The AI customer support agent that shows its work
HelpShelf answers your customers from your own content — with citations on every confident answer — and offers a verified human path when it should. This page explains what an AI customer support agent actually is, how one works, and how to pick one without getting burned.
What is an AI customer support agent?
An AI customer support agent is software that answers customer questions automatically by retrieving relevant information from a company's documentation and generating a response. When it cannot verify the result, a responsible agent stops clearly and can offer a user-controlled human path instead of claiming a person was contacted automatically.
How an AI customer support agent works
Three stages: ingest, answer, stop or route. Where a vendor cuts corners on one of them is where the product bites you later.
1. Ingest your knowledge
You point the agent at what you already have: your website, help docs, existing help desk articles, or your codebase — HelpShelf can build the knowledge base from a repo with one npx command. Content is classified by trust: publisher-authored and imported sources rank above crawled original pages, and AI-drafted articles are clearly separated and held for review.
2. Answer with citations
When a customer asks a question, the agent retrieves the most relevant passages using hybrid search — exact keyword matching plus semantic similarity — and generates an answer grounded in them. Every confident HelpShelf answer carries numbered citations back to the source, so customers (and you) can verify it instead of taking a bot at its word.
3. Stop safely, then offer help
No agent knows everything, and the good ones know it. When confidence is low, HelpShelf stops guessing. If your tested Telegram destination is reachable, it offers an editable handoff; only after the visitor confirms does Telegram receive bounded conversation context.
What to look for in an AI customer support agent
Six questions that separate an agent you can trust from an expensive way to frustrate customers.
Citations and proof
Ask to see where an answer came from. If the agent cannot show its sources, you cannot audit it and your customers cannot trust it. Citations turn "the bot said so" into "the docs say so — here is the link."
Escalation quality
Every vendor demos the happy path. Test the unhappy one: ask something the bot cannot know and watch what happens. Look for a clear stop, a visitor-controlled path to a person, and an honest delivery receipt — not a loop or a silent claim that somebody was contacted.
Pricing model
Seat pricing punishes hiring, per-resolution pricing scales with volume, and metered conversations can bill you for noise. Whatever the model, compute your cost per resolved conversation and check exactly how human-routing attempts, failures, and spam are metered.
Control over sources
You decide what the agent answers from. Look for a real separation between published source content, crawled original pages, and AI-drafted articles — and the ability to keep unreviewed content out of sensitive answers entirely.
Setup and maintenance
The best agent is the one that is actually live. A script tag and a content scan beat a six-week bot-flow implementation, and keeping the agent current should be a re-sync, not a project.
Legibility to AI assistants
Your buyers increasingly ask ChatGPT, Claude, and Perplexity about your product before they ever visit. A platform that publishes llms.txt and an MCP endpoint lets those third-party AIs answer from your real docs instead of guessing.
Will an AI agent replace my support team?
No — and you should distrust any vendor that says otherwise. An AI customer support agent handles the repetitive layer: the password resets, the where-is-this-setting questions, the pricing clarifications that make up most of your volume. What remains is the work that actually needs judgment — angry customers, edge-case bugs, refunds, accounts on fire. Your team ends up doing more of that work, not less of it.
It is also worth naming the elephant: people are wary of support bots, and the wariness is earned. Ask customers what they hate about automated support and the answers are consistent — not that a bot answered, but that the bot was wrong, or trapped them in a loop, or refused to hand them to a person. The fear is not automation. It is bad automation with no exit.
That is a design problem, and it is solvable: require real documentation for each complete answer, show citations so answers can be verified, keep a visible path to visitor-confirmed human help when a tested destination is reachable, and hold anything uncertain for review instead of sending it. That is how HelpShelf is built — the agent earns trust by proving its answers, and by knowing when to stop.
AI agent vs. traditional chatbot
The word "chatbot" covers two very different products. Here is the practical difference.
| Feature | HelpShelf | Scripted chatbot |
|---|---|---|
| Answer source | Your docs, site, and codebase — retrieved per question | Scripted flows and canned replies |
| Proof | Citations on every confident answer | |
| Unknown questions | Safe stop; optional visitor-confirmed handoff | Loops, rephrases, or guesses |
| Human handoff | Visitor-confirmed via verified Telegram | |
| Sales questions | Answers, then offers visitor-controlled follow-up | |
| Legible to AI shoppers | llms.txt + per-site MCP server | |
| Setup | One script tag, or npx helpshelf from a repo | Flow-builder configuration |
Where HelpShelf fits
HelpShelf is our product, so read this section knowing that — but the fit is specific enough to state plainly. HelpShelf is an AI customer support agent for small self-serve SaaS teams: it answers from your content with citations on every confident answer, withholds unsupported widget answers, and offers an editable, visitor-confirmed handoff when your tested Telegram destination is reachable.
The same agent also works the sales side: it answers presale questions on your pricing page, captures a lead after a confident answer, and reports the buying questions your visitors are asking. And it makes your product legible to AI shoppers — every site gets an llms.txt file and its own MCP server, so ChatGPT, Claude, and Perplexity can answer questions about your product from your real docs.
- Free plan — no credit card required
- Pro is $49/mo with 10,000 AI searches and 2,000 AI chats each month
- Scale is $149/mo with 100,000 AI searches and 20,000 AI chats each month
- Rated 4.7 on Capterra
- Live via one script tag — or built from your codebase with npx helpshelf
Go deeper
Researching the category? These guides cover the comparisons and adjacent tools.
Best AI agents for customer support
Nine vendors compared honestly — pricing models, categories, and who each one actually fits.
AI customer support software: buying guide
The three categories of tools, an evaluation checklist, and how pricing converged on outcomes.
AI knowledge base
The knowledge base that answers instead of storing — hybrid search, trust tiers, citations.
AI help desk
The help desk where supported widget answers cite their sources and visitors control when bounded context is sent to tested Telegram.
Answers that sell
The same agent on your pricing page: cited presale answers, proactive engagement, lead capture.
HelpShelf vs Intercom
How a usage-limited agent compares to the best-known name in the category.
Frequently asked questions
An AI customer support agent is software that answers customer questions by retrieving relevant information from documentation and generating a grounded response. When it cannot verify an answer, a responsible implementation stops and may offer a user-controlled route to a human instead of claiming that somebody was contacted automatically.
No. It handles the repetitive layer — password resets, how-do-I questions, pricing clarifications — while judgment, empathy, and account access still require people. Evaluate exactly how each product offers human help, what the visitor must confirm, and whether context and delivery state are preserved.
A traditional chatbot follows scripted flows: if the visitor says X, show Y. An AI agent retrieves relevant passages from your actual content for each new question and generates an answer from them. The practical differences are coverage (no flow to pre-build), honesty (good agents cite supported answers), and failure behavior (stop clearly and offer an explicit human path instead of looping).
The good ones stop guessing. In the widget, HelpShelf withholds an unsupported answer and can offer an editable, visitor-confirmed handoff when your tested Telegram destination is reachable. Its answer policy requires support from published sources before a complete answer is shown.
Models vary widely: per-seat help desk plans with AI add-ons, per-resolution pricing, metered conversation allowances, and message-credit plans. HelpShelf has a free plan with no credit card, 500 AI searches, and 100 AI chats each month. Pro is $49/month for 10,000 searches and 2,000 chats.
Grounding and guardrails. The agent should only answer from retrieved passages of your real content, show citations so every claim is checkable, keep unreviewed AI-drafted content out of sensitive answers, and hold low-confidence replies for human review instead of sending them. If a vendor cannot explain their guardrails concretely, keep looking.
With HelpShelf, minutes: create a site, let it scan your website or docs (or run npx helpshelf against your codebase), and paste one script tag. There are no conversation flows to design. Most of the ongoing work is reviewing held answers and filling the content gaps the agent reports.
Yes — and it should, because visitors do not sort their questions into departments. HelpShelf detects buying intent, answers presale questions with citations, pins sales mode on pricing pages, and can capture a lead right after a confident answer. Owners see the buying questions in a Sales Pulse view.