Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
Platform Features

AI Agents That Search the Web: When to Use Live Answers

AI SNS CHAT EditorialPublished 7 min read

How web-searching AI agents answer questions your knowledge base can't, with guardrails, cost examples and how AI SNS CHAT handles live web search.

Every support knowledge base goes stale. A vendor changes a setting, a carrier updates a policy, or a customer asks about something you never wrote down, and your AI agent either guesses or says "I don't know". Letting an agent search the web can close that gap, but only if you decide carefully where it may look and when it may answer. This article explains how web-searching agents work, where they help, where they hurt, and how to set one up in AI SNS CHAT.

Why a knowledge base alone goes stale

A knowledge base is a snapshot. You upload help articles, a price list and a returns policy, and the agent answers from them. That is the right default: answers are consistent, on-brand and under your control.

The weakness is time. Anything outside your documents is invisible to the agent, and anything inside them is only as fresh as your last update. Typical gaps include:

  • Third-party product questions, such as "how do I reset this setting in the operating system your software runs on?"
  • Fast-changing public information, like a vendor's current instructions or a public notice.
  • Long-tail questions that no one thought to document, because only one customer in a hundred ever asks them.

Without a fallback, the agent has two bad options: invent something plausible, or refuse. A web search gives it a third option: look it up, answer from a real page, and say where the answer came from.

What "searching the web" means for an agent

Web search for an agent is not the model "knowing" the internet. The model has a training cutoff, so on its own it cannot know what changed last week. Search adds a step at answer time:

  1. The customer asks a question.
  2. The agent checks its knowledge base first.
  3. If nothing relevant is found, it sends a search query to a search provider.
  4. It reads the returned pages or snippets and writes an answer based on them.
  5. Ideally, it cites the page it used so the customer can verify.

There are two broad designs, and the difference matters.

Fallback search

The agent searches only when the knowledge base has no answer. Your own documents always win. This keeps the agent predictable and keeps cost down, because most questions never trigger a search.

Search as a tool

The agent decides for itself, mid-conversation, that it needs fresh information and calls a search tool. This is more flexible, good for questions such as "is this service down right now?", but you give up some predictability, so it needs clearer instructions about when to use it.

  1. 01Customer asks a question
  2. 02Agent checks your knowledge base
  3. 03No answer found, so it searches approved sources
  4. 04Reply cites the page used
Fallback web search, step by stepIllustrative example

When live search helps, and when it does not

Search is a tool for specific jobs. Use this comparison to decide.

Good fit
Poor fit
Third-party software or device how-to questions
Your prices, discounts and contract terms
Public information that changes often
Refund, warranty or legal policy
Long-tail questions outside your documents
Anything about a customer's own account or order
Checking a vendor's current instructions
Medical, legal or financial advice
Where web search fitsIllustrative example

The rule of thumb: your own business facts should come from your own sources. If a customer asks what your return window is, an answer pulled from a random web page is a liability, even if it happens to sound right. Web search is for facts that live outside your walls.

Guardrails that keep web answers trustworthy

An agent with unrestricted web access is a different risk from an agent restricted to a short list of sites. Four controls matter most.

Restrict the sources

Search results include forums, outdated blog posts and marketing pages. Limiting the agent to official support sites for the products your customers actually use removes most of the noise. A narrow, authoritative list beats the open internet for support work.

Keep your own topics off the web

Define a list of subjects the agent must never look up externally: your pricing, your policies, your competitors, your internal processes. For those, the knowledge base is the only source, and a missing answer should lead to a human handoff, not a search.

Cite the source

A reply that says "according to this page" lets the customer check it, and lets your team audit it later. If the agent cannot point to a page, it should not present the answer as fact.

Plan for the miss

Sometimes the search returns nothing useful. The agent should say so plainly and offer a handoff to a person. A graceful "I couldn't confirm that, let me bring in a colleague" is much better than a confident wrong answer.

What it costs: credits, latency and a worked example

Web answers cost more than knowledge-base answers, in both money and time. The search call adds a delay, and many platforms charge more for a reply that used live data. It is worth modelling this before you switch it on.

Here is a hypothetical. Suppose a shop receives 300 chats a month, and 60 of them end up needing a web lookup. If a normal reply costs 1 credit and a web-answered reply costs twice that, the month looks like this:

Knowledge-base replies (240 x 1)
240
Web-answered replies (60 x 2)
120
Total credits
360
Example: credits used by 300 monthly replies (hypothetical)Illustrative example

In this example, 20% of the replies drive one third of the credit use. That is not a reason to avoid web search; it is a reason to keep it for the questions that need it, and to keep improving the knowledge base so that repeat questions stop reaching the web at all. A good habit: review which questions triggered a search each month, and turn the common ones into Q&A entries.

Rolling it out safely

A simple sequence keeps risk low.

  1. Start with a short source list. Pick the two or three official sites your customers ask about most.
  2. Test in a sandbox first. Ask the questions you collected, plus a few trick questions about your own pricing and policies. Confirm those never go to the web.
  3. Check the citations. Open every cited page and confirm it really supports the answer.
  4. Go live on one channel. Website chat, WhatsApp or Messenger are all good first places, but pick one so you can watch the results.
  5. Review weekly at first. Read the conversations where search was used, rate good and bad answers, and adjust the source list.
  6. Convert repeats into knowledge. Anything asked more than a few times belongs in your own data sources.

Doing this in AI SNS CHAT

AI SNS CHAT offers two routes to live web answers, and they work differently, so it helps to know which you need.

Built-in web search on LINE WORKS

For agents deployed on LINE WORKS, there is a web search setting under Behaviour → Web search. It works as a fallback: when the knowledge base has no answer, the agent can look the question up on the web. Its behaviour is deliberately conservative:

  • It defaults to Microsoft and Apple support sites.
  • It never searches for in-house topics.
  • It cites the page it used.
  • A reply answered from the web costs 2x credits by default.

The search providers behind this feature are configured by the platform administrator, not on the customer side. If you want to know which provider is in use, or you need the credit multiplier changed, that is a conversation with the platform admin.

The Tavily action for any agent

If your agent runs on another channel, such as the chat widget, WhatsApp or Messenger, the Tavily action gives any agent real-time web search. You add it from the Actions / Integrations area, alongside the other built-in actions such as Collect leads, Stripe, Shopify, Slack, Cal and Calendly. Because it is an action, the agent decides when to use it, so describe in the agent's persona exactly what it is for and what it must not look up.

Plan gates apply. Integrations and actions need the Hobby plan or higher, and the number of actions per agent grows with the plan, from 5 on Hobby up to 20 on Corporate. The Free plan has none. See /pricing for current plans.

A sensible setup

  1. Build your knowledge base first, under Data sources: files, text snippets, your website, Q&A, or connectors such as Google Drive, Confluence or Notion. Web search should cover only what these cannot.
  2. In Playground, test the questions you collected, including ones about your own pricing, and use the A/B model compare if you want to see how two models handle live results.
  3. Add the persona rules: what the agent may look up, what it must hand off, and that it should name its source.
  4. Set up human handoff under Helpdesk → Scheduling so that a miss reaches whoever is on shift.
  5. After going live, read Activity → Chat logs, use thumbs up and down, and watch credits per agent on the Usage page.

Web search in AI SNS CHAT is not a single switch that works identically everywhere. Check which channel you deploy on, and choose the built-in LINE WORKS search or the Tavily action accordingly. More detail is on /features, and the channel list is on /connections.

Frequently asked questions

Will a web-searching agent answer questions about my own prices and policies from the internet?

It should not. On LINE WORKS, the built-in search never looks up in-house topics. With the Tavily action, you control this through the agent's persona, so state clearly that pricing, policies and account questions come only from your knowledge base or a human.

Does web search work on WhatsApp, Messenger and the website chat widget?

The built-in Behaviour → Web search setting applies to LINE WORKS only. For other channels, add the Tavily action to the agent, which gives it real-time web search. Which plan you need depends on the Actions limits, and integrations require Hobby or higher.

How much more does a web-answered reply cost?

On the built-in LINE WORKS search, a reply answered from the web costs 2x credits by default, where a normal reply is 1 credit. Your credit use per agent is visible on the Usage page, so you can see the effect after a few weeks.

Can I choose which websites the agent searches?

The built-in LINE WORKS search defaults to Microsoft and Apple support sites. The search providers themselves are set by the platform administrator. For the Tavily action, guide the agent through its persona instructions about which kinds of sources to rely on.

#web search#AI agents#knowledge base#customer support#platform features

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.