小型企業網站 AI 聊天機械人指南：方案、成本與限制

The simplest safe way to put an AI FAQ chatbot on a small business website is to use a managed agent builder such as Microsoft Copilot Studio, point it at a handful of public pages you control, embed it with Microsoft's supplied snippet, and cap its spending. If you have a developer, a custom build with OpenAI's ChatKit is the other documented route. This guide compares both, explains how they are billed, and lists the limits and security steps you should settle before going live.

Product details were checked on Microsoft Learn and OpenAI's developer documentation on 28 October 2026. AI products change quickly, so confirm prices and menu names on the day you build.

Who this is for and what you will have at the end

This is for owners and office managers of small businesses who answer the same customer questions repeatedly (opening hours, delivery, returns, pricing structure), and for the IT person asked to "add a chatbot".

By the end you will have:

• A way to choose between a no-code agent builder and a custom build.
• A prepared FAQ knowledge base that suits either.
• A cost model you can calculate on paper before you spend anything.
• A short security and testing checklist for the launch.

Decide what the bot should and should not do

An FAQ bot is useful for repeat, factual questions with a single correct answer from your own content. It is a poor fit for anything that needs account access, a judgment call, or legal accuracy, such as refunds above a limit, medical or financial advice, or disputes.

Write down:

1. The 20 to 40 most common questions from your inbox, phone log or contact form.
2. The pages that answer each one. If no page answers it, write the answer first. A chatbot cannot be more accurate than its source.
3. The questions it must hand off to a person, and how (phone number, contact form, email).

If your answers are inconsistent between pages, fix the pages before building anything.

Option 1: Microsoft Copilot Studio

Copilot Studio lets you build an agent in a browser, ground it in knowledge sources and publish it to channels including a website.

What it can use as knowledge

Microsoft's knowledge sources page lists public websites, uploaded documents, SharePoint, Dataverse and connector-based enterprise data. For a public FAQ the relevant ones are public websites and uploaded documents. In generative mode, Microsoft lists support for up to 25 public websites. For websites you need to confirm which sites your organisation owns that Bing will search through Copilot Studio. Public website knowledge searches the query on Bing and only returns results from the websites you provide.

Two settings matter for accuracy:

• Allow ungrounded responses in the Generative AI settings. When it is off, the agent blocks a response generated in a turn where it used no knowledge source or tool, and the fallback topic triggers instead. Microsoft states this does not guarantee the agent never uses general knowledge, so it reduces rather than eliminates the risk.
• Use information from the web, which lets the agent search all public websites indexed by Bing. For a business FAQ, leave it off, because you want answers from your pages only.

You can also set a content moderation level from Lowest to Highest; the default is High.

Costs

Copilot Studio is billed by Copilot Credits. The billing page lists these rates for the standard harness:

Microsoft's own example for a customer support agent on a website is four classic answers and two generative answers per conversation, which comes to 8 credits per customer. At 900 customers a day that is 7,200 credits a day. Use that arithmetic with your own numbers. If you expect 30 conversations a day with three generative answers each, that is 30 x 3 x 2 = 180 credits a day.

This article does not list a price per credit pack because the Microsoft pages it draws on do not state one. Microsoft points to its Copilot Studio Licensing Guide and a usage estimator for current figures; use both before committing.

Licensing facts that affect a small business:

• The free trial lets you create agents and test them in the test chat panel, but you cannot publish the agent. When the trial expires you can extend it by 30 days, and the agent keeps working for up to 90 days after expiry.
• Microsoft says self-service trial sign-up may be disabled by your IT administrator, and that a personal email address can be rejected; use a work or school account.
• When consumption exceeds prepaid capacity, enforcement starts at 125 percent and custom agents are disabled until capacity is increased or reset. With pay-as-you-go linked to an Azure subscription, overage is billed instead. Administrators can set monthly limits for individual agents in the Power Platform admin center under Licensing, Copilot Studio, Manage Agents.

A disabled agent shows users a message such as "This agent is currently unavailable. It has reached its usage limit." Decide whether you prefer that or a surprise bill.

Publishing to your website

Microsoft's publishing article describes two things: a demo website for testing, which is not for production and whose URL you should not share with customers, and an iframe-based embed for the live site. In outline:

1. Open the agent, select Settings, then Security, then Authentication. For a public FAQ, Microsoft's demo steps choose No authentication.
2. Publish the agent.
3. Retrieve the agent's token endpoint (Microsoft documents this under customising the default canvas).
4. Use Microsoft's Web Chat sample page, replace the <YOUR TOKEN ENDPOINT> placeholder in the line const tokenEndpoint with your endpoint, test it in a browser, then give the file to your web developer.

The live-site sample is a full HTML page, so your developer will lift the relevant script and markup into your site template. Test on mobile as well, because the sample's chat window is sized for desktop and switches to full screen below 768 pixels wide.

Security

Microsoft's web channel security page says that without secured access, anyone who knows the agent ID can connect through the demo website and custom website channels. The Require secured access setting is under Settings, Security, Web channel security. Changes can take up to two hours to take effect. Credentials from this setting authorise access to the Direct Line channel; they do not authenticate the end user or restrict access to one website.

The rule that matters most: never place the Direct Line secret in code that runs in the browser. Exchange the secret for a token in your server code, using the generate-token request Microsoft documents. Tokens are valid for one conversation and expire unless refreshed.

Option 2: Custom build with OpenAI ChatKit

OpenAI's ChatKit documentation describes ChatKit as an embeddable chat user interface for agent experiences. There are two paths: a custom server integration, where you run ChatKit on your own infrastructure with the ChatKit Python SDK connected to an agent service, and an integration that uses a workflow hosted in Agent Builder.

The page states that Agent Builder is scheduled to shut down on November 30, 2026, and recommends the custom server integration for new work. That is a reason to avoid starting a new project on the hosted workflow path.

For the hosted path the documented flow is: your server creates a session by calling the ChatKit sessions endpoint with your API key, a workflow ID and a unique value for each end user, then returns a client secret to the browser. The API key stays on the server. The documentation says your server must authenticate users and supply the identifier.

Costs: the ChatKit page does not describe pricing, so check OpenAI's pricing page directly. Because the integration calls OpenAI's API with your account, plan to cap spend there with the controls your account offers.

This option suits a business with a developer or an agency already maintaining the site. For most small businesses with no developer, Option 1 is less work.

Option 3: Skip the AI

If the questions are few and stable, a well-organised FAQ page with a search box often does the job with no running cost and no risk of an invented answer. Consider it seriously if your top 20 questions fit on one page. It can also become the knowledge source for a bot later.

Prepare the knowledge base

Whatever you pick:

• Keep to public, approved content. Do not upload customer data, internal price lists or anything you would not publish.
• One question per heading, with the answer directly beneath it in plain sentences.
• Put dates, prices and policy numbers in one place so you update them once.
• Remove outdated pages from the sources the bot can search.
• Write instructions telling the bot to answer only from the provided pages, to say when it does not know and to offer your contact details. In Copilot Studio, Microsoft also recommends asking the model to include citations, because with ungrounded responses off, an answer without a citation can be withheld.

Related reading: an AI hallucinations verification checklist for business and an AI acceptable use policy template.

Test before launch

Build a test sheet of 40 real questions in three groups: the FAQ questions, questions the bot should refuse or hand off, and tricky ones with wrong assumptions ("Do you deliver on Sundays?" when you do not). Run each through the bot and mark every answer correct, partly correct, wrong or withheld. Fix the source pages for any wrong answer, not just the bot's instructions, then retest. Ask two colleagues to try to break it.

Also check:

• The handoff message works and shows the right phone number or form.
• Your privacy notice says that chats are processed by an AI service. Confirm what your own legal obligations are; this guide is not legal advice.
• Chat transcripts: decide who can read them and how long you keep them.

Troubleshooting

• The agent answers "I couldn't find that" for questions you know are on the site. Check that the site is in the knowledge list, confirm ownership as required, and add citation instructions. Intermittent withheld answers are a documented behaviour when ungrounded responses are off.
• The embed does not appear. Confirm the agent is published, the token endpoint is correct and your site's security settings allow the chat script to load.
• Costs are higher than forecast. Long conversations and extra features multiply credits. Cap usage per agent, shorten instructions and narrow the knowledge sources.
• The agent becomes unavailable. Enforcement may have triggered at 125 percent of prepaid capacity. Reallocate or buy capacity, or enable pay-as-you-go.
• Changes to security settings seem ignored. They can take up to two hours.

Launch checklist

• The 20 to 40 questions are answered accurately on public pages.
• Knowledge sources are limited to your own reviewed pages.
• Ungrounded responses and open web search are off.
• A spending cap or monthly limit is set.
• The secret stays on the server; the browser only receives tokens.
• A human handoff is visible in the chat.
• Test sheet passed on desktop and mobile.
• A named person reviews the transcripts weekly for the first month and fixes the sources.

References

• Knowledge sources summary, Copilot Studio: https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio
• Billing rates and management, Copilot Studio: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management
• Get access to Copilot Studio: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-licensing-subscriptions
• Publish an agent to a live or demo website: https://learn.microsoft.com/en-us/microsoft-copilot-studio/publication-connect-bot-to-web-channels
• Configure web and Direct Line channel security: https://learn.microsoft.com/en-us/microsoft-copilot-studio/configure-web-security
• ChatKit guide, OpenAI: https://developers.openai.com/api/docs/guides/chatkit

This guide was researched and drafted with AI assistance and checked against the vendor documentation listed in its References. Read the maintained version and related guides on EpiFive.