AI Chatbot Suite

The build

Most support load is not hard: it is the same two hundred questions asked twenty thousand times — where's my order, what do you charge, when does admission close. The Chatbot Suite answers that reality with a proper product: a widget and channel bots wired to one retrieval layer over the business's own documents, tuned to refuse rather than bluff, and built to hand over mid-conversation the moment a human should take the thread. Its signature feature is the gap queue: everything the bot couldn't answer, counted and ranked — the FAQ your customers are writing for you. This is the customer-facing twin of our Automation Copilot, and it has been running our own inbox since launch.

What we handled

• Web widget in one embed script — themeable, mobile-first, Bangla/English/Banglish in the same thread
• WhatsApp Business and Messenger on the same brain — answers never drift between channels
• Knowledge ingestion from FAQ pages, PDFs, sheets and policies — re-sync on change, not on hope
• Citations on every answer; below-threshold confidence falls back to “I don’t know — want a human?”
• Human handover with full context: the agent reads the thread, not a transcript of a transcript
• Gap analytics: unanswered questions ranked by frequency — the support roadmap writes itself

Two hundred questions, twenty thousand times

Open any small business inbox and the shape is unmistakable: maybe two hundred distinct questions, arriving twenty thousand times a year. “Is the laptop still under warranty?” “What do you charge for a two-bedroom shoot?” “When does admission close?” Each one takes ninety seconds and a mood, and together they eat the hours that should go to the work itself. Off-the-shelf bot templates failed this problem a decade ago because they were decision trees pretending to be conversations — so most businesses gave up on bots, not on the idea.

Retrieval changed the middle of that sentence. A bot that reads your actual FAQ page, your fee sheet, your policy PDFs — and quotes them — is not the same product as the “Hi 👋 press 1 for sales” trees people remember. The Suite was built to that standard from day one: fast to the point of boring, cited to the point of auditable, and humble to the point of trustworthy. It answers the two hundred and gets out of the way of everything else.

Three channels, one brain

Customers do not consolidate their questions onto one platform, so neither do we. A web widget goes on the site via a single embed script — themeable, mobile-first, and happy to switch mid-thread from Bangla to English to Banglish. WhatsApp Business and Messenger connect the same brain to the numbers customers already text. Because retrieval, thresholds and tone rules live in the middle, an answer cannot drift between channels: fix the fee sheet once and every surface is correct on the next question.

The conversation design is deliberately unglamorous: short answers, a document chip under every claim (“Source: Fee Policy, updated 12 Bhadro”), quick-reply chips only where they speed things up, and a handover button always one tap away. A support bot earns its keep by resolving, not by entertaining.

Knowledge that stays fresh on its own

The ingestion layer watches the places answers actually live: FAQ pages, Google Docs and Sheets, published PDFs, product listings, the calendar. Documents sync on change, each answer cites the version it used, and anything stale past its review date loses the right to be quoted — the bot would rather say “I don't know” than cite last year's pricing. Owners see the whole base as a simple list with sync dots, not a vector-database dashboard nobody asked for.

Ingestion is also scoped honestly: we start with the twenty documents that generate eighty percent of questions and grow from there. A bot over a complete-enough knowledge base deflects most traffic; a bot over the entire messy archive takes longer to train and — this is the part competitors will not tell you — answers worse while it does.

The gap queue is the product

Every unanswered or low-confidence question is counted, clustered and ranked: “uniform vendor timings — asked 14 times this week”. That queue is the single most honest analytics view a support organization gets: it is your customers writing your FAQ for you, complete with frequency data. Each row has a one-click “create answer” that drafts from wherever a partial answer exists, and an owner approves it in the same place.

We instrument deflection the way you would want a mechanic to show their work — per question type, per channel, per week, with sampled transcripts a human can re-score. If the bot's number drops, the gap queue says why, and no one has to take the vendor's word for anything.

When it quits, it quits well

The moment a thread turns emotional, legal, medical or just stubborn, the bot's job ends and it says so — in the same breath offering a human. Handover carries the full thread, what was already checked, and what the customer is actually asking for, delivered to the inbox your team already works in: WhatsApp for the restaurant floor, Helpdesk for the school office. If no human is on shift, the bot books a contact window and logs a promise it will not break.

Confidence below threshold triggers the same graceful exit. A wrong answer costs trust three times what no answer does — customers forgive “I don't know, Mr. Faruk will call you at six”; they never forgive a confident lie about a refund policy. That asymmetry is a hard rule in this product, not a setting.