1. Home 
  2. > Blogs 
  3. > Practical Guide

AI Agent Knowledge Base for B2C Teams: Stop AI Quoting Old Prices and Policies

George Wong

Β·

9 min read
AI Agent Knowledge Base: Stop AI Quoting Old Prices

TL;DR: Respond.io AI Agents recommend products and confirm prices only from knowledge sources that resync on a schedule

  • The problem: AI agents quote old prices and policies when they answer from a file uploaded once, a website read once or general internet knowledge.

  • The fix: respond.io's AI Agents reply, recommend products and confirm prices using only the knowledge sources you add to each agent. Website sources resync automatically on a schedule you set, and you can name the source the AI Agent checks first, such as your latest price list.

  • Missing answers go to a person: when the answer isn't in its sources, the AI Agent hands the chat to your team instead of guessing.

  • Best fit: mid-market B2C businesses handling 500+ customer chats a month whose prices, stock or policies change at least monthly.

  • Not a fit: teams whose prices, policies and catalog rarely change and where a person already answers every price question within minutes.

For mid-market B2C businesses, the decision is whether your AI agent answers from knowledge that stays current, and whether respond.io, a customer conversation platform with built-in AI Agents, is the way to get there.

When an AI agent quotes last month's price, a teammate has to step in and a sale that should close in one reply needs three. The AI agent usually works. The knowledge behind it is out of date. A wrong answer costs twice: once in the sale that stalls, and again in the time your team spends fixing the reply.

Why AI agents quote old prices and policies, and when to replace yours

AI agents give outdated or wrong answers for four reasons. Each one starts with how the agent was set up, not with the AI model itself.

  1. The file was uploaded once. Many teams train their AI agent on a PDF of FAQs or a price list. When prices change, the PDF stays the same. The AI keeps quoting the old figures until someone remembers to upload a new file.

  2. Website changes never reach the AI. Your team updates prices, promotions and stock on the website. The AI agent read the site once, during setup. It has no way of knowing the page has changed.

  3. The AI fills gaps with general knowledge. When the uploaded file has no answer, some AI agents fall back on what the model learned from the internet. The reply sounds confident and has nothing to do with your business. Businesses evaluating AI agents regularly ask for the opposite: an agent that gives the proper information, not random answers.

  4. Too many overlapping sources confuse the agent. A catalog, a staff training document and a long set of instructions often cover the same products in different words. With no rule for which source wins, the AI gives a generic answer or mixes old and new details.

The cost shows up in two places. Customers told the wrong price or the wrong stock level either leave or ask you to honor what the AI said, which lowers conversion. Teammates spend their day checking and correcting AI replies, so the cost of handling each conversation rises instead of falling.

Two signs tell you it's time to replace your current setup:

  • A teammate has corrected an AI reply about a price, promotion, stock level or policy in the last week.

  • A price change made on your website is still missing from the AI's answers a day later.

If either is true, the problem is the knowledge setup. Changing the prompt will not fix it.

How respond.io AI Agents answer only from your approved knowledge sources

Respond.io is a customer conversation platform that includes AI Agents. Those agents pull only from approved knowledge sources, and the way they do it answers failure modes three and four above.

Answers come only from the sources you add to each AI Agent

When you create or edit an AI Agent, you choose which of your workspace's knowledge sources it can use. Only those selected sources are used to generate replies to contacts. Sources can be your website, documents (.pdf, .txt, .md, .csv, .docx, .pptx and .ppsx files) or both.

Per-agent sources keep knowledge separate where it should be. A sales AI Agent can answer from your price list and product catalog. A support AI Agent can answer from your return policy and warranty terms. Neither AI Agent sees the other's documents, so a sales question isn't answered from your warranty terms.

Diskat, a retailer of electronics and household goods in Guatemala, runs its sales conversations this way. Its AI Agent, Diky, "responds instantly to product questions, collects essential customer details such as name, address and phone number, confirms prices and delivery times", and passes the chat to a person only for logistics or tracking. AI Agent Diky now handles 90% of Diskat's sales conversations end to end, and Diskat keeps an 81.4% conversion rate.

RAG retrieves the source before the reply is written

Respond.io's AI Agents use retrieval-augmented generation (RAG). In plain terms, the AI Agent first finds the passages in your knowledge sources that match the customer's question. It then writes the reply from those passages. RAG grounds the response to reduce hallucinations, meaning answers the AI makes up.

The order matters. An AI agent that writes first and checks later can still produce a confident guess. An AI Agent that retrieves first has your current wording in front of it before it writes a word. Each task is also handled by the most suitable AI model rather than one model for everything. Replies come back in the same language the customer wrote in.

Missing answers go to a human, not a guess

Retrieval only helps if the AI Agent knows what to do when nothing matches. Respond.io's AI Agents hand the conversation to a human instead of guessing when the answer isn't in their sources. Guardrails set what each agent can and can't say, and sensitive topics or low-confidence replies trigger escalation.

AI Agents hand chats to human teams by role, online status or workload. Set up the Add comments AI Skill once and the AI Agent leaves an internal note on what the customer asked with every handoff. Any teammate can take over a conversation at any time. A gap in your knowledge becomes a handoff the customer barely notices, not a wrong answer they act on.

For a shorter answer to the same question, see whether respond.io's AI Agents give wrong or made-up answers.

How respond.io keeps AI Agent knowledge current when prices and stock change

Respond.io keeps AI Agent knowledge current in four ways: website sources resync on a schedule, you replace files when a price list changes, you name the source the AI Agent checks first, and HTTP requests check live stock. None of them needs anyone to retrain the AI Agent.

These fix failure modes one and two. Grounding stops the AI from making things up, but it doesn't stop the AI from quoting a source that is out of date.

Website sources resync on a schedule

A website knowledge source can be resynced on demand or set to an automatic sync schedule, for example weekly or monthly. When your team changes a price or a policy on the website, the next sync replaces the old content in the AI Agent's knowledge.

For urgent changes, such as a flash sale that starts in an hour, a manual Resync pushes the update straight away. The crawler goes three levels deep into your site by default and can be set up to 100 levels, so product pages several clicks from the homepage are included.

Without a schedule, a website source is just as static as an uploaded PDF. It reflects the site on the day someone added it.

Replace files when the price list changes

Documents work differently from website sources. When your price list or policy document changes, you replace the uploaded file with the new version and remove the old one.

Leaving both versions in place is how most outdated answers start. The AI Agent has two price lists, both look valid, and nothing tells it which is current. Replacing the file removes that choice.

Name the source the agent checks first

Respond.io's AI Agents don't prioritize knowledge sources on their own. When the instructions don't name a source for a situation, the AI Agent searches all of its sources. That is where an old catalog can outrank a new price list.

You control this in the AI Agent's instructions. Name a source for a type of question, such as "use the Current Price List source for pricing questions", and the AI Agent searches only the named sources. If any question in a customer's message has a named source, the AI Agent searches only those sources for the whole message. Customers often ask two things at once ("how much is it and do you deliver on Sunday?"), so it pays to name the source for each common question type.

Picture a parts distributor with a manufacturer's spec sheet and its own product pages. It can name the spec sheet for technical questions and the website for prices. The AI Agent then checks the right document for each kind of question instead of blending them.

Check live stock with HTTP requests

Some information changes too often for any sync schedule. Stock levels, order status and appointment slots can change by the hour.

For these, respond.io's AI Agents can check live stock by sending HTTP requests to your inventory system during the conversation. Only Tourism, a travel agency in Dubai, uses its AI Agent to answer visa requirement questions and to retrieve each applicant's visa application status through an API integration. The AI Agent now handles 80% of Only Tourism's visa inquiries.

This needs an API on your side. If your stock data has no API, the AI Agent can only answer from your last synced source. Plan that limit in: either sync the stock page often or have the AI Agent pass stock questions to a person.

How to check which source an AI Agent answer came from before customers see it

The fastest way to trust an AI Agent is to see where each answer came from. When you test an AI Agent in respond.io, replies can show a "{#} sources" label, so you can verify which knowledge sources were used to write each one.

A wrong source points straight to the fix. If a pricing answer cites last year's catalog, you replace the file or name the current price list in the instructions. You don't need to guess why the AI said what it said.

Before you put an AI Agent on price and policy questions, run a short check:

  1. Ask the five price, stock and policy questions your customers send most often.

  2. Read the source label on each reply and confirm it came from the current document.

  3. Fix the source or the instruction, then ask the same questions again.

After launch, keep an eye on sync status. A website source can show "Resync failed" when the last sync didn't finish. The AI Agent still uses the older content in that case, so a failed sync is worth checking the same day.

Who should move AI answers to respond.io, and who shouldn't

The mechanisms above matter most when your information changes often and your AI answers customers on several channels.

Move AI answers to respond.io if any of these apply:

  • You are a mid-market B2C business whose prices, promotions, stock or policies change at least monthly.

  • Your customers ask about prices, availability and policies on WhatsApp, Instagram, Facebook Messenger and other channels, and you want every channel answered from the same set of knowledge sources.

  • Your team already corrects AI replies, or answers the same price questions by hand because they don't trust the current bot.

Keep what you have if either of these applies:

  • Your catalog and policies rarely change, and a person answers every price question within minutes. A lightweight inbox or WhatsApp Business app replies may be enough.

  • Every answer depends on live data and you have no API to expose it. The AI Agent can only answer from the last sync, so the accuracy problem moves rather than disappears.

Plan for the setup work. Accurate answers take some preparation. Split knowledge sources by topic so each one covers one subject. Strip footers, disclaimers and unrelated text before uploading. Name sources in the AI Agent's instructions for your most common questions. Give one person ownership of the sync schedule and the source list. Live stock or order lookups need developer time to connect your system.

When to put respond.io AI Agents on price and policy questions

Act now if a teammate corrected an AI reply about price, stock or policy in the last week, or a website change was still missing from AI answers a day later, and your team handles 500+ customer chats a month. Respond.io's scheduled resync keeps website sources current, named sources in the instructions keep the latest price list in charge of pricing questions, and the handoff covers anything the sources don't answer.

Wait if your prices rarely change and a person already answers every price question within minutes. Track how often your team corrects or repeats price answers. Once that becomes a weekly task, the case for synced knowledge is made.

Rule respond.io out for this job if every answer needs live data you can't expose through an API.

To start, list the questions your team corrects most often. Add the sources that answer them to an AI Agent, test with the source label and turn on a resync schedule. You can create a respond.io account and run that test on your own price list. For a broader comparison of options, read our guide to the best AI agents for B2C customer conversations.

Turn customer conversations into business growth with respond.io. ✨

Manage calls, chats and emails in one place!

FAQs about AI agent knowledge bases

How often do respond.io AI Agents resync website knowledge sources?

You set the schedule when you add or edit a website source, for example weekly or monthly. You can also click Resync at any time to pull in an urgent change, such as a new promotion.

Resync is for website sources. To update a document, replace the uploaded file with the new version and remove the old one.

Can an AI agent answer with live stock levels instead of the last synced price list?

Yes, if your stock data has an API. Respond.io's AI Agents can send HTTP requests to your inventory system during the conversation and answer with the current figure.

Without an API, the AI Agent answers from your last synced source. In that case, sync the stock page often or have the AI Agent pass stock questions to your team.

What does a respond.io AI Agent do when the answer isn't in its knowledge sources?

It hands the conversation to a person instead of guessing. Sensitive topics and low-confidence replies also trigger escalation.

Your team picks up the chat with an internal note on what the customer asked, once you've told the AI Agent to leave handoff notes.

Can each AI agent have its own knowledge base?

Yes. Each respond.io AI Agent uses only the knowledge sources you add to it. A sales AI Agent can answer from your price list and catalog, while a support AI Agent answers from your return and warranty policies.

The same split works on any channel, including a WhatsApp AI agent for B2C customer conversations.

Does an AI agent knowledge base replace my CRM or product database?

No. Respond.io is a customer conversation platform that includes AI Agents and CRM features. Your product database and CRM stay the source of record for products, stock and customer data.

The knowledge base is what the AI Agent reads to answer customers. Website sources resync from your site, and HTTP requests can check live data in your own systems.

Can my team correct an AI Agent answer in the same inbox?

Yes. Respond.io is a customer conversation platform, and its shared inbox is one part of it, alongside AI Agents and CRM features. AI Agent conversations sit in that inbox alongside your team's. Any teammate can take over a conversation at any time.

If a correction reveals an outdated source, update or replace that source so the AI Agent answers correctly next time.

Do AI Agents use the same knowledge on WhatsApp, Instagram and Messenger?

Yes. Respond.io is a customer conversation platform that includes AI Agents, and it is not a WhatsApp-only tool. Its AI Agents work across WhatsApp, Instagram, Facebook Messenger, TikTok, Telegram, Viber, LINE, RCS and other channels, and each AI Agent answers from the same knowledge sources wherever it runs.

Customers get the same price and policy answer wherever they ask.

Share this article
Telegram
Facebook
Linkedin
Twitter
George Wong
George Wong
George Wong is a Communications Strategist at respond.io with deep experience in growth and product marketing. Since joining the company as a Content Manager in 2022, he has helped shape the go-to-market strategy for key product launches, refined messaging across channels and driven brand positioning through content and campaign initiatives. George specializes in turning complex product features into compelling narratives that drive business impact.
3x Your Business Results with Respond.io πŸš€