B2B Chatbot Buying Guide: 12 Questions to Ask

How to use this list

Do not send these as a questionnaire. Questionnaires are answered in marketing language and every box gets ticked. Use them as the agenda for two sessions — one commercial, one technical — and insist that answers are either demonstrated live or committed to in writing.

Three habits separate a real evaluation from a procurement ritual. Bring your own difficult question rather than accepting the demo's. Ask what happens when each capability fails, because the failure behaviour is the product and the success behaviour is the brochure. And separate the platform from the deployment: a capable platform badly configured fails exactly like a weak one.

Knowledge and accuracy

1. 11. Where do answers come from?Ask the vendor to point at the specific piece of your material that produced a given reply. A system that records its sources lets you correct a wrong answer once. A system that cannot is a black box you will argue with for the length of the contract.
2. 22. What happens when it does not know?The correct behaviour is to say so and offer a person. Watch it happen with a question your material genuinely does not answer. Do not accept a description of this behaviour — ask to see it.
3. 33. How does material get in, and how is it updated?Website crawling, document upload, manual entries, a product catalogue, bulk import. Ask about limits by number, not in the abstract: how many pages, how large a file, how many records per request. Then ask who updates it in month three and how long that takes.
4. 44. Can it reach live data?Stock, order status, appointment availability and balances change faster than any indexed document. Ask whether your own API can be called mid-conversation, and what the bot says when that call times out.

Channels and conversation

1. 15. Which channels, and is the behaviour the same on each?This is where vague answers cluster. Capabilities genuinely differ per platform — voice replies, image understanding, comment handling and auto-reply rules are not available everywhere. Ask for the matrix, per channel, in writing, and be suspicious of a table with ticks in every cell.
2. 26. Is it one bot or several?If each channel has its own configuration and its own knowledge, you have bought several products to maintain. Ask whether a change to an answer propagates everywhere, and watch it happen.
3. 37. How much control do you have over what it says?Tone, length, when to end a conversation, which fields to extract, which languages. If changing the wording of one answer requires a support ticket, the system will be out of date within a quarter.

Handover and human control

1. 18. How does a conversation reach a person?Ask to see it: what triggers it, what the operator sees, and whether the customer has to repeat themselves. If the transcript does not travel, the automation has added a step rather than removed one.
2. 29. Can a human take over mid-conversation, and does the bot then stay quiet?Two parties answering at once is the most confusing outcome for a customer. Ask how the bot is paused, for how long, and how it is resumed.

Integration, analytics and data

1. 110. What is written after each conversation, and where?A transcript, a summary, extracted fields, a lead. Into the vendor's own CRM, into yours, or into an export you process later — the third is where most 'integrated' products actually sit. Ask specifically how a returning customer is matched, and what happens when they use a new number.
2. 211. What can you measure, and can you export it?Volumes, channel split, out-of-hours arrivals, conversations closed without a person, leads produced. Ask whether test conversations are excluded by default, and whether AI usage cost is visible per conversation — an AI line item you cannot break down becomes a monthly surprise.
3. 312. What is the exit path, and what does the pricing do at scale?Your transcripts, records and customer data should be exportable in a usable format without a negotiation. On price, do not ask for a rate card — take your busiest month, add a plausible growth figure, and ask for it to be priced. Vendors answer this readily when the model is simple and evasively when it is not, and the evasion is itself the signal.

Security and governance, asked properly

This section is where buyers most often accept reassurance instead of answers. Four things should be requested in writing during procurement, from any vendor including this one.

• Where data is stored and processed, and under whose jurisdiction.
• How long transcripts and recordings are retained, and whether you can set that period.
• Which subprocessors are involved, including which model providers see your conversations.
• Whether vendor-side access to your data is logged, and whether you can see that log.

A vendor who discourages putting these in writing has told you something useful. A vendor who has not yet been asked and says so honestly is in a different category from one who answers confidently and vaguely.

How to compare finalists

1. 1Pick three disqualifiers firstFor most buyers: it cannot answer from our material accurately, it cannot write to our CRM, or it cannot hand over cleanly. Anyone failing one of those is out regardless of how well they do elsewhere.
2. 2Run the same conversation with eachSame material, same awkward question, same channel. Differences invisible in a feature matrix appear within two exchanges.
3. 3Weight deployment effort, not just capabilityAsk each finalist how long until the first real conversation and what you have to supply. The gap here between vendors is usually larger than the gap in capability.
4. 4Compare price lastComparing price across products that do materially different things is how organisations buy the cheapest thing that does not work.

How Vexvon answers this checklist

On knowledge, material comes from five sources: your site crawled to as many as 4,500 pages, PDFs up to 25 MB, manual or bulk entries at up to 200 per request, a product catalogue with variants and images, and an Instagram archive. Each piece is targeted at the chatbot, the call agent, or both, and is classified into one of thirteen types. Retrieval runs over Milvus with six results by default, each company's material isolated by partition key, and the system records which pieces of knowledge produced a given answer.

On channels, every channel calls the same AI engine, so an answer corrected once is corrected everywhere. Capabilities do differ by platform and the differences are stated rather than glossed: Instagram is the deepest channel, with voice replies, reel and image analysis; several other channels support incoming voice messages transcribed to text but not spoken replies; auto-reply rules exist on some channels and not others. Language is detected per message, and the company can restrict the list.

On handover, three explicit routes exist — a customer request raising a notification, an agent's stop character pausing the AI for thirty minutes, and switching the AI off for a conversation — with conversation memory of the last fifteen messages travelling with it. On data, each conversation writes a transcript, extracted fields and a one-sentence summary to a customer record shared across channels, with phone matching in five stages, a buyer-intent filter before a lead is created, and onward delivery to Bitrix24 or the built-in nine-stage CRM.

On analytics, reporting covers volumes, channel split, out-of-hours arrival, hourly heat maps and leads produced, exportable to a file, with AI cost and tokens logged across seventeen distinct purposes. On governance, prompt versions are tracked with one active at a time and a path back. The security items in the section above — storage location, retention, subprocessors, access logging — should be requested in writing during procurement, as they should be from any vendor.

Frequently asked questions

1. 1What should a B2B chatbot cost?There is no useful benchmark, because products meter differently — conversations, seats, channels, AI usage or a mix. The only meaningful comparison is a priced worked example using your own expected volume, which is fair to ask of every vendor.
2. 2What is the most commonly skipped question?What happens when the bot does not know. It is the behaviour that defines the product in daily use, and it is never in the demo.
3. 3How long does deployment usually take?The connection is quick. The work is the material the bot answers from, the conversation design and the routing rules, so an honest estimate depends more on the state of your documentation than on the software.
4. 4Should we run a pilot?Yes, on one channel and one clearly defined set of questions, with success criteria agreed in advance. A pilot without written criteria becomes an opinion contest.
5. 5What data and integrations are required?Material to answer from, a CRM or notification destination, customer matching on phone or email, and a handover path into a queue your team actually works in.
6. 6When should a human take over?On an explicit request, on commercial negotiation, on complaints, and whenever the material does not cover the question — and the transcript must travel with the conversation.

Take this into a technical session

The version of this conversation that produces a decision is not a sales demo. It is a working session with your material on the table, your CRM in the room, and the two or three questions you most dread being asked by a customer. An hour spent that way settles more than a month of feature comparison.