One task at a time
Start with a limited role, such as searching a catalogue or drafting a support response. The clearer the scope, the easier it is to evaluate performance.
Access with a purpose
The assistant needs only the sources and tools required for its task. Define who can see which data and which actions require approval.
Uncertainty needs a process
If sources are insufficient, the agent should recognise that. If it encounters an exception, it needs a clear route to a person. A well-organised request for help is preferable to a convincing answer without supporting evidence.
Operations continue after handover
Sources change, new questions arise and models evolve. Plan for knowledge updates, evaluation and an accountable owner. An agent is a process that needs maintenance.
Define the assistant's remit precisely
The label “customer service assistant” leaves too much open. Specify what it can answer, from which sources and for which audience. Also document what falls outside its remit: changing an order, promising compensation or making a final commercial decision. A clear definition makes behaviour testable in real cases and understandable to the team supporting it.
Every tool needs usage limits
When an agent can use tools, its read and write permissions must be clear. Being able to look up an order does not mean it should be allowed to change it. Actions with external consequences need appropriate approval and checks. Permissions should not rely solely on a natural-language instruction to the model.
Test when the assistant should refuse to act
An evaluation set needs cases where the assistant should stop. Information may be missing, the request may concern a different product or the user may ask for an action outside its remit. The correct response could be clarification or escalation. This capability matters because it prevents the process from continuing without a sound basis.
A human handoff should be useful
A simple “speak to someone” message often fails to solve the problem. Design the information passed to the responsible person: the request, what has already been checked and where the assistant stopped. Include only what is needed to continue. The team then receives useful context, and the customer does not have to repeat the entire conversation.
Frequently asked questions
How does an agent differ from a chatbot?
A chatbot typically provides a conversation. Depending on its implementation, an agent can use tools and carry out steps towards a task. The real distinction lies in the system's capabilities and controls, rather than its name.
Can it operate without human review?
Autonomy is designed around each task. Drafting and information retrieval may allow more freedom. Actions that create commitments or change important data need suitable limits and approval levels.
How does it learn about our business?
We give it specific instructions and access to approved sources. We do not assume it already knows your products, policies or procedures. Organising and updating those sources is an essential part of implementation.
Can it give an incorrect answer?
Yes. We therefore include evaluation, source references, uncertainty handling and human escalation. These checks reduce failures without guaranteeing zero errors. Suitable use depends on the requirements of the particular task.