Answer tier-one questions from your own docs

Most support questions have been answered before, usually in writing, usually by you. The problem is that the answer is in a document nobody opens and the question arrives at 2am.

This playbook closes that gap without letting anything reach a customer unreviewed.

Put the answers where they can be found

Open Documents and upload what you already have: help articles, policies, your refund and returns rules, setup guides, the FAQ nobody maintains.

If your answers currently live in someone's head rather than in a file, this is the moment that costs you an afternoon and saves you the next year. Write the ten most-asked questions and their answers into one document. It does not have to be polished.

Check that it answers correctly before it answers anyone

Use Ask and put your ten hardest real questions to it. Not typical ones. Hard ones, and especially the ones where the honest answer is "it depends".

A customer on the annual plan is 40 days into their term and wants a
refund because they say the product does not do something they were
promised. According to our documents, what are they entitled to, and
which policy says so?

You are checking two things: that it finds the right policy, and that it does not invent a reasonable-sounding one when your documents are silent. The second is the dangerous case, and the fix is a document, not a prompt.

A gap you discover here is a gap in your policy, not in the tool. If three of your ten questions have no answer in writing, you have just found the most valuable afternoon of work available to you, and it has nothing to do with AI.

Give the employee its brief

Hire or nominate a support employee and give it a standing brief. The two rules that matter are what it must cite and when it must stop.

You answer customer support questions using only our documents.

Always search the documents before answering. Quote or reference the
document your answer comes from.

If the documents do not answer the question, say so and hand it to a
human. Do not reason from general knowledge about how companies usually
handle this. A wrong answer delivered confidently costs us more than a
slow one.

Hand off to a human immediately, without answering, if the customer is
angry, if they mention legal action or a chargeback, if they are asking
for an exception to policy, or if the question involves their specific
billing details.

Match the customer's language. Keep answers short.

Start it where a mistake is cheap

Each employee has its own real email address, so the natural first deployment is a low-stakes inbox rather than your main support queue.

Better still, run it in draft: let it answer into a queue you review before anything sends. Watching twenty drafts is a much better way to calibrate than reading twenty prompts.

Feed what it learns back into the docs

The handoffs are the useful output. Every question it could not answer is a missing document.

Every Friday, review the questions you handed to a human this week
because our documents did not cover them. Group them by theme, and
draft a new help article for the three most common. Save the drafts to
the Help Drafts folder.

What good looks like

First response time drops to under a minute at all hours, and every answer that went out can be traced to a document you wrote.

The honest measure after a month is not how many tickets it answered. It is how many articles you now have that you did not have before.

Also covers

Escalating when someone is frustrated. The handoff rules above already name anger as a trigger, and it is worth doing better than keyword matching.

Hand a thread to a human, without answering, when the customer sounds
frustrated: repeated contact about the same issue, shorter and sharper
replies than earlier in the thread, phrases like "again" or "still",
explicit mention of leaving.

Say which signal you saw. Do not attempt to de-escalate first.

That last instruction matters. An automated attempt at soothing an angry customer is the thing most likely to make them angrier, because it reads as being handled rather than heard.

Replying in the customer's own language. One line in the brief:

Reply in the language the customer wrote in. If you are not confident you
can answer accurately in that language, answer in English and say you are
happy to continue in their language with a colleague.

The caveat is the point. A confident reply in imperfect Portuguese about a refund policy is worse than a clear one in English.

Keeping first response under a minute at 2am. This is not a separate configuration. It is what happens once an employee holds the mailbox: it does not keep hours, so the 2am question gets the same answer as the 2pm one.

What is worth deciding is whether that first response sends unattended or waits for approval. Overnight is the strongest argument for letting the safest category send, because the alternative is an eight-hour wait — but keep the handoff rules above absolutely intact. Fast and wrong at 2am is worse than slow.