How to Build an Approved Answers Library for Your AI Reply Agent

Most bad AI replies are not badly written. They are confidently wrong. The prospect asks whether you integrate with their CRM, whether you can start next month, or what the contract term is, and the agent produces a fluent answer that nobody on your team ever agreed to.

Tone can be trained. Facts have to be supplied. That is the job of an approved answers library: a short, owned document that lists what your AI reply agent is allowed to say about the questions prospects actually ask, and what it must hand to a person instead.

This guide covers how to build one from your real replies, how to structure each entry so an agent can use it, how to connect it to escalation, and how to keep it from going stale.

What an approved answers library is (and is not)

An approved answers library is a set of entries, one per recurring question, each with an approved answer, the conditions where that answer applies, and a clear rule for when to escalate instead.

It is not:

• A brand voice guide. Voice decides how the reply sounds. The library decides what it claims. You need both; our guide to training an AI reply agent to match your brand voice covers the voice side.
• Your whole website pasted into a prompt. Marketing pages are written to persuade, not to answer a specific question in two sentences. They also contain claims that need context a reply cannot give.
• A script. Entries give the agent the facts and the boundaries. The agent still writes a reply that fits the thread.

The test of a good library is simple: if a new salesperson read it on day one, could they answer most replies correctly and know exactly which ones to pass on?

Step 1: Pull the questions from your real replies

Do not start from a blank page or from what you think prospects ask. Start from your inbox.

Export a few weeks of replies across every mailbox and LinkedIn account your campaigns use. If your replies are spread across several sending tools, this is much easier from a unified inbox than from each tool separately. Then:

1. Drop the replies that need no facts. Out of office, unsubscribe, “not interested” and simple “yes, let’s talk” replies are handled by categories, not by answers. Our reply categories and triage workflow covers those.
2. Highlight every question. Anything the prospect asks, explicitly or implicitly. “Does this work with our setup?” counts. So does “We tried something like this before and it didn’t work.”
3. Group the questions by topic. You will usually end up with a handful of clusters: price, integrations and setup, security and data, timelines and onboarding, proof, and comparisons with alternatives.
4. Rank the clusters by how often they appear. Write the entries for the frequent ones first. A library that answers the common questions well beats one that answers everything vaguely.

Step 2: Get answers from the people who own them

Every answer needs an owner: the person who can say “yes, we can say that.” For most teams that looks like this:

Ask each owner to write or approve their answers in plain language. Short beats complete. If an owner will not commit to an answer, that is useful too: the entry becomes “always escalate.”

If you run replies for clients, the client owns every one of these answers, and collecting them is part of onboarding. The client reply desk onboarding checklist has the full list of what to gather before the first reply goes out.

Step 3: Write each entry in a format the agent can use

A paragraph of facts is hard for an agent to apply correctly. A consistent structure is easier for the agent and for the person reviewing its replies. Each entry should have six parts:

1. Question pattern. The question in a few common phrasings, taken from real replies. “How much is it?”, “What does this cost?”, “Do you have pricing?”
2. Approved answer. One to three sentences, in plain language, with nothing the owner has not signed off.
3. When it applies. Segment, region, product or plan conditions. An answer that is true for small teams may not be true for enterprise accounts.
4. Never say. The specific claims to avoid. Discounts, delivery dates, features that are on the roadmap, results from other customers.
5. Escalate if. The conditions where the agent stops and hands off. “Asks for a contract redline”, “mentions a procurement deadline”, “existing customer.”
6. Next step. What the reply should ask for: a call, a document, a question back.

Here is the shape of one entry, as a template rather than real wording:

Question pattern: “Does it work with our CRM?” / “Can it sync to our CRM?”

Approved answer: Yes for the CRMs on our integrations page. Replies and meetings sync to the contact record.

When it applies: Any segment, listed CRMs only.

Never say: That an unlisted CRM is supported, or that a custom integration can be built.

Escalate if: The CRM is not on the list, or they ask about custom fields or data mapping.

Next step: Offer a short call to check their setup.

Keep entries short. If an entry needs more than a few sentences of approved answer, the question probably needs a person.

Step 4: Decide what the agent never answers on its own

Some questions should not get an AI answer however good the library is. Write these down as their own list, separate from the entries, so nobody has to infer them:

• Contract terms, legal language and anything that looks like a negotiation
• Specific security questionnaires, audits or data processing agreements
• Custom pricing, discounts and anything about budget approval
• Existing customers, strategic accounts and anyone on a do-not-automate list
• Complaints, hostile replies and anything that mentions a lawyer or regulator
• Requests to stop contacting, which need removal rather than a sales answer; see handling unsubscribe and opt-out requests

This list and the “escalate if” lines in each entry together make up your escalation policy. Our guide to AI reply agent human escalation rules goes deeper on the routing.

Step 5: Match the library to how much autonomy the agent has

The same library supports different operating modes. Decide which one you are in per topic, not for the whole inbox:

• Draft and approve. The agent drafts every reply, and a person approves before it is sent. This is where a new library should start.
• Send covered answers, approve the rest. Replies that only use approved entries go out on their own. Anything outside the library is drafted and held for approval.
• Send most, escalate the exceptions. The agent replies on its own and only the never-answer list reaches a person.

Most teams move pricing and security more slowly than integrations or scheduling, because the cost of a wrong answer is higher. AI inbox agent autonomy levels lays out the full ladder.

The approval step only works if it is fast. A draft that waits a day in someone’s email defeats the point of having an agent. Put approvals where your team already works: with Underfive, the inbox agent drafts the reply and sends it to Slack, where a person approves it before anything goes out. See Slack notifications for cold email replies for a clean channel setup.

Step 6: Test the library before it goes live

Before you switch anything on, run the library against replies you already handled:

1. Take a sample of recent replies that contain real questions.
2. Have the agent draft answers using only the library.
3. Have the answer owners review the drafts and mark each one as correct, wrong, or “should have escalated.”

Wrong answers usually mean an entry is vague or missing a “when it applies” condition. Missed escalations usually mean the “escalate if” lines are too narrow. Fix the entries and run the sample again. Do not move on until the owners are comfortable with what they see.

Step 7: Keep it current

A library that was right in spring can be wrong by autumn. Prices change, integrations launch, a security document gets updated, a case study gets withdrawn. A stale entry is worse than no entry, because the agent will repeat it with confidence.

Build three habits:

• Every edit in approval is a signal. When a reviewer changes a draft, ask whether the library caused the mistake. If it did, fix the entry the same day.
• Every escalation is a candidate entry. When the same question gets escalated repeatedly and the answer is always the same, write it into the library with its owner’s approval.
• Review on a schedule. Once a month, each owner reads their entries and confirms or updates them. Add a “last reviewed” date to every entry so stale ones are easy to spot.

Also tie library updates to the events that change answers: a pricing change, a new integration, a new security document, a product launch. Put “update the approved answers library” on the checklist for each of those.

Common mistakes

Pasting the website into the agent. It feels complete and produces replies full of claims that need context, or that are out of date.

Answers without boundaries. “Yes, we integrate with that” with no list of what “that” covers leads the agent to say yes to everything.

No owner. Entries without an owner never get updated, and nobody notices when they go wrong.

One library for every client or segment. Agencies and teams selling to very different segments usually need separate libraries, or at least separate “when it applies” conditions.

Measuring nothing. Track how often drafts in each topic are edited or escalated. Topics with frequent edits need better entries before the agent gets more autonomy.

Where to start this week

Pick the three questions your prospects ask most. Write one entry for each using the six-part format, get the owner to approve it, and run the agent in draft-and-approve mode on just those topics. Add an entry every time an escalation repeats. In a few weeks you will have a library that reflects what prospects actually ask, approved by the people who are accountable for the answers.

Underfive brings every mailbox and LinkedIn account into one unified inbox, with AI inbox agents that draft replies and send them to Slack for a person to approve before they go out. See how it works or book a demo.