Skip to content
Atal AI

It only knows what you told it

Most support bots are a language model with a system prompt and a hope. This is built the other way round: a retrieval system your documents fill, with a model on the end that has no route to anything else.

NHNorthwind HostingReplies in seconds
Hi — ask me anything about our servers or IP leasing.
do you have a 32GB dedicated server in Stockholm?

Yes. NW-32S — $99/mo is available in Stockholm with 32 GB as standard:

  • Xeon E3, 1 TB SATA
  • 50 TB bandwidth, 5 IPv4
can someone call me about bulk pricing?
Of course — I’ll pass this to the team. Where can we reach you?

Your name

Email address

SendSkip
Ask a question…Send

What goes in decides what comes out

There is no training step and no model of your business. Your files are read, split into sections and indexed. A question retrieves the sections that answer it, and those sections are the only thing the model is shown.

That is the whole architecture, and it is why the bot cannot tell a customer something you never wrote down.

Formats
PDF, DOCX, XLSX, CSV, Markdown, HTML and plain text. Or point it at a website and let it read the pages itself.
Updates
Change a file and re-read it. Nothing retrains, and nothing needs rebuilding.
Injection
A section that looks like an instruction is withheld from the prompt entirely rather than wrapped in a fence — fencing provably fails when the question resembles the injection.
Depth
You set how much of the corpus a reply may carry. One FAQ page can be read whole; a four-hundred-page manual cannot, at any price.
PDFPricing 202618 sections
DOCXShipping policy6 sections
WEBsupport.northwindhosting.com41 pages
MDReturns FAQ9 sections

“Standard delivery is 3–5 working days, and returns are free within 30 days.”

Answered from Shipping policy and Returns FAQ

Every message ends one of three ways

Answers

When the retrieved sections clear the confidence bar. The reply cites what it used, in whatever language the visitor wrote in.

Refuses

When they do not. It says so plainly rather than assembling something plausible out of the nearest material.

Hands over

On request, or after a refusal. A ticket is raised with the full transcript, and replies come back into the same thread.

ESC-01042

Bulk pricing for 40 licences

Needs a person
Why it stopped
No pricing above 25 seats in the corpus
Waiting
2 minutes

Full transcript attached · replies land back in this thread

A handover that carries the context

Whoever picks it up sees what was asked, what the bot could not answer and why, and whatever contact details were captured — rather than a cold “customer needs help”.

  • Live handover first, a ticket when nobody is free
  • Details asked for at the moment somebody wants help
  • Email replies land back on the right ticket

It collects contacts at the only moment people give them

Not a form in the corner that nobody fills in. The bot asks after it has already been useful, or at the moment somebody wants a person — which is when a visitor is most willing and least likely to type a fake address.

What it asks for
Any of name, email and phone. A plumber wants a number far more than an address, and the form should reflect that.
When it asks
After a set number of messages, or only when somebody escalates. Browsing visitors are never interrupted.
Whether it insists
Required blocks the answer until it is filled in; optional answers first and lets them skip. Both are defensible; one is your call.
Where it goes
The Contacts page, a CSV export, or straight to your CRM over a webhook. Consent is recorded with every one.

Contacts this week

3 new

  • Ana Ruiz

    Asked for a call

    Bulk pricing, 40 licences

  • [email protected]

    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 wears your brand, not ours

A support widget that looks like the vendor’s rather than the customer’s is an advert somebody else is paying to host. Colour, shape, position, avatar, display name and greeting are all yours — and our badge comes off entirely.

  • Three widget styles and two corner positions
  • Upload an avatar, or link one
  • A live preview beside the settings, before anything saves
  • Notification sound on or off, per bot

Classic

Soft

Square

243 models, charged by what a reply costs

A reply is metered on the tokens it actually spends, not a flat fee per message. So a short question on a cheap model costs a fraction of a long one on a flagship, and the difference is visible before you choose.

  • A different model per bot, and a fallback for when it is unavailable
  • Top up mid-month without changing plan
  • Unused top-up credits do not expire

Cost of one reply

same pool, different rate

DeepSeek V4 Flash1 credit
GPT-4o mini2 credits
GPT-4.124 credits
Claude Sonnet 4.536 credits

Switching model changes how fast you spend, never what you are charged for.

And the parts nobody puts on a landing page

  • Multiple bots per account
  • Team members with roles
  • Two-factor authentication
  • Restrict a widget to your own domains
  • Lead capture with consent recorded
  • Conversation history and search
  • Weekly digest by email
  • Webhooks and email notifications
  • Thumbs up and down on every answer
  • Full transcript export
  • EU data residency
  • Invoices you can actually file