RAG Agents vs. a Standard FAQ Chatbot: How Do You Choose the Right One?

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Title: RAG Agents vs. a Standard FAQ Chatbot: How Do You Choose the Right One?

RAG 代理与标准 FAQ 聊天机器人：如何做出正确的选择？

Originally published at jy-labs.com. Updated 2026-10-07. A standard FAQ chatbot matches a customer’s question to a script someone wrote in advance, while a RAG agent retrieves the answer from your live documents and writes a sourced response on the spot. In October 2026 there is a third option between them: AI support agents built into Intercom and Zendesk, which run retrieval over your help center and bill $0.99 to $2.00 per resolved conversation. This post prices all three, shows where each one wins, and covers the rulings that make a wrong answer your legal problem.

本文最初发布于 jy-labs.com，更新于 2026 年 10 月 7 日。标准 FAQ 聊天机器人将客户的问题与预先编写的脚本进行匹配，而 RAG（检索增强生成）代理则从您的实时文档中检索答案，并当场撰写带有来源引用的回复。2026 年 10 月，两者之间出现了第三种选择：内置于 Intercom 和 Zendesk 中的 AI 支持代理，它们通过检索您的帮助中心内容来工作，并按每条已解决的对话收取 0.99 美元至 2.00 美元的费用。本文将对这三种方案进行定价分析，展示各自的优势，并探讨那些将错误回答转化为法律责任的裁决。

The problem Every vendor calls its product an AI agent, so a business owner comparing options sees near-identical claims on three systems that behave nothing alike in production. A scripted FAQ bot answers the twenty questions you gave it and nothing else. Intercom Fin and Zendesk AI agents answer from your help center and charge per resolution, so the bill grows with your volume. A custom RAG agent costs more up front, searches documents a help center cannot hold, and leaves you in control of what it refuses to answer. Pick on price alone or on a demo, and you pay twice: once for the wrong system and again for the one you needed first.

问题所在：每个供应商都称其产品为 AI 代理，因此企业主在比较选项时，会看到三个在实际生产中表现截然不同的系统，却有着近乎相同的宣传。脚本化 FAQ 机器人只能回答您预设的二十个问题，除此之外一概不知。Intercom Fin 和 Zendesk AI 代理根据您的帮助中心内容进行回答，并按解决次数收费，因此账单会随着业务量的增加而增长。定制的 RAG 代理前期成本较高，但能搜索帮助中心无法容纳的文档，并让您能够控制它拒绝回答的内容。如果仅凭价格或演示来选择，您最终会支付两次费用：一次是为错误的系统买单，另一次是为您真正需要的系统买单。

The approach What each system does under the hood A scripted FAQ bot runs on decision trees or keyword matching. A customer types a phrase, the bot finds the closest pre-written answer, and returns it word for word. Nothing gets generated. A RAG agent searches your documents for the passages relevant to the question, then writes a new answer from what it found, with a citation back to the source. One retrieves a fixed answer. The other retrieves raw material and writes.

方法：各系统的工作原理。脚本化 FAQ 机器人基于决策树或关键词匹配运行。客户输入短语，机器人找到最接近的预写答案并逐字返回，没有任何生成过程。RAG 代理则在您的文档中搜索与问题相关的段落，然后根据找到的内容撰写新答案，并附上来源引用。前者检索的是固定答案，后者检索的是原始素材并进行创作。

The vendor AI agents (Intercom Fin, Zendesk AI agents) are RAG agents too, with two constraints. The vendor controls the retrieval pipeline, and the knowledge sources are the ones the vendor supports, which in both cases centers on your help center plus connected files. Zendesk lists Google Drive and PDFs as external sources. Intercom lists third-party systems reached through APIs, data connectors, or MCP for actions like order lookups and refunds. If your answers live in a help center and a handful of connected tools, these products cover most of what a custom build did in 2024.

供应商 AI 代理（Intercom Fin、Zendesk AI 代理）本质上也是 RAG 代理，但有两个限制。供应商控制检索流程，且知识来源仅限于供应商支持的范围，通常以您的帮助中心及关联文件为核心。Zendesk 将 Google Drive 和 PDF 列为外部来源。Intercom 则列出了通过 API、数据连接器或 MCP 访问的第三方系统，用于执行订单查询和退款等操作。如果您的答案存储在帮助中心和少数几个关联工具中，这些产品涵盖了 2024 年定制开发方案的大部分功能。

What off-the-shelf support agents do and cost in October 2026 List prices from the vendors’ own pricing pages this week: Intercom Fin. $0.99 per outcome. A resolution counts when no further help is requested after Fin’s last answer. Conversations Fin hands to your team with no outcome are free. Lead qualifications bill at $9.99 each. On Intercom’s own helpdesk, seats start at $29 per month. Fin also runs on Salesforce, HubSpot, Freshworks, Zoho, Gorgias, and others with no seat cost, with a 50-outcome monthly minimum. Copilot for human agents is $35 per user per month. Intercom reports an average resolution rate of 76 percent across 12,000+ customers, which is a vendor figure, not an audit.

2026 年 10 月现成支持代理的功能与成本。本周来自供应商定价页面的标价如下：Intercom Fin，每次结果收费 0.99 美元。当 Fin 最后一次回答后客户未再寻求进一步帮助，即计为一次解决。Fin 转交给团队且未达成结果的对话免费。潜在客户资格认证每次收费 9.99 美元。在 Intercom 自己的帮助台，席位费每月 29 美元起。Fin 也可在 Salesforce、HubSpot、Freshworks、Zoho、Gorgias 等平台上运行，无需席位费，但每月至少需达到 50 次结果。人工坐席的 Copilot 每用户每月 35 美元。Intercom 报告称其 12,000 多家客户的平均解决率为 76%，这是供应商提供的数据，而非审计结果。

Zendesk AI agents. Included in every Suite and Support plan, billed on successful outcomes. Suite Team is $55 per agent per month billed yearly and includes 5 automated resolutions per agent per month. Suite Professional is $115 and includes 10. Beyond the allowance, committed resolutions cost $1.50 and pay-as-you-go resolutions cost $2.00. Copilot is $50 per agent per month on Professional and above. Zendesk’s own help article defines a billable resolution as one the AI handled without escalation, verified by an LLM, counted 2 hours after the first message on messaging and 72 hours after the first email.

Zendesk AI 代理。包含在所有 Suite 和 Support 计划中，按成功解决次数计费。Suite Team 计划每年结算时每坐席每月 55 美元，包含每坐席每月 5 次自动解决额度。Suite Professional 计划为 115 美元，包含 10 次额度。超出额度后，预付解决次数费用为 1.50 美元，按需付费为 2.00 美元。Professional 及以上版本的 Copilot 每坐席每月 50 美元。Zendesk 的帮助文档将可计费的解决定义为：AI 在无需人工介入的情况下处理完成，经 LLM 验证，且在消息传递中首次消息 2 小时后或电子邮件首次发送 72 小时后统计。

Worked example. Say you handle 1,000 support conversations a month and the agent resolves 70 percent, so 700 resolutions. On Fin that is $693 in outcome fees plus seats. On Zendesk Suite Team with three agents, you pay $165 for seats, get 15 resolutions included, and pay $1,027.50 for the remaining 685 at the committed rate, so about $1,193 a month. Both meters have the same property: the better the agent performs, the bigger the bill. Budget on resolutions, not conversations.

案例演示。假设您每月处理 1,000 次支持对话，代理解决了 70%，即 700 次解决。使用 Fin，费用为 693 美元的结算费加上席位费。使用拥有三名坐席的 Zendesk Suite Team，您需支付 165 美元的席位费，包含 15 次解决额度，剩余 685 次按预付费率支付 1,027.50 美元，总计每月约 1,193 美元。两种计费方式都有一个共同点：代理表现越好，账单越高。请按解决次数而非对话次数来做预算。

Where a scripted bot still wins For ten to twenty stable questions (store hours, return policy, shipping cost), a scripted bot is cheap, launches in a week or two, and is hard to get wrong, because a human wrote every word in advance. A wrong answer is impossible unless a human typed it. If your support volume stays low and your answers rarely change, a per-resolution meter and a retrieval pipeline both add more than the problem needs.

脚本化机器人的优势所在。对于十到二十个稳定的问题（如营业时间、退货政策、运费），脚本化机器人成本低廉，一两周即可上线，且不易出错，因为每一个字都是人工预先编写的。除非人工输入错误，否则机器人不可能给出错误答案。如果您的支持需求量小且答案很少变动，那么按解决次数计费和检索流程都会增加不必要的复杂性。

Where a scripted bot breaks The moment a customer phrases a question a way nobody scripted for, the bot returns the closest generic match or admits it cannot help. Businesses with large or changing document libraries (product catalogs, policy manuals, technical documentation) hit this wall every week. Every new product, policy change, or edge case means another manual script update, and the backlog of unscripted questions grows faster than your team writes for it. Vendor AI agents fixed this for the help-center case. The next section covers the cases they did not fix.

脚本化机器人的局限性。一旦客户提出的问题方式超出了脚本预设范围，机器人就会返回最接近的通用匹配项，或者承认无法提供帮助。拥有庞大或不断变化的文档库（产品目录、政策手册、技术文档）的企业每周都会遇到这个问题。每一个新产品、政策变更或边缘情况都意味着需要手动更新脚本，而未被脚本覆盖的问题积压速度往往超过了团队的编写速度。供应商 AI 代理解决了帮助中心场景下的这一问题。下一节将介绍它们未能解决的情况。

When a custom RAG agent still earns its cost A custom build at JY Labs falls in the Build Sprint tier, $8K to $15K one time, with optional retention at $499 to $999 a month. A typical deployment runs six to seven weeks, with a working prototype by week three. Against the Zendesk example above, that is about one year of the meter at 1,000 conversations a month. The custom build wins in five situations: Your answers do not live in a help center. Contracts, scanned PDFs, an ERP, a ticket archive, a shared drive with 50,000 files. A regional law firm put 50,000+ contracts, briefs, and case files behind a RAG agent and cut attorney research time by 80 percent, from 3.2 hours to 38 minutes a day. No support-desk product indexes that corpus. The users are your staff, not your customers. Per-resolution pricing assumes a support ticket. An internal knowledge agent for onboarding, policy lookups, or sale

定制 RAG 代理的价值所在。在 JY Labs 进行定制开发属于“构建冲刺”阶段，一次性费用为 8,000 至 15,000 美元，可选的维护费用为每月 499 至 999 美元。典型的部署周期为六到七周，第三周即可提供工作原型。与上述 Zendesk 的例子相比，这大约相当于每月 1,000 次对话一年的计费成本。定制开发在以下五种情况下更具优势：您的答案不在帮助中心内。例如合同、扫描的 PDF、ERP 系统、工单存档或包含 50,000 个文件的共享驱动器。一家区域性律师事务所将 50,000 多份合同、摘要和案件档案接入 RAG 代理，将律师的研究时间缩短了 80%，从每天 3.2 小时减少到 38 分钟。没有任何支持台产品能索引此类语料库。用户是您的员工而非客户。按解决次数定价的前提是存在支持工单。用于入职培训、政策查询或销售的内部知识代理。