Answers from your own documents
Not from the open internet. Upload the PDFs, price lists and help pages you already have — Atal AI answers your customers from those, in their own language, and fetches a person the moment it cannot.
500 credits a month free, no card. 243 models on the top plan.
Yes. NW-32S — $99/mo is available in Stockholm with 32 GB as standard:
- Xeon E3, 1 TB SATA
- 50 TB bandwidth, 5 IPv4
Your name
Email address
Answers in seconds
Captures qualified leads
Hands over to your team
Works around the clock
It cites what it used
Every answer names the sections behind it. If your documents do not cover the question, it says so rather than assembling something plausible.
One script tag
No build step, no plugin, no retraining. It renders in a shadow root, so your CSS cannot break it and it cannot break your site.
You pick the model
243 of them on one credit pool, from the cheapest to the most capable. Switch whenever you like — it changes how fast you spend, not what you are charged for.
How it works
Three things happen, and the second one is the difference
01
It reads what you give it
PDFs, Word files, spreadsheets, or your website address. Everything is split into sections and indexed — and that index is the only place an answer may come from.
- Re-read a file after you change it; nothing to retrain
- Instruction-shaped content is quarantined, not fenced
“Standard delivery is 3–5 working days, and returns are free within 30 days.”
Answered from Shipping policy and Returns FAQ
“I don’t have that in my reference material.”
The sentence most support bots will not say. Below the confidence bar it refuses and opens a ticket, rather than reaching for the nearest paragraph and inventing a returns policy out of it.
02
It refuses when it should
A bot that answers everything is a bot that invents things, and one invented policy costs more trust than a hundred correct answers earn.
03
It hands over with the context
Live first, a ticket when nobody is free. Whoever picks it up sees what was asked, why the bot stopped, and how to reach the visitor.
- Email replies land back on the same ticket
- Never promises a person when there is not one
ESC-01042
Bulk pricing for 40 licences
- Visitor
- [email protected]
- Why it stopped
- No pricing above 25 seats in the corpus
- Waiting
- 2 minutes
Full transcript attached · replies land back in this thread
What it will not do
Said here, rather than discovered in month two. If any of these is a dealbreaker, better to know before you upload anything.
Answer from the open internet
Only your documents. There is no fallback to general knowledge, ever.
Pretend to be a person
It says what it is. A support bot passing as human is a complaint waiting to happen.
Promise a callback at 3am
It offers a handover only when somebody can actually take it.
Charge for a reply it could not give
A refusal costs a fraction of a credit, not a full one.
The conversations it cannot finish are worth something
A visitor asking a question your documents do not cover is not a failure — it is somebody who wants to talk to you. The bot asks for a name and an email at exactly that moment, which is the only moment anybody willingly gives one.
- Ask for a name, email or phone — or only on escalations
- Consent recorded with every contact
- Export to CSV, or push to your CRM by webhook
Contacts this week
3 new
Ana Ruiz
Asked for a call
Bulk pricing, 40 licences
Left an email
Stockholm, 32GB, 100TB
Priya N.
Asked for a person
Migration from a competitor
Consent recorded with each one · exports to CSV, or straight to your CRM by webhook
It should look like you, not like us
Your colour, your name, your avatar, your greeting. Three shapes, two positions, and an option to turn our own badge off entirely — the widget on your site should not be advertising ours.
- Accent colour, corner style and position
- Upload an avatar, set a display name and a welcome line
- Preview every change beside the settings, before you save
Classic
Soft
Square
Charged by what a reply costs
Not a flat fee per message. A reply is metered on the tokens it actually spends, so a short question on a cheap model costs a fraction of a long one on a flagship — and you can see the difference before you choose.
Every model draws on the same monthly allowance. Nothing adds a second bill.
Cost of one reply
same pool, different rate
Switching model changes how fast you spend, never what you are charged for.
Live this afternoon
- 01
Upload what you have
Drag in your documents, or give it your website address. Indexed in minutes.
- 02
Set the tone and the limits
Pick a persona and a model. Decide when it should fetch a person, and what to ask for first.
- 03
Paste one line
Before your closing body tag. WordPress, Shopify, Webflow or plain HTML.
<script src="https://api.atalai.com/widget.js" data-key="wk_live_…" async></script>
243 models, one allowance
Use the cheapest for routine questions and the most capable where it matters. Both spend from the same pool.
Your people pick up where it stops
Everything the bot could not finish lands in one inbox, oldest first, with the transcript and the reason it stopped already attached. Nobody has to ask the customer to explain again.
- Reply in the widget while the visitor is still there
- Or by email, and it threads back to the same ticket
- Unlimited team members on every plan
Needs a person
Ana Ruiz
2m
Bulk pricing for 40 licences
Stockholm capacity
Priya N.
1h
Migrating from a competitor
Questions people ask first
How long does it take to set up?
An afternoon. Upload your documents or point it at your website, pick a persona and a model, then paste one script tag. Indexing a typical knowledge base takes a few minutes, and there is no training step to wait for.
What stops it inventing an answer?
It can only see the sections retrieved from your own documents — there is no fallback to general knowledge. When nothing retrieved clears the confidence bar it refuses and offers a person instead, rather than assembling something plausible from the nearest material.
What happens when it cannot answer?
It says so, asks for a way to reach the visitor, and opens a ticket carrying the full transcript and the reason it stopped. If somebody on your team is available it offers a live handover first. Replies by email land back on the same ticket.
Which languages does it work in?
It replies in whatever language the visitor writes in, from documents written in any language. There is nothing to configure and no per-language setup — the model handles the translation as part of answering.
How is it priced?
One pool of credits a month, spent per reply. A reply is metered on the tokens it actually uses, so a short question on a cheap model costs a fraction of a long one on a flagship. Switching model changes how fast you spend, never what you are charged for.
Can I use my own AI model?
You choose from the models on your plan and set a different one per bot, plus a fallback for when your first choice is unavailable. Bringing your own API key or self-hosted model is not supported today.
Where is my data stored, and can I get it out?
In the EU. Conversations, documents and contacts export to CSV at any time, and webhooks can push events to your own systems as they happen. Deleting a document removes it from the index immediately.
Will it look like your product or mine?
Yours. Colour, corner style, position, avatar, display name and greeting are all configurable, and the “Powered by Atal AI” line can be turned off entirely.
Put your own documents behind your support
Free to start. Paid plans from $99.99 a month, and you will know within an afternoon whether it answers your customers properly.