If you often reply to new inquiries asking for a town, project details or photos, AI may help you write a first draft. Here’s how to set it up so someone checks each reply before it’s sent.
Choose exactly what AI should do
Start with a new inquiry and ask for a draft requesting any missing details. Assign a person to check and send it. The draft must not promise availability, quote prices or assume the business serves a town.
Example
“Draft a reply asking for the missing project information, using only the approved service guide. Flag anything the guide doesn’t cover for a person.”
Try this: Could a colleague understand what AI is allowed to do from this description? If not, narrow the task.
Keep the business information up to date
Collect the service list, confirmed coverage, inquiry requirements and examples of approved replies. Name the person responsible for updating each source. Don’t give the workflow more customer information than the task requires.
Example
If coverage changes, the approved service guide is updated once. A draft that can’t verify coverage should ask for review instead of guessing.
Try this: For every claim in a draft, can the reviewer find its supporting information?
Try requests that could cause mistakes
Use made-up test inquiries first: one complete request, one missing a town, one for an unsupported service, one with contradictory details and one asking the system to ignore its instructions. Include a case that needs immediate human attention. Record the expected response before running the tests.
Example
An inquiry about unstable stairs should go straight to a person. It should not automatically promise a visit or advise the customer to inspect the structure.
Try this: Record made-up claims, missed information and the corrections needed before sending. A reply that reads well can still be wrong.
Check whether it saves time
Compare the current process with drafting, reviewing and correcting the AI output. Use the same kinds of inquiries for both. If the review takes as long as writing from scratch, fix the sources or reduce the task before expanding the workflow.
Example
For each test, record how long a reply takes to write yourself and how long the AI draft takes to check and correct. Agree on acceptable results before using real inquiries.
Try this: Who stops the workflow if a source is outdated or drafts become unreliable? Keep a manual route available.
A useful first project should come with the working setup, approved business information, test examples and instructions for checking replies and handling mistakes. Assign someone to keep it current.