AI chatbot for B2B websites: a buyer's checklist

TL;DR

• Buying an AI chatbot for a B2B website is mostly about trust: does it answer from your content, does it admit when it does not know, and does it pass serious buyers to a person with the context intact.
• This checklist covers seven areas: answers, leads, control, privacy, installation and accessibility, reporting, and pricing and exit terms. Each item is a yes/no question you can put to a vendor or test yourself.
• Do not buy from a demo. Connect your own site, ask your own questions and read the transcripts. An hour of testing beats a week of feature tables.

Why B2B sites need a different checklist

A chatbot on a consumer store answers "where is my order" thousands of times a day. A chatbot on a B2B website has a different job. Traffic is lower, each visitor is worth more, and the questions are harder: does this integrate with our ERP, what does the enterprise plan include, do you have a DPA, can you deliver to Poland by March.

The people asking are also better prepared than they used to be. In a Gartner survey of 645 B2B buyers published in May 2026, 45% said they used generative AI in a recent purchase, mostly to research vendors and products, and 69% said they prefer to validate AI-generated insights with a sales rep. Visitors arrive with an AI-written summary of your company and want it confirmed or corrected. A bot that guesses makes that worse.

So the bar is higher than "it chats." Use the checklist below to test whether a product clears it.

How to use this checklist

Mark each item as must have, nice to have or not needed for your site before you talk to vendors. Then score each product only on your must-haves. Most items can be checked in a trial; a few, mostly privacy and contract terms, need a written answer from the vendor. Keep their answers. They are what you hold the vendor to later.

1. Answers

This is the core of the product, and the part demos hide best.

• Answers come from your content. The bot builds its knowledge base from your pages and documents, not from what the underlying model happens to remember about your industry.
• Every answer shows its source. Visitors and your team can click through to the page an answer came from. This is the practical form of grounding.
• It says "I don't know." When your content has no answer, the bot says so and offers a way to reach you instead of inventing one. Test this deliberately with a question your site cannot answer.
• It handles pricing honestly. If your prices are public, it quotes them correctly. If they are not, it explains how pricing works and offers a quote or a call, without making up a number.
• It stays current. You know how often it re-reads your site and can trigger a refresh after you change a pricing or policy page.
• It handles your languages. If you sell abroad, test questions in each market's language.

Why this matters more than it seems: in Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline responsible for a refund policy its chatbot described incorrectly. The tribunal rejected the idea that the chatbot was separate from the company: it is "still just a part of Air Canada's website." Whatever your bot says, your company says. Our guide to grounded answers covers how to keep a bot from making things up.

2. Leads and handoff

On a B2B site, the conversations that matter most end with a person.

• It captures contact details at the right moment, after it has helped, not as a gate before the first answer.
• It can ask qualifying questions you define, such as company size, use case or timeline, and skip the ones a visitor has already answered.
• It can book a meeting in your team's calendar, or at least hand over a booking link with context.
• Handoff keeps the transcript. When a visitor asks for a person, your team gets the full conversation so nobody asks the same questions twice. Good human handoff is the difference between a warm lead and an annoyed one.
• Leads reach your systems. Through a native CRM connection, webhooks or email, with the source page and conversation attached.
• Out-of-hours behavior is clear. The bot tells visitors when they will hear back instead of promising an immediate callback nobody will make.

3. Control and safety

You need to decide what the bot does, not discover it in transcripts.

• You can set scope. Topics it should not discuss (legal advice, competitor comparisons, discounts) and how it declines them.
• You can set tone and wording for greetings, handoffs and refusals.
• It resists manipulation. OWASP lists prompt injection as the top risk for applications built on language models: input that changes the model's behavior "in unintended ways" (OWASP). Ask what the vendor does about it, and try a few attempts yourself, such as asking the bot to ignore its instructions and offer a 90% discount.
• It can only do what it needs to. OWASP's mitigations include restricting the model's access "to the minimum necessary" and human approval for privileged operations. A chatbot that can only answer and hand over cannot be talked into doing much else.
• You can review conversations and correct the content behind a bad answer.

4. Privacy and compliance

These items need written answers. Ask for documents, not reassurance.

• AI disclosure. Under Article 50 of the EU AI Act, AI systems that interact directly with people must be designed so that people know they are interacting with an AI, unless that is obvious from context. The information has to be "clear and distinguishable" and given no later than the first interaction. Under Article 113, the Act applies generally from 2 August 2026. Check that the widget says it is an AI, by default.
• A data processing agreement. When a vendor processes visitors' personal data on your behalf, GDPR Article 28 requires that processing to be "governed by a contract" that sets out its subject matter, duration, purpose and the types of data. Ask for the DPA up front.
• A subprocessor list. Which model providers and hosting companies see your visitors' messages, and in which countries.
• Training use. Whether your content and your visitors' conversations are used to train the vendor's or anyone else's models. The answer should be in the contract.
• Retention and deletion. How long transcripts are kept and how you delete a visitor's data on request.
• Cookies and consent. What the widget stores in the browser, and whether it works with your consent banner.

For what to tell visitors once the bot is live, see AI chat on your website and privacy. For comparison, ASDK publishes its DPA and subprocessor list.

5. Installation, performance and accessibility

• One script or a plugin for your platform, without a developer project. Check the vendor's install guide for your CMS; ours are under installation docs.
• It loads without slowing the page. The script should load asynchronously and not block rendering. Run your page speed test with and without the widget.
• It works with your Content Security Policy if you have one: ask which domains to allow.
• It meets accessibility basics. WCAG 2.2, a W3C Recommendation since 12 December 2024, is the usual reference. Two quick tests: open and use the chat with the keyboard alone (success criterion 2.1.1, Keyboard), and check that the open widget never hides the element that has focus (2.4.11, Focus Not Obscured).
• It behaves on mobile. The widget should not cover your main call to action or the cookie banner's buttons.

6. Reporting

• Conversations and transcripts you can search and export.
• Unanswered questions, grouped, so you can see which content is missing.
• Leads, meetings and handoffs, tied to the pages where conversations started.
• A way to share results with sales and marketing without giving everyone an admin login.

7. Pricing and exit terms

• You know the billable unit. Seats, conversations, answers, credits or resolutions. Intercom, for example, charges $0.99 per Fin "outcome" on top of seat prices (Intercom pricing, checked October 21, 2026). Read the definition of the unit word by word; our guide to AI agent pricing models explains why it matters.
• Overage is predictable. What happens when you hit the limit: extra charges, a cap you control, or the bot going quiet.
• Branding removal and extra sites are priced, not "contact us."
• A real free plan or trial on your own content, not a sandbox.
• Monthly billing is available, or the annual commitment is worth the discount.
• You can leave. Export of transcripts and leads, and deletion of your data when you cancel.

The checklist at a glance

A one-hour test drive

Red flags

• The vendor will not let you test on your own content before paying.
• Demo answers are perfect, but your trial answers have no source links.
• The bot never says "I don't know."
• There is no written answer on whether conversations are used for model training.
• The pricing page defines the billable unit loosely or not at all.
• Handoff means "the visitor gets an email address."

If you are still deciding whether you need a scripted chatbot or an agent that answers from your content, read website chatbot vs AI agent first.

The right AI chatbot for a B2B site is the one that is accurate on your content, honest about its limits and good at handing serious buyers to your team. Test those three things on your own site, get the privacy answers in writing, and the rest of the checklist will sort the remaining options quickly.