Give an AI agent a job you can review.

Start with the output

“Handle enquiries” is too broad for a first implementation. A more useful scope is: collect required details, prepare a draft response and create a review task. The team can inspect that output and decide whether it helps.

Separate known facts from missing details

Define the information needed before a response can be prepared. For a quotation request that might include scope, location, quantity and timing. The agent should ask for missing information or flag uncertainty rather than inventing a requirement or price.

Make approval an actual step

A draft is not a sent message. Decide who approves it, what they can edit and what happens after approval. The permission to retrieve records does not automatically imply permission to send an email or modify those records.

Evaluate awkward inputs

Test incomplete enquiries, conflicting details, duplicate submissions and requests outside the intended scope. Review whether the proposed next action is useful and whether the agent asks for help at the right point. Record failures as well as successful examples.

Measure the whole workflow

Compare the time spent gathering details, reviewing drafts and correcting errors with the existing process. A fast generated response is not enough if a person needs longer to correct it. Establish the baseline before making time-saving claims.

See a worked example

The enquiry demo on this site uses fictional data and shows the approval boundary explicitly. Nothing is sent. Use it to discuss the steps your own workflow would require before connecting live tools.