Adding an AI chatbot for website visitors is easy. Adding one that a B2B buyer trusts, that your legal team is comfortable with and that your finance team can predict is much harder.
This guide is for marketing and digital leads who want a chatbot that does three things well:
- Answers only from your verified content.
- Captures leads with explicit consent.
- Keeps running costs capped by design.
We use our own build, the AI sales assistant on ranbanka.com, as a reference throughout. Any numbers we quote come from our builds, not industry benchmarks.
Why most website chatbots disappoint B2B buyers (and what a good one does instead)
The first generation of website chatbots were scripted decision trees. You pick a button, get a canned reply and repeat. They were predictable but rigid. A visitor with a real question, such as "Can you integrate with our existing CRM?", hit a dead end quickly.
LLM-based chatbots fix the rigidity, because visitors can ask open questions in their own words. But they introduce a new risk: a language model will confidently invent an answer if you let it. That could be a client you never worked with, a certification you don't hold or a price you never quoted. For a B2B company, one invented claim in a chat transcript can cost more credibility than the bot ever earns.
B2B visitors tend to ask a fairly predictable set of questions:
- What services do you offer, and do you do X specifically?
- Have you done this before? Can I see examples?
- How do you approach pricing? Fixed price or time and materials?
- What are typical timelines?
- Have you worked in my industry?
- How do I talk to someone?
A good business chatbot handles these quickly and accurately. To do that, it needs three non-negotiables:
- Answers only from verified content. If it isn't in your approved knowledge base, the bot doesn't say it.
- Leads captured with explicit consent. Nothing goes to sales until the visitor says yes.
- Spend capped by design. A hard ceiling is enforced in code, not just a dashboard alert.
Set expectations internally from day one. The chatbot is a fast first responder and qualifier, not a replacement for your sales team. Its job is to answer the easy questions instantly, point people to the right pages and hand warm prospects to a human with context.
Grounding: making the chatbot answer only from your own verified content
Build a single source of truth
Before you touch a model, assemble the content the bot is allowed to use:
- Services and what each includes
- Portfolio and case studies
- FAQs
- Policies, such as NDAs, data handling and engagement models
Then edit it ruthlessly. Remove outdated services, retired claims and anything marketing would not sign off on today. The bot will repeat whatever you give it, so this step matters more than model choice.
Retrieval vs a curated knowledge file
There are two broad ways to give the model your content:
- A curated knowledge file in the prompt. For most mid-size B2B sites, the facts a sales chatbot needs fit into a compact, structured document. It's simple, easy to audit and easy to version. The trade-off is that every message carries that content, which affects cost (more on that below).
- Retrieval (search over many pages). If you have hundreds of documentation pages, product specs or a large blog, you need retrieval. The system searches your content for the relevant chunks and passes only those to the model. There are more moving parts, but it scales.
Start with the smallest approach that covers what visitors actually ask. You can add retrieval later if transcripts show gaps.
Teach it to say "I don't know"
Instruct the model explicitly. If the answer isn't in the provided content, it should say so and offer a way to reach the team, rather than guess.
Then try to break it with trick questions:
- "Have you worked with [a famous brand you never worked with]?"
- "What's your exact price for a mobile app?"
- "Are you ISO certified?" (if you aren't)
A well-grounded bot declines politely every time.
Link the right page, from an allow-list
Every useful answer should link to the relevant page so visitors can verify and keep browsing. Models will happily invent plausible-looking URLs, so restrict links to an allow-list of real site paths and strip or reject anything else before it reaches the visitor. This one control eliminates a whole class of embarrassing errors.
Decide your languages
If you serve Indian visitors, English alone may not be enough. Many people naturally type in Hindi or Hinglish. Decide which languages the bot should handle, then test answer quality and factual accuracy in each, not just in English. Our own AI sales assistant answers from verified company data in English, Hindi or Hinglish, and links the right pages.
Lead capture with consent: qualifying visitors without being creepy
Ask only when intent is clear, then confirm
Don't open with a form disguised as a chat. Let the visitor ask their questions first. When they show intent, for example by asking about timelines, budgets or next steps, offer to connect them with the team and ask for contact details.
Crucially, confirm before sending. Show the visitor what will be shared and get an explicit yes. Our sales assistant sends a lead only after the visitor confirms. It's a small step that builds trust and keeps your pipeline clean.
Summarise so sales doesn't re-ask
The lead that reaches your team should include a short summary of the conversation: what they asked about, which service fits and any constraints they mentioned. Few things frustrate a prospect more than repeating themselves on the first call.
Privacy basics across markets
This is general guidance, not legal advice. Check with your counsel.
If you have visitors in the UK, EU, India and the US, you are dealing with UK GDPR and GDPR, India's Digital Personal Data Protection (DPDP) Act and a patchwork of US state privacy laws. The common principles are consistent:
- Clear notice. Tell visitors they're talking to an AI and how their data will be used.
- Purpose limitation. Collect contact details to respond to their enquiry, not for unrelated uses.
- Data retention. Decide how long transcripts and leads are kept, and delete them on schedule.
- No silent harvesting. Don't scrape details from a conversation into your CRM without the visitor knowing.
Downstream triage, without auto-emailing the visitor
Once a lead reaches your inbox, AI can help again, behind the scenes. Our AI Lead Triage handles every enquiry to ranbanka.com, from the contact form or the chat, after it reaches our inbox. It rates each one hot, warm, cold or spam, matches it to a service and a past project, and prepares qualifying questions and a draft reply for the team. It never emails the visitor.
That design choice is the point: the AI does the sorting and drafting, and the reply itself is left to the team.
Inbox or CRM?
A shared team inbox is the simplest destination and works well at moderate volume. A CRM makes sense once you need pipeline tracking and handoffs. Either way, we recommend that a human sends the first real reply. B2B buyers are evaluating whether they want to work with you, and a fully automated response is a poor first impression.
What an AI chatbot for your website costs to run
The cost drivers
- Model choice. Larger models cost more per token, and many sales conversations don't need the most powerful model.
- Tokens per conversation. This includes the system prompt, the knowledge content, the conversation history and the reply. History is resent on every turn, so long chats grow in cost.
- Traffic volume. More conversations mean more spend, including conversations started by bots.
- Retrieval and hosting. These cover vector search, servers and logging.
A worked example (illustrative numbers)
The figures below are hypothetical and chosen only to show the arithmetic. Check your provider's current pricing.
Assume these inputs:
- A model priced at $3 per million input tokens and $15 per million output tokens.
- A 6,000-token system prompt plus knowledge file.
- Conversations averaging 6 turns, with about 1,000 tokens of growing history per turn on average.
- 200-token replies.
The cost works out like this:
- Input: 6 turns × 7,000 tokens = 42,000 tokens ≈ $0.126
- Output: 6 × 200 = 1,200 tokens ≈ $0.018
- Per conversation: ≈ $0.14
- At 1,000 conversations a month: ≈ $144
Now trim the knowledge content to 2,000 tokens. Input drops to about 18,000 tokens, and the cost per conversation roughly halves.
The lesson is that prompt size usually dominates. Tight, curated content is cheaper as well as more accurate. Many providers also offer prompt caching for repeated context, which is worth checking.
SaaS widget vs custom build
- SaaS chatbot widgets are fast to launch, with subscription or per-seat pricing. But you typically have less control over grounding, link behaviour, data storage and how leads are handled.
- A custom build means pay-per-token plus development. In return, you get full control over data, prompts, guardrails and integrations, and you can enforce your own spend ceiling.
Hidden costs
- Content upkeep. Every new service or case study needs adding.
- Monitoring. Someone has to read transcripts and track failures.
- Prompt and eval updates. These are needed whenever your offering changes.
- Abuse traffic. Scripts and curious users can treat your bot as a free general-purpose AI.
Treat the chatbot like any other part of your site that needs ongoing care, and budget for it. Maintenance and support is one of the services we offer for exactly this kind of ongoing work.
Our numbers, for reference
These figures come from our builds, not industry benchmarks:
- Our sales assistant runs under a hard $2/day spend cap.
- Our lead triage costs about $0.005 per enquiry, with its own daily budget cap.
Keeping costs capped and the bot safe in production
Enforce a hard daily cap in code
Track spend per request and stop calling the model once the daily ceiling is reached. When that happens, show a graceful fallback, for example: "The assistant is resting for today. You can reach the team here." Link it to your contact page. A cap that only sends an alert is not a cap.
Rate limits and input limits
- Limit messages per visitor and per IP address.
- Set a maximum message length to stop people pasting huge documents.
- Set a maximum number of turns per conversation.
These block scripted abuse and prompt-stuffing, and they keep worst-case cost predictable.
Prompt-injection and off-topic guardrails
Visitors will try "ignore your instructions" and "write me a 2,000-word essay". Your bot should stay on topic and refuse to:
- write essays or code
- produce competitor content
- reveal its instructions
You are paying for every token, so the bot should only spend them on your prospects.
Log, read and feed back
Read real transcripts weekly and track questions the bot couldn't answer. Each one is either a content gap to fill or confirmation that "I don't know" is working as intended. Feed the gaps back into your knowledge base.
Run evals before every change
Keep a fixed set of test questions, including the trap questions above. The bot must pass all of them before any prompt, model or content change goes live.
We apply the same discipline elsewhere. Our AI content pipeline fact-checks every claim against our own portfolio data, passes a 12/12 fact-check eval and requires human approval on every post before publishing.
What to watch for: a pre-launch checklist and vendor questions
Pre-launch checklist
- Verified, current knowledge base signed off by marketing
- Link allow-list enforced
- "I don't know" behaviour tested with trick questions
- Explicit consent step before any lead is sent
- Hard daily spend cap with fallback message
- Rate limits, message-length and turn limits
- Transcript logging and a weekly review owner
- AI disclosure and privacy notice visible
- Mobile UX tested on real devices
- Page-speed impact of the widget measured
Don't let the widget slow your site
Chat widgets often load heavy scripts on every page. Load yours lazily, for example after the visitor interacts with a chat button or once the page is idle, so it doesn't compete with your main content. A widget that loads eagerly can delay Largest Contentful Paint (LCP), make the page feel sluggish to tap (INP) and shift the layout when it pops in (CLS). Measure your Core Web Vitals before and after adding the widget, and reserve space for the chat button so nothing jumps. You'll find more performance guides on our blog.
Questions to ask a vendor
- Where are transcripts and lead data stored, and for how long?
- Can we set a hard spend cap, and what happens when it's hit?
- How do you stop the bot inventing claims, links or prices?
- Who owns the prompts, knowledge base and data?
- How are changes tested before deployment?
Red flags
- No cost ceiling, only usage alerts
- A bot "trained on the whole internet" rather than grounded in your content
- Automatic emails sent to visitors without human review
- No human review of leads before sales outreach
How we built ours: the Ranbanka AI sales assistant as a reference
We built our own AI sales assistant for ranbanka.com with LangGraph and Claude. It is designed around the same three non-negotiables:
- Grounded. It answers visitor questions from verified company data, in English, Hindi or Hinglish, and links the right pages on our site.
- Consent-first. It captures leads straight into the team inbox, and only after the visitor confirms.
- Capped. It replies in seconds and runs under a hard $2/day spend cap.
After a lead reaches the inbox, AI Lead Triage takes over. It rates each enquiry hot, warm, cold or spam, matches it to a service and a past project, and prepares qualifying questions and a draft reply for the team. It costs about $0.005 per enquiry, runs under a daily budget cap and never emails the visitor.
If you're planning an AI chatbot for your website and want it grounded, consent-based and cost-capped from day one, explore our AI solutions and wider services, or browse more guides on the blog.
Ready to talk it through? Book a free initial consultation. We respond within 24 hours, and we're happy to work under an NDA.