A WhatsApp chatbot that knows your business

The useful question is not whether a bot can answer, but what happens on the answers it gets wrong. That is where most WhatsApp chatbots fail.

A WhatsApp chatbot is an automated conversational agent that answers customer messages on WhatsApp — either by following rules you build on a canvas, or by retrieving answers from your own approved content — and hands the conversation to a human when it cannot answer confidently.

ConvoFly gives you two kinds of bot, and most stores end up running both. Understanding the difference is the whole decision.

Rule-based flows: predictable, and you own every word

A flow is a canvas of blocks you wire together. It matches keywords or store events, branches on conditions, asks questions, saves the answers and takes actions. Every reply is text you wrote, so it cannot say anything you did not approve.

This is the right tool for anything transactional — confirming a cash-on-delivery order, collecting a delivery address, qualifying a lead, taking a booking. You want those exact, not creative.

The knowledge-base bot: answers you didn't script

The knowledge-base bot answers open questions from content you have approved: pages crawled from your own website, uploaded documents, and Q&A pairs you write or mine from past chats. It retrieves the relevant material and answers from it, rather than inventing an answer.

Two controls make it safe to run. A confidence floor means it only answers when the retrieved match is good enough — below that it escalates rather than guessing. And handoff passes the conversation to a human in the same shared inbox, so the agent sees everything the bot already said.

Where the knowledge base comes from

The bot is only as good as what you give it, and there are three sources. A crawl of your own website picks up your shipping policy, returns policy and product pages. Uploaded documents cover the things that are not on the site — a price list, terms, a supplier agreement. Q&A pairs are the highest-quality source, because you write both halves, and they can be mined from your own past conversations rather than written from scratch.

Start with the Q&A pairs for your top ten questions and add the crawl afterwards. A crawl alone produces a bot that technically has the information and phrases it like a web page. AI answers from your own website covers what the crawler picks up and how to keep it current.

If you serve customers in more than one language, note that the bot answers in the language of its sources unless you branch — multilingual WhatsApp automation covers the patterns that work.

AI Copilot: the option people underrate

Copilot drafts a reply grounded in your knowledge base and shows it to the agent, who edits and sends. Nothing goes out without a human. For teams nervous about a bot talking to customers unsupervised, this captures most of the speed with none of the risk, and it is often the right first step.

Building one

  1. Start with your five most repeated questions. Read last month's chats; do not guess.
  2. Build those as a rule-based flow with keyword triggers. Exact answers, no AI needed.
  3. Add your knowledge base — point the crawler at your site, upload your policy documents, and approve a starting set of Q&A pairs.
  4. Set the confidence floor deliberately. Start strict. A bot that escalates too often is an annoyance; one that answers wrongly is a refund.
  5. Wire the handoff to the right team, and make sure a human actually watches that queue.
  6. Test in the simulator before publishing, then review what it answered in the first week and mine the misses back into your Q&A pairs.

Set the confidence floor deliberately

The floor is the single most consequential number in the whole feature, and the default of 0.40 is a starting point rather than an answer. Raising it means the bot answers less often and is right more often; lowering it means the reverse.

The right setting depends on what a wrong answer costs you. A bot answering "what are your opening hours" can run low — being wrong costs a follow-up message. A bot answering "can I return this after 40 days" should run high, because being wrong costs a refund and an argument. If you only ever change one setting after launch, change this one, and change it based on the week of answers you have read rather than on a feeling. Testing a WhatsApp automation safely covers exercising the bot before customers do.

The mistakes that make customers hate chatbots

No visible way out. Every flow needs an obvious route to a human, and a human who is actually watching. A bot that traps people is worse than no bot.

Pretending to be a person. Customers work it out, and then they distrust everything that came before. Be plainly a bot and be useful.

A confidence floor set too low. A bot that answers everything will answer some things wrongly, and a wrong answer about a refund policy costs more than an escalation.

Never reviewing what it said. The knowledge-base bot logs what it answered and what it escalated. Reading a week of that and feeding the gaps back into your Q&A pairs is the difference between a bot that improves and one that quietly annoys people for a year.

Measuring a chatbot honestly

The vanity metric is "questions answered". The number that matters is resolution without escalation — conversations the bot closed where the customer did not come back with the same question a day later. Alongside it, watch your CSAT scores on bot-handled chats against agent-handled ones. Auto-CSAT surveys are available from the Growth plan, and Diamond adds an AI customer-experience score on every resolved conversation, which is the fastest way to spot a bot doing damage at volume.

What a WhatsApp chatbot cannot do

Meta's rules bind bots exactly as they bind humans. A bot cannot message someone first without an approved template, cannot re-open a conversation more than 24 hours old except with a template, and cannot exceed your number's messaging tier — see Meta's messaging limits. Anyone selling you a WhatsApp bot that "reaches your whole contact list any time" is describing a policy violation.

Two more limits are worth stating because they surprise people. A bot cannot exceed the 1,000 / 10,000 / 100,000 unique-customer tiers that govern your number, and a bot sending badly-targeted messages damages the quality rating that decides which tier you sit in. Nor can a bot send a template that has not been approved — Meta's template documentation covers what gets rejected, and why templates get rejected covers the patterns that actually trip review.

The knowledge-base bot also only knows what you have given it. It does not browse the internet at answer time, and it will escalate rather than improvise on a question your content does not cover — which is the behaviour you want.

What it costs

Rule-based flows are on every plan from $12/month. The knowledge-base bot, AI Copilot and AI blocks unlock on Growth at $50/month, which includes a monthly pool of AI credits — one credit per bot answer, Copilot draft or AI block. Every plan includes a 14-day free trial.

Rule-based flow vs knowledge-base bot

Rule-based flowKnowledge-base bot
Answers questions you scripted Yes Yes
Answers questions you didn't script No Yes, from your content
Every word pre-approved by you Yes Generated from your sources
Best for transactions (COD, address, booking) Yes Not suited
Best for open FAQs Rigid Yes
Escalates when unsureBy designBelow a confidence floor
Available fromStarter, $12/moGrowth, $50/mo

Figures checked August 2026. WhatsApp platform limits and charges are set by Meta and can change — always confirm against Meta's own documentation.

Common questions

Will the chatbot make things up?

The knowledge-base bot answers from content you have approved — your crawled website, uploaded documents and Q&A pairs — and escalates to a human when the best match falls below a confidence floor you set, rather than guessing. Rule-based flows send only text you wrote.

Can the bot start a conversation with a customer?

Only with an approved WhatsApp template, and only to contacts who have opted in. This is Meta's rule for every provider. Inside the 24-hour window after a customer messages you, the bot can reply freely.

How does handoff to a human work?

The bot assigns the conversation to a team or agent in the same shared inbox, and the agent sees the full transcript including everything the bot said. There is no separate bot console to check.

Do I need AI to run a WhatsApp chatbot?

No. Rule-based keyword flows are on the $12/month Starter plan and handle transactional jobs better than AI does. AI helps with open-ended questions and messy phrasing, and unlocks on Growth at $50/month.

What languages does it handle?

The AI blocks include translation and intent detection that work across languages, and the knowledge-base bot answers in the language of your source content. Rule-based flows reply with exactly the text you write, so multilingual flows use branching on the customer's language attribute.

Your customers are already on WhatsApp.
Now your business is too.

Set up your inbox, connect your store and launch your first automation — today.

Start free trial