Quick answer: A chatbot resolves questions when it is grounded in a clean knowledge base and hands off cleanly when it is unsure. Feed it well-structured docs, use RAG so it answers only from your content with citations, test against real questions, and set clear escalation triggers. With
BuiltABot, you train on your docs and watch your resolution rate climb.
Most chatbots do not actually resolve anything. Secret-shopper research across the industry has repeatedly found that chatbots resolve only a small fraction of the conversations they handle — one widely-cited study put it in the single digits. That is the dirty secret behind the buyer’s real fear: “Will this bot actually work, or just annoy my customers into giving up?”
The good news is that the gap between a bot that frustrates and one that resolves is not magic. It comes down to a handful of fundamentals: the quality of the knowledge base, grounding the answers, testing against real questions, and handing off cleanly when the bot cannot help. This guide walks through each one so you can build a knowledge-base chatbot that actually resolves — and measure it. (Resolution figures cited here are external industry data; benchmark against your own baseline.)
Build on these knowledge base guides
Why Most Chatbots Fail to Resolve (the Resolution Gap)
When independent testers pose real customer questions to live chatbots, the resolution numbers are sobering — a large share of conversations end without the customer getting a useful answer. Treat that low bar as the thing to beat, not a reason to avoid automation. The failures cluster into three causes:
- Thin or messy knowledge base: the bot cannot answer what the business never documented clearly.
- No grounding: ungrounded models invent plausible-sounding answers, destroying trust the first time a customer catches one.
- Dead-end design: when the bot is stuck, it loops or apologizes instead of escalating — so the conversation fails even when a human could have helped.
Every one of these is fixable, and the fixes compound. Nail them and your bot moves from the frustrating majority to the resolving minority.
Resolution Starts With the Knowledge Base
Garbage in, garbage out. No model, however advanced, resolves questions your documentation never answered. Before you tune anything else, get the knowledge base right.
What to feed it
- Help center and docs: your primary source of truth — crawl it.
- FAQ pages: already written in question form, ideal for retrieval.
- PDFs and manuals: product guides, policies, spec sheets.
- Your website: pricing, features and policy pages customers ask about.
Structure it for retrieval
- One topic per article so the bot retrieves a clean, focused answer.
- Front-load the answer — put the resolution in the first lines, detail after.
- Use conversational titles (“How do I reset my password?”) that match how people actually ask.
- Prune the noise — exclude marketing pages and boilerplate that dilute retrieval.
Full mechanics are in our knowledge base connection guide and training-on-your-data guide.
Build a Chatbot That Actually Resolves
Train BuiltABot on your knowledge base and watch your resolution rate climb. 14-day free trial, no credit card required.
How to Set Up a KB Chatbot That Resolves (Step by Step)
Step 1: Ingest your best sources
Connect your help center URL and upload key PDFs in your BuiltABot dashboard. Start with the content that answers your top 20 questions — coverage beats volume.
Step 2: Structure and prune
Review what was indexed. Remove irrelevant pages, split mega-articles, and confirm the content is in answer-first shape. This single step moves resolution more than any prompt tweak.
Step 3: Ground every answer (RAG)
Keep the bot constrained to your content. BuiltABot uses RAG so answers come from your documents with citations, and the bot admits uncertainty instead of inventing. This is the difference between “confident and wrong” and “resolved.”
Step 4: Test against real questions
Before launch, run the 15–20 questions customers actually ask. Confirm each is resolved correctly and cites the right source. Ask off-topic questions to verify the bot escalates rather than guesses.
Step 5: Measure, then close gaps
Launch, then watch the unanswered-questions report. Every unresolved query is a documentation gap — fill it, re-crawl, and resolution climbs week over week.
When the Bot Should Hand Off (and Why That Still Counts)
Resolution does not mean the bot answers everything. A clean handoff is a win: the customer gets helped, just by a human. Design escalation deliberately:
- No confident answer in the knowledge base → offer a human.
- Explicit request (“talk to a person”) → escalate immediately.
- Sensitive issues (billing, complaints, security) → route to a human with full context.
The anti-pattern is the dead-end loop. Resolve first, escalate cleanly, never trap the customer. See our human handoff guide for trigger design.
Measuring Resolution Rate
If you do not measure resolution, you cannot improve it. Track these:
| Metric | Why it matters |
|---|
| Resolution rate | Share of conversations handled without a human — the headline number |
| Escalation rate | How often the bot hands off — high rates flag KB gaps |
| Unanswered questions | A ranked backlog of docs to write next |
| Satisfaction (thumbs) | Which topics resolve well vs. frustrate |
Pair resolution with ticket deflection to see the downstream cost savings as resolution climbs.
Build a KB Chatbot That Resolves With BuiltABot
The resolution gap is an opportunity: most competitors are shipping bots that frustrate. A grounded, well-documented, cleanly-escalating assistant beats that bar handily.
- Start free: create a BuiltABot account (14-day trial, no card).
- Connect your knowledge base: crawl your help center, upload key docs.
- Ground and test: RAG answers with citations, validated against real questions.
- Set handoff triggers: resolve first, escalate cleanly.
- Measure and improve: close gaps from the unanswered-questions report.
Explore the knowledge base integration solution or view pricing from $29.99/month.
Why do chatbots give wrong answers?
Most chatbots give wrong answers for one of two reasons: either they are answering from a weak or disorganized knowledge base, or they are generating responses without grounding them in your actual content. A general-purpose model that is not constrained to your documents will confidently invent answers. The fix is RAG (Retrieval-Augmented Generation), which forces the bot to retrieve and answer only from your real knowledge base — and to say it does not know when the answer is not there.
What is a good chatbot resolution rate?
Resolution rate is the share of conversations the bot fully handles without a human. Widely-cited industry secret-shopper research has found many chatbots resolve only a small fraction of conversations — a low single-digit to low double-digit percentage — which sets a low bar that a well-built, docs-grounded bot can beat substantially. Rather than chasing a universal number, benchmark against your own baseline and track resolution climbing as you improve the knowledge base, test coverage and handoff rules.
How do I train a chatbot on my knowledge base?
With BuiltABot you connect your help center, documentation, FAQ pages, PDFs and other sources, and the AI indexes them into a searchable knowledge base. You can crawl a URL or upload files, combine multiple sources into one bot, and set a re-crawl schedule so content stays fresh. The step-by-step process is covered in our guide on how to connect a knowledge base to a chatbot and how to train a chatbot on your data.
How do I stop my chatbot from making things up?
Use a chatbot built on RAG that answers only from retrieved content and cites its sources. BuiltABot grounds every answer in your indexed documents; when the information is not present, the bot acknowledges it rather than guessing, and can escalate to a human. Keeping your knowledge base clean, well-structured and current is the other half — a hallucination is often just the bot filling a gap your documentation left open.
When should a chatbot escalate to a human?
A chatbot should escalate when it cannot find a confident answer in the knowledge base, when the user explicitly asks for a person, or when the conversation involves a sensitive or high-stakes issue (billing disputes, complaints, account security). A clean handoff — passing full conversation context to an agent — still counts as a successful resolution for the customer. The goal is resolve-first, escalate-cleanly, never dead-end.
How is this different from a generic AI chatbot?
A generic AI chatbot answers from a broad model and is prone to confident-but-wrong replies. A knowledge-base chatbot that resolves is constrained to your actual content via RAG, cites sources, and hands off cleanly when it is unsure. The difference shows up directly in resolution rate and customer trust: grounded bots resolve more and frustrate less.
