Grounded support, one script tag
Answer your customers from what you have already written.
Reads your docs, pages and recordings. Answers from those alone — or says so, and shows what to write next.
What it reads
- DOCX
- XLSX
- CSV
- TXT
- MP3 / MP4
- URL
- Sitemap
- Pasted text
Built for businesses whose customers ask the same questions every day.
Online stores
“Where is my order?” “What’s your returns window?”
Answered from your shipping and returns policies — at 2am, without a ticket.
Software teams
“How do I connect our CRM?” “Why won’t my invite send?”
Answered from your docs and changelogs, so your team writes instead of repeating.
Agencies
One bot per client, each with its own knowledge and colours.
Run them all from one login, with roles and per-workspace costs you can bill back.
Service businesses
“Do you service my area?” “What does a call-out cost?”
Answered from the price list and the FAQ that live in a PDF on someone’s laptop.
If your customers’ questions are answered somewhere in your documents, it can answer them tonight.
01 — Today
The answer exists. Nobody can reach it.
It is in the refund policy nobody has opened since March. In one paragraph on the pricing page. In the head of whoever wrote the onboarding deck. In the recording of last Tuesday’s training call. None of it is reachable at eleven at night — and increasingly none of it is reachable by your newest hire either.
And the obvious fix makes it worse.
Put a general-purpose assistant on the page and you get one that is fluent, confident, and wrong about your business. It will invent a refund window. It will describe a feature you cut. Now your team is handling the tickets your bot created.
How long do I have to return something?
You have 30 days from delivery for a full refund, no questions asked.
InventedThe real window is 14 days, and it is written on the returns page it was never given.
A support bot that is wrong one time in twenty is not ninety-five per cent useful. It is a liability with a chat bubble.
BotIQ exists because the fix for that is not a better model. It is a smaller question: what is this bot allowed to answer from?
How an answer is produced
02 — Ingest
Start from what you already have.
Hand over the documents, the help pages, the sitemap, the call recordings. Long files are cut into overlapping passages, so an answer arrives with the sentence before it and the sentence after — not a fragment that happened to match.
Audio and video are transcribed first, which is how the part nobody wrote down gets in.
03 — Index
Every passage gets a position.
Passages about the same subject end up near each other — not because anyone tagged them, but because nearness is what the position means. No keywords, no intents, no decision trees. You add material; the map redraws itself.
no tags · no intents · no dialogue trees
04 — Retrieve
A question is a position too.
Your customer’s question is placed on the same map, and only the passages nearest to it are ever shown to the model. It cannot answer from anything else, because it is never given anything else.
question → nearest passages → answer
05 — The miss
And sometimes there is nothing nearby.
Ask about something you never uploaded and the question lands in empty space. Nothing is close enough to retrieve — so no model is called at all, and there is nothing for it to be confidently wrong with.
A miss costs nothing, because nothing was spent on it.
06 — The moment worth paying for
What it does when it does not know.
Every other support bot treats a miss as an embarrassment to hide. Here it is the most useful thing the product produces.
What the visitor sees
“I don’t have information about that. Please contact support and someone will help.”
The words you wrote, returned exactly — not a generated apology in a voice you did not choose.
What you get
The question your bot could not answer this morning is at the top of what to write this week. Write it once, upload it, and it stops being asked — the count stops rising and the gap closes.
See what that list looks like07 — The loop closes
A content backlog written by your own customers.
Every question the bot could not answer is recorded and counted. Ranked by how often it was asked, so the top of the list is the most valuable paragraph you could possibly write this week.
Answer one in place — it becomes a source, goes through the same chunker and embedder as everything else, and the next person who asks gets a real answer. If it ever stops matching, you are told.
What it costssample data · not a real workspace
Conversations
Every transcript, summarised automatically once it goes quiet.
Confidence
Per answer: how strong the retrieval was, and which passages it used.
Cost
Per workspace and per model, built from recorded token usage.
Volume
Answers against your plan, over time, with the busiest days marked.
08 — Ship it
On your site in five minutes, in your colours.
One script tag, pasted once, by whoever looks after the site. It renders inside a shadow DOM, so your CSS and the widget’s cannot reach each other — nothing of yours breaks it, and it breaks nothing of yours.
<script src="https://botiq.ai/widget.js" data-workspace="ws_9f2c…" defer ></script>
Every element and every state — launcher, header, bubbles, input — is styled in a visual editor over a live preview of the real widget. Domains are allowlisted, so the snippet only runs where you said it could.
Northwind Coffee
Answers from their own documents
Hi! Ask me anything about Northwind Coffee — orders, shipping or your machine.
How long do I have to return a grinder?
30 days from delivery, unused and in its original packaging. Start the return from your order confirmation email and we cover the shipping label.
Answered from returns-policy.pdf
Try asking
Sample knowledge base: three documents from a fictional coffee company.
Ask it the third question. It is the one it turns down — in the same words a real workspace would have written.
Try it on your own document — free plan, about five minutes.
09 — The rest of the job
The work you would otherwise be doing by hand.
Answering questions is the interesting half. This is the half that decides whether you are still running it in six months.
You never find out what the bot could not answer.
A backlog you did not have to write
Unanswered questions are recorded and ranked by how often they were asked, and answering one in place turns it into a source.
Spend arrives as one invoice with no breakdown.
You can see where the money went
Cost per workspace and per model, built from recorded token usage priced against a dated table — so last quarter stays correct when a provider changes its rates.
Embedded widgets look bolted on.
It looks like your site, not like ours
Every element and state is editable over a live preview of the real widget, and compiled to CSS scoped to its own shadow root.
Everyone with the login can change everything.
Roles, enforced on the server
Invite by email, assign a role, remove someone. Checked server-side, and an organisation can never be left without an owner.
Your other systems never hear about a conversation.
It tells the rest of your stack
Signed webhooks for conversations, summaries and handoffs, delivered with backoff and a log you can read. A versioned HTTP API for sources and chat.
Getting your data out is a support ticket.
Leaving is as easy as arriving
Export the whole workspace to one file, set a retention window and let old transcripts age out on their own, or delete a conversation and have its messages go with it.
10 — The proof
One frame, three things you can hold it to.
This is your index as the product sees it: passages clustered by subject, a question inside one cluster, and only the nearby passages lit. Everything below follows from that picture.
Free plan, no card.
- Only retrieved passages reach the model. What the widget answers from is the set above your confidence threshold — never its own memory of the internet.
- No passage clears the bar, no model call happens. The visitor reads your fallback text verbatim, and you pay nothing for the question.
- Every miss is written down and ranked. The questions it could not answer become an ordered list of what to write next.
11 — Pricing
One price. Everything included.
No credits sold in packs, no per-conversation fee, no charge to remove branding. An unanswerable question costs nothing, because no model is called for it.
Free
Free
One bot, no card, no expiry.
- Bots
- 1
- Answers a month
- 500
- Recording minutes
- —
- History kept
- 90 days
Starter
$39 /mo
For one business with real traffic.
- Bots
- 3
- Answers a month
- 8,000
- Recording minutes
- 60
- History kept
- 365 days
Growth
$129 /mo
Several bots, or one busy one.
- Bots
- 10
- Answers a month
- 40,000
- Recording minutes
- 300
- History kept
- 365 days
Business
$399 /mo
Agencies and larger corpora.
- Bots
- Unlimited
- Answers a month
- 150,000
- Recording minutes
- 1,200
- History kept
- 730 days
An answer is one reply on the default model. A more expensive model spends more of the allowance per answer, and you choose the model per bot. Retention is adjustable per bot within your plan’s bounds.
12 — Bounded on purpose
Why this is different from putting a model on your page.
| Concern | BotIQ | A general assistant |
|---|---|---|
| Where answers come from | Only passages retrieved from your own material, with the words you wrote when nothing matches. | A general model answers from training data, confidently, whether or not it knows your business. |
| When it does not know | It says so, and the question joins a ranked list of what to write next. The gap closes and stays closed. | It guesses, you find out from the customer, and nothing is recorded. |
| Who has to be involved | Whoever writes the help content. One developer, once, for five minutes, to paste a script tag. | An ingestion pipeline, a vector store, an auth model and a chat interface — then the support work. |
| What it costs to run | One price per plan, with per-workspace spend visible and a miss costing nothing at all. | One provider invoice at the end of the month, and every unanswerable question paid for. |
Your content is yours
Sources and conversations live in your workspace and are never used to train a shared model. Export the lot to a single file whenever you want.
Separation the database enforces
Row-level security on every table, with the workspace derived from an authenticated session or a validated Origin — never from a request body.
An audit trail for changes
Who changed the prompt, who invited whom, who removed a transcript. Recorded with the actor and the fields that changed.
Deletion that actually deletes
Remove a conversation and its messages go with it. Set a retention window and old transcripts age out on their own.
There are no customer logos or testimonials on this page. The product is young and we do not have permission to publish any yet — so rather than invent them, here is what we will put in writing.
What happens when the bot does not know something?
It returns the fallback message you wrote — by default, that it does not have the information and to contact support. It does not fill the gap from general knowledge. The question is added to a knowledge-gaps list, counted by how often it has been asked, so you can see what to write next.
What can I feed it?
PDF, Word, Excel, CSV and plain text files; audio and video, which are transcribed first; a single URL; or a whole sitemap. Processing runs in the background, so large sources keep going after you close the tab.
Is my content used to train a model?
No. Your sources and conversations stay in your workspace and are used only to answer your visitors’ questions. You can export everything to a single file, or delete a conversation, at any time without asking us.
How is one customer’s data kept away from another’s?
Postgres row-level security on every table, not a filter in application code. The workspace a request belongs to is derived from an authenticated session, or for the public widget from an Origin header checked against your allowed domains — it is never read from the request body.
What does it cost?
There is a free plan that needs no card. Paid plans raise the monthly answer allowance, the number of bots and the recording minutes — one price each, with nothing sold separately. A question the bot declines costs nothing, because no model is called.
Can I choose the model?
Yes — models from OpenAI, Anthropic and Google, picked per workspace, with speed, quality and per-million-token pricing shown next to each. If a cheaper model of the same quality tier would have covered a workspace’s spend, the cost page tells you.
How do I put it on my site?
One script tag. The widget renders inside a shadow DOM, so your styles and its styles cannot affect each other. Colours, copy, position and launcher are edited in a designer with a live preview.
Can it talk to my other systems?
There is a REST API for chat and for adding sources, authenticated with scoped API keys, documented at /developers with a downloadable OpenAPI file. Conversation events can also be posted to your own endpoint as signed webhooks, with the recent delivery log visible in the dashboard.
Who on my team can change what?
Four roles — owner, admin, editor and viewer — enforced by database policies rather than by the screens. A viewer cannot rewrite a prompt even by calling the API directly. Changes are recorded with who made them.
How do I get help?
Email support on every plan. We are a young product and answer our own support, which means you reach someone who can actually change the thing you are asking about.
14 — Start
Find out tonight whether it is any good.
One document is enough. About five minutes, and it costs nothing.
Free plan, no card required. Export or delete your data whenever you like.
- 01Create an account with your email address.
- 02Name your workspace. That is your first bot.
- 03Upload one document, or paste one URL.
- 04Ask it something you already know the answer to.
- 05Then ask it something you never gave it.