A customer asks whether your crew can visit a building after hours. Your website says to call for availability. An old email template says evening appointments are available, but your dispatcher knows they are only offered for certain jobs. If a website assistant reads all three versions, the problem is not that it needs a friendlier tone. The business has not agreed on the answer. Before adding AI to customer inquiries, make a small library of answers your staff can stand behind. The same work can improve the website, the front desk, and the handoff to the person who can actually help.
Start with one channel and one concrete purpose. For a service business, that might mean helping people find the right way to request an estimate. It does not have to mean quoting prices, checking calendars, or resolving complaints. Ask the people who respond to inquiries to bring a short list of repeat questions from a typical week. Write down the questions in customer language without copying names, phone numbers, addresses, or account details. Compare the list with the top questions on your site. The useful first batch is often the intersection: questions customers ask repeatedly for which your business already has a stable, public answer.
Then find the source for each answer. A location page may establish which towns you serve. A contact page may describe the best way to reach the office. A signed internal policy might explain what your team can promise about response times, but that does not make the entire policy public. Record the source, what the assistant is allowed to say, who can approve that wording, and when it was last checked. Put the approved text in one place. If there is a disagreement, leave the answer out of the first release and ask the owner to settle it. Copying both versions into the tool just moves the disagreement to the customer.
An answer record can fit in one row. Give it a question, a short public answer, a source link, an owner, and a review date. Add a boundary and a handoff path. For example, the row for estimates might say that customers can send a brief description through the contact form, that someone will confirm whether the work is within scope, and that the assistant must not name a price or book a visit. The boundary is as important as the answer. It tells the team what not to do when a customer asks the same question with more urgency or gives an unusually detailed description.
Make the public source easy to read on its own. If the site promises that an inquiry goes to the right department but does not say which requests belong there, fix the page first. Put service areas, contact hours, and limitations near the action a visitor needs to take. Remove duplicate wording from pages that no longer own the answer, or link back to the authoritative page. Good website design makes the next step obvious even if the assistant is offline, the visitor prefers not to use it, or a screen reader user skips directly to the contact information.
Decide what the assistant should do when it cannot answer. A useful fallback is specific: explain what it cannot confirm and show the next available channel. A visitor asking if a technician is coming today does not need a generic apology. They need to know that the website cannot see their appointment and where to check. Test that path on a phone. Does the contact button actually open? Is the office number selectable? Does a form acknowledge that it received the request, without making a promise about when someone will reply? Make sure staff know who picks up the inquiry after it leaves the site.
Keep customer details out of the answer library. The public guidance can say that a service coordinator will follow up, but it should not contain example records copied from real bookings. Use fictional details for testing and only the minimum fields the destination actually needs. Do not prompt visitors to put passwords, payment information, or sensitive documents in a general inquiry box. If the business needs to collect more information later, send the customer to its established process. Keep the library of public answers separate from staff notes, customer records, and internal instructions. The assistant should never need access to those sources just to tell someone where to start.
For each approved question, write at least one normal test and one awkward version. Try a typo, two questions in the same sentence, or a request that sounds like the approved question but is outside scope. Ask for a same day appointment where no date can be confirmed. Ask whether an unfamiliar town is included instead of naming a town the business already lists. Try a vague request like, Can you help me today? A good response uses the approved answer when it fits, asks a narrow follow up when that would help, and sends uncertain cases to staff without inventing availability.
Review the tests with the person who would handle the real inquiry. Judge them on three things: whether the answer is accurate, whether it leaves out any important condition, and whether the next step works. Note the exact reason for each failure. If a link opens a page that no longer has the relevant form, fix the destination. If the assistant makes an unsupported promise, tighten the answer boundary and retest. If a single question is asked several ways, save those test cases together so the next update cannot quietly undo the fix. Keep the release small enough that someone can actually run the checks again.
When a website needs to pass requests into a shared queue, document the handoff separately from the answer. The customer might be allowed to ask about estimate requests, but the actual routing could depend on location, service type, or office hours. Map that path from the form to the receiving team, including what happens when routing is uncertain or delivery fails. If you are considering custom workflow automation, start by agreeing on which information is required and which system owns it. A clear answer on the front end does not help if the request is quietly lost behind it.
Give a named person the ability to change or pause the rollout. When hours change for a holiday, update the authoritative page, replace the approved response, and run the affected test cases. Do not keep two active answers and hope the newest one wins. Set a short review cycle at first, then adjust it to how often your business information changes. An owner should be able to see what changed, who approved it, and what customers were told in the test. The NIST generative AI risk management profile offers a broader reference for teams formalizing human oversight and review.
For the first few weeks, record questions the assistant cannot answer, broken destinations, and any promises staff had to correct. Use categories or short anonymous summaries rather than retaining unnecessary customer details. Review the patterns with your team. If customers keep asking about service areas, the answer might belong on the service page. If they cannot find a request form, move the link closer to the relevant explanation. If the team is seeing requests for work you do not offer, add a plain statement of the boundary. The goal is not to make every interaction end with an AI answer. It is to help the customer take the right next step.
If the assistant was part of an AI generated website draft, review the rest of the site at the same time. Placeholder links, inconsistent contact details, and copied service language can create conflicting answers even when the library is well written. The work of turning an AI website draft into a production site includes making sure those pages, forms, and integrations agree. Treat the answer library as one piece of that review. Walk through the public pages without opening the assistant and make sure a customer can still learn what you do and contact the right person.
To begin, pick five questions the team heard this week. Give each one an approved answer, a source, an owner, and a rule for when to hand it off. Then test the answers against the website and the office process. Start with the items you can explain without seeing a customer record or making a commitment. Keep a copy of the test cases so a future update starts with evidence instead of memory. You will have a useful reference even if you never turn on AI. If you do, you will know exactly which questions it can answer, which ones it should pass on, and who is responsible when the business changes.
