What is a grounded AI assistant for business?
A grounded AI assistant for business answers questions from your own documents, catalogues and systems, cites where each answer came from, and says "I don't know, let me get a person" when the answer is not there. It is the opposite of a general chatbot. It does not improvise, it does not quote a price it has not been given, and it does not make up a policy. It sits inside the software you already run: your website, your customer portal, your mobile app, your kiosk, and answers by text or by voice.
AtashiTech's assistant, AtashiTech AI, is available now. It answers the chat on atashitech.com from our own catalogue, and it is built into our visitor, induction and health products as the safety and help assistant. We bring the same engine to your existing systems: we ingest your documents and data, connect the assistant to your web and mobile applications, add voice where it helps, and hand difficult questions to your people with the conversation attached. There is no free trial for this product; it is set up per company. Contact sales and we will scope it with you.
One assistant for every language your visitors speak
A visitor asks in their own language, typed or spoken. The assistant answers in the same language from the same documents, shows where the answer came from, and reads the reply aloud in a voice for that language. You keep one set of documents, not one per country.
- English
- Deutsch
- Français
- Español
- Italiano
- Nederlands
- Polski
- Português
Example conversation. The questions and answers come from the documents you supply.
Source: site-safety-rules.pdf, page 4
Why most company chatbots get switched off
The chatbot that knew everything
A general-purpose chatbot is plugged into the website. It is fluent, confident and wrong. Asked what a product costs it invents a figure. Asked about a return policy it describes a policy the company does not have. Asked something it cannot know, it answers anyway. The support team spends its time correcting the bot, legal asks who approved it, and within a quarter the button is removed. The staff who were supposed to be freed up are back to answering the same twenty questions by email.
The assistant that knows your documents
A grounded assistant is given your material: product pages, manuals, policies, price lists where you choose to include them, safety rules, FAQs, the catalogue from your own system. Every answer is built from passages in that material and carries a citation to them. If the material does not contain the answer, the assistant says so and offers a person, passing the whole conversation to your inbox so the customer does not repeat themselves. Commercial questions you have not approved for the assistant are not answered at all; they go straight to a person. That rule is enforced in code, not requested in a prompt, because we learned the difference the hard way on our own website.
From your documents to a live assistant in four steps
Ingest
Upload documents in the admin portal or connect a source that syncs on a schedule: a product catalogue from your ERP or e-commerce API, a knowledge base, a policy folder. Documents are split into passages and indexed in your own tenant. Nothing from one company is ever visible to another.
Tune
Set the assistant's name, tone and languages, which topics it may answer and which must go to a person, the opening sentence, and the length of spoken replies. Test it in the portal against real questions before anyone sees it.
Connect
Add the assistant to your website with the floating widget, to your Android or iOS app with the native component, to a kiosk or tablet, or to your own portal through the API with a service credential. Voice uses press-to-talk and natural speech in the customer's language.
Hand over and learn
When the assistant cannot help, the conversation lands in your agent inbox with everything the customer said. The portal shows what was asked, what was answered, what was cited and what was escalated, so you add the documents that are missing.
What is included
Answers with citations
Every answer is built from passages in your documents and names them. A customer, or an auditor, can see where it came from.
Refuses instead of inventing
No passage, no answer. The assistant says it does not know and offers a person. Topics you mark as sensitive are never answered by the model at all.
Human handoff
Escalation into your existing inbox or our live-chat console, with the transcript attached and an alert to your team.
Voice in and out
Press-to-talk speech recognition and natural spoken replies, shortened to what a person wants to hear, in the languages you enable.
Floating widget
The same assistant button on your website and in your apps, opening a voice-first chat panel, with your branding.
Scheduled sync
Connect a catalogue or knowledge base and the assistant re-indexes on a schedule, so it never answers from last year's product list.
API and service credentials
Each of your applications gets its own credential; your keys for the language and speech services stay on the server and never reach a browser or an app bundle.
Multi-tenant by design
Departments, brands or your own customers each get an isolated tenant with its own documents, assistant and credentials, under one portal.
Languages
Answers in the language of the question; spoken voices per language; your documents can be in more than one language.
Conversation analytics
What was asked, answered, cited and escalated, by day and by topic, so you see the gaps in your documents and the questions worth a new page.
Your data stays yours
Documents and conversations live in your tenant on servers you choose. Your material is not used to train public models.
Bring your own model keys
The language and speech services run on your own accounts, so usage is billed to you at the provider's rates and can be rotated without touching the assistant.
Available wherever your customers already are
One assistant, one knowledge base, the same answers on every channel.
The assistant is in production today on our own website and inside our own products. Channels marked On request are integration work we scope with you.
Adding AI to the systems you already run
Most companies do not need a new system; they need an AI assistant for business questions on top of the one they have. That is what we call AI migration: keeping your ERP, CRM, portal, website and apps exactly as they are, and placing a grounded assistant on top of them. The work follows a fixed pattern. We map the sources of truth: which documents, which database tables, which APIs. We agree what the assistant may and may not answer, and who receives what it cannot. We connect the assistant to each channel your customers and staff already use, with your branding and your languages. We run it in parallel with your people for a period, review every escalation, and fill the gaps in the documents before it answers unaccompanied.
Because the engine is already in production on our own website and inside our visitor, induction and health products, the migration is integration and content work, not research. Typical first assistants: a product and policy assistant on a website, an internal help assistant for staff over procedures and manuals, a safety assistant on a factory kiosk, a patient-information assistant in a care app, and a support assistant inside a customer portal that knows the customer's own orders through your API.
Assistants already running on this engine
This website
The chat button on atashitech.com answers from our catalogue and hands over to a person when it cannot. Try it.
Safety assistant
Inside our induction products: answers from the site's own rules and manuals, by text or voice, on kiosks and phones.
Care assistant
In a family health-monitoring app, with press-to-talk voice for older users and a strict rule never to give medical advice outside its documents.
Customer support
A support assistant in a customer portal that answers from policies and, through your API, from the customer's own account.
Internal help desk
Staff ask about procedures, HR policies and system how-tos and get the paragraph, with the source, instead of a ticket.
Product advisor
On e-commerce and catalogue sites: which product fits, what it includes, what the documentation says, never a price you did not publish.
Shaped to your company from the start
This product is always delivered as a customised version. There is no self-serve trial, because an assistant is only as good as the documents and the rules behind it.
Every engagement starts with a scope: your sources, your channels, your languages, your escalation path, and the questions the assistant must never answer. Integration with your ERP, CRM, help desk or e-commerce platform, single sign-on for staff assistants, on-premise or in-country hosting, and WhatsApp, Teams or Slack channels are all part of that scope. Delivery is by the team behind our cloud and integration services and web application development. Tell us which system should start answering questions and we will come back with a scope, not a sales pitch.
Give your existing systems an assistant that only tells the truth
Available now. Set up per company, in your systems, with your documents.
Contact sales