An AI chatbot for lead generation works on a B2B website when it does a few narrow jobs: it answers pre-sales questions from your own content, asks a handful of qualifying questions, and hands good-fit buyers to a calendar or a person quickly. A bot that tries to answer everything ends up qualifying nobody. Judge it by meetings held and pipeline, not by how many conversations it starts.
What an AI chatbot should and should not do on your site
Most B2B chatbots fail the same way: launched as a general assistant, they greet every visitor on every page and produce conversations that never become meetings. Write the bot a job description instead, as you would for a new SDR.
| Job | Bot does it? | Why |
|---|---|---|
| Answer pre-sales questions: plans, integrations, security basics, who it’s for | Yes | These questions stall buyers between browsing and booking |
| Qualify fit with a few questions | Yes | Sales gets context before the first call |
| Book a meeting or connect to a rep | Yes | This is the output that matters |
| Route support requests, job seekers and vendors | Yes, briefly | Keeps them out of the sales queue |
| Negotiate price or offer discounts | No | Creates commitments nobody approved |
| Promise roadmap items or custom features | No | Sets up a broken promise in the first call |
| Answer contract terms or security questionnaires | No | Needs a person and a paper trail |
Placement matters as much as scope. Put the bot where intent is highest: pricing, demo and contact, integration and comparison pages, and docs. On blog posts, keep it a quiet launcher with no auto-open message. A greeting that pops up everywhere trains visitors to close it unread.
Grounding answers in your docs, pricing and case studies
A bot running on a general model will answer questions about your product from whatever it absorbed from the web, which may be outdated, wrong or about a competitor. Grounding means the bot retrieves passages from a curated set of your content and answers only from those passages.
What goes into the knowledge base
- The current pricing page, plan limits and what each plan includes
- Product documentation and the full integration list
- Your public security and compliance page
- Case studies, each tagged with the customer’s industry, size and use case
- An approved-answers document: the questions sales hears every week, each answered in one to three sentences, written by sales and product together
- Your positioning against alternatives, stated factually
Tagging pays off: when a visitor says they run a 40-person agency on a legacy tool, the bot can point to the closest case study. A form can’t.
What stays out
Old pricing, draft roadmaps, sales decks with discount guidance, Slack exports and anything containing customer data. If you wouldn’t publish it on the website, it doesn’t belong in the bot.
Operating rules
- Answer only from retrieved content. If retrieval finds nothing relevant, the bot says so and offers a person.
- Link the source. Every factual answer links to the page it came from, which also moves visitors deeper into the site.
- Give the knowledge base an owner. Re-sync whenever pricing, packaging or docs change. Stale pricing in a chatbot is worse than no chatbot.
Qualification questions and scoring
Give before you ask. Let the bot answer at least one question, then qualify in three to five short questions, woven into the conversation rather than fired off as a form in disguise.
Don’t ask for anything enrichment can tell you. Company size, industry, location and often tech stack can come from the work email domain. The build for that sits in AI lead enrichment and routing. Spend the questions on what only the visitor knows:
- What are you trying to solve? Open text; the bot classifies it into your core use cases.
- What do you use today? A current tool or a spreadsheet tells sales a lot about the deal.
- A scale question tied to your pricing metric. Seats, locations, order volume, whatever you charge on.
- When do you need something in place?
- Your role in the decision. Only if enrichment can’t infer it.
Then combine fit and intent into a score. The points below are a hypothetical example to adapt, not a benchmark:
| Signal | Source | Example points |
|---|---|---|
| Company size inside your ICP range | Enrichment | +30 |
| Use case matches a core use case | Bot classification | +25 |
| Needs a solution within 90 days | Question | +20 |
| Decision-maker or evaluator role | Question or enrichment | +15 |
| Viewed the pricing page this session | Page context | +10 |
| Personal email domain | -20 | |
| Student, job seeker, vendor or support request | Bot classification | Route out, no score |
In this example, 70 or more is tier A, 40 to 69 is tier B, and anything lower is tier C. Keep the model simple enough that a sales rep can read it and agree with it. If sales doesn’t trust the tiers, they’ll ignore them.
Handoff: calendar booking, live chat and CRM
The handoff is where the value is. A qualified visitor who waits for an email reply cools off, and one who repeats everything on the first call feels the bot wasted their time.
- Tier A: show a calendar inside the chat, routed to the right rep by segment, territory or round robin. Scheduling tools with routing rules, including the meeting features built into some CRMs, handle this.
- Tier B: if a rep is online, offer a live handoff with the transcript attached. If not, offer the calendar for an SDR or a first-call slot.
- Tier C: send the most relevant guide or case study, offer to email a summary, and enroll them in nurture.
- Not a fit: link to support, careers or partnerships and end politely.
Every conversation should write back to the CRM: contact and company created or updated, qualification answers as properties, score and tier, lead source set to chat, the page where the chat started, and the session’s UTM values. For booked meetings, attach a short AI-written brief: what the visitor asked, their answers, and the case studies the bot sent. Reps read a five-line brief. They don’t read a 30-message transcript.
Guardrails: hallucinations, pricing promises and tone
Treat anything the bot says as something your company said in writing. Before launch, check:
- Instructions forbid discounts, custom quotes, roadmap commitments and legal or compliance guarantees
- The bot says it doesn’t know and offers a person when retrieval comes back empty
- Pricing answers quote the pricing page and link to it
- Competitor questions get factual, approved positioning with no disparaging claims
- The bot says up front that it’s an AI assistant, and keeps answers short and in the site’s tone
- After several turns without progress, the bot offers a person or a calendar
- It never asks for sensitive data, and retention matches your privacy policy
- A test set of tricky prompts (discount requests, “can you guarantee,” attempts to override instructions, off-topic questions) runs before launch and after every content change
Then read transcripts: weekly for the first month, a monthly sample after that. Review catches wrong answers early and shows how buyers actually describe their problem.
Measuring impact on pipeline, not conversation counts
Conversation volume rises whenever you make the bot more intrusive, so it tells you almost nothing on its own. Track the chain from conversation to revenue:
| Metric | What it tells you | Watch out for |
|---|---|---|
| Conversation rate by page | Where the bot gets used | Meaningless on its own |
| Qualified conversation rate | Share reaching tier A or B | Loose scoring inflates it |
| Meetings booked from chat | Direct output | Compare with the form on the same pages |
| Meeting held rate | Whether chat bookings are real | Easy bookings can mean more no-shows |
| Chat-sourced pipeline | Revenue impact | Double counting with form fills |
| Unanswered question rate | Content gaps | Rising rate means the knowledge base is stale |
The real question is whether chat adds meetings or just moves form fills into a chat window. If traffic allows, show the bot to half of visitors on high-intent pages and compare total demo requests per visitor, form and chat combined. On lower traffic, compare a longer before-and-after window and treat the result as directional.
The unanswered-questions log is a bonus: it lists what your website fails to explain, and the fix often belongs on the page itself. How to get more demo requests covers those page-level changes.
Custom build vs off-the-shelf options
| Off-the-shelf | Custom build | |
|---|---|---|
| Examples | Intercom Fin, HubSpot’s chatbot tools, Qualified | LLM API, your own retrieval, n8n or code, embedded widget |
| Time to launch | Typically weeks | Typically longer |
| Control over retrieval, prompts and scoring | Limited to vendor settings | Full |
| CRM integration | Strong when it’s native to your CRM | Whatever you build and maintain |
| Cost shape | Per seat, per conversation or per resolution, by vendor | Build cost plus model usage and upkeep |
| Best for | Standard funnels, no engineering time | Multiple products, complex qualification, unusual stacks |
My rule of thumb: start with the tool your CRM or support platform already offers if it can do four jobs well: answer, qualify, book and write back to the CRM. Build custom when your qualification logic doesn’t fit the vendor’s settings, or when you want the same workflow to enrich, score, route and brief the rep in one pass. That combined workflow is a common build in my AI automation work.
Get it built
If your site gets traffic but too few qualified meetings, I can scope and build a chatbot that answers, qualifies and books, wired into your CRM. Start with a Growth Audit, $1,500 fixed and credited if we continue. See pricing or get in touch.