it up did not intend.
For a business owner, the important distinction is how much freedom the system has. Preparing a reply, promising a delivery date and sending that promise to a customer each carry different consequences.
Before connecting an agent to your business, decide where that freedom should end. Our companion guide explains three limits to set before an AI agent uses your business tools.
Thinking of connecting AI to your business tools? Tell us what you want it to do, and we can work through the access and checks it needs.
Technical source: Anthropic’s explanation of workflows, agents and tools. The customer enquiry above is an illustrative example, not a client case study.