An assistant that answers from your documents, and from nothing else.
It answers questions around the clock, drawing on your own documents and staying inside the limits you set. When it cannot find the answer, it says so and hands the conversation to a person instead of inventing something.
- Answers from your documents
- Hands over to a person
- Fixed price after the consultation
Four signs that too much of your day goes into answering.
The same questions every day
Opening hours, delivery times, guarantees, how to send something back. When the enquiries come down to a handful of questions, somebody is typing out the same replies every morning.
Enquiries after closing time
The message arrives at half past nine in the evening and the reply goes out the next morning. By then the sender has often asked somebody else as well.
Knowledge nobody can find
Manuals, instructions, price lists and briefing notes sit somewhere on the shared drive. People still ask a colleague, because that is quicker than digging through folders.
Customers in several languages
German, French, Italian, English: in Switzerland that is the norm rather than the exception. Writing every answer out by hand in every language multiplies the same work.
From the window on your website to the log that shows what people really ask.
An assistant for your customers
A window in the bottom right of your website, in your colours and under your name. It answers questions about your offer, your process and your terms, Sunday evening included.
An assistant for your team
The same machinery behind the login, working from internal material: instructions, price lists, procedures. New staff ask the assistant instead of interrupting the person at the next desk.
Answers from your own documents
We read in what you already hold: manuals, tariffs, terms and conditions, frequently asked questions, product sheets. The assistant searches there first and writes its answer from what it finds.
Several languages in one conversation
It replies in the language of the question, even when your documents exist only in German. If somebody switches language halfway through, it switches with them.
Handover to a person
On sensitive topics, and whenever it does not have the answer, it passes the conversation on: to your contact form, to an email address or into your ticketing system, with the exchange so far attached.
A log of every conversation
You see what people genuinely ask and where the assistant had nothing to go on. Each gap becomes a new section in your documents, and the assistant gets better as you go.
Your documents are the source, not half the internet
The difference between a useful assistant and a dangerous one comes down to a single rule: it may only say what your documents say. In practice that means a search happens before every answer.
What we read in
Price lists, product sheets, terms and conditions, manuals, frequently asked questions, internal instructions. PDF, Word or your website directly. When something changes you swap the document, and the assistant is up to date at once.
How it searches before it speaks
Your documents are broken into short passages and stored so they can be found by meaning rather than by keyword. A question comes in, the assistant pulls the relevant passages first, then writes the answer from those passages alone.
When nothing matches
If the search turns up nothing usable, it does not fall back on guesswork. It says it does not know and offers the route to a person. That is the single most important line in the whole set-up.
An assistant without limits is a risk, not an advantage
What the assistant says, your organisation says. So before it goes live we agree with you what it talks about, what it does not, and when it hands the conversation over. That list is the real work.
What it will not say
No price that is not already published somewhere. No discounts and no commitments about dates. No legal, tax or medical assessment. And nothing at all on subjects that have nothing to do with your business.
What it passes on
Complaints, contract questions, anything involving money or deadlines, and every enquiry where a misunderstanding would be expensive. There, handing over to a person is not a failure, it is the entire point.
Why the limits pay off
An assistant that answers everything will eventually be wrong exactly where it hurts. One that admits it does not know is believed on the rest. Trust comes from what it leaves out.
Assistant, or a chatbot with scripted replies?
In a sales pitch both are called a chatbot. Day to day they are two entirely different things.
Questions nobody anticipated
Assistant on your documents
It understands the question however it is phrased and answers, provided the documents cover it.
Chatbot with scripted replies
What is not in the tree does not exist. The visitor gets a menu or a shrug.
Upkeep when prices change
Assistant on your documents
Read in the new document and you are done. The answers follow automatically.
Chatbot with scripted replies
Every affected reply is edited by hand, separately in each language.
Languages
Assistant on your documents
The same documents serve every language. It replies in the language of the question.
Chatbot with scripted replies
The whole reply tree is rebuilt and maintained again for each language.
Tone of the reply
Assistant on your documents
Written freely in your tone and shaped to the question actually asked.
Chatbot with scripted replies
Exactly the sentence somebody typed once, whoever happens to be asking.
When it does not know
Assistant on your documents
It says so and hands over to a person, with the conversation so far attached.
Chatbot with scripted replies
It offers the same menu again until the visitor closes the window.
Effort to set up
Assistant on your documents
The work sits in the documents and in the limits, not in writing replies out.
Chatbot with scripted replies
Every path is written out, and the effort grows with each new question.
Honestly: if it comes down to three questions and your opening hours, a simple chatbot will do and costs less to keep. If that is your situation, we will say so at the initial consultation.
The choice follows how sensitive your material is, nothing else.
Models
OpenAI, Anthropic, Mistral, open models such as Llama or Qwen where the data demands it
Document search
Vector databases, PostgreSQL with pgvector
Integration
Your website, email, ticketing system
Control
Conversation log, limits, measurement
When we advise against it
- When there is nothing written down for it to answer from: with no documents all that is left is invention, and that is precisely what we are avoiding.
- When the subject carries legal or medical responsibility and nobody is going to read the answers over.
- When the expectation is that it will sell on its own, with nobody looking at what it writes to people.
- When your data may not leave the building: then the route is not this one but AI running on your own server.
What does an AI assistant cost?
It depends on how much material is read in, how many languages it answers in and what it is connected to. That is why we do not name a figure before we know. At the free initial consultation we settle the scope, and you then receive a fixed price for exactly that scope. On top of that there is a running cost per question at the model provider, unless the assistant runs on your own server.
Free initial consultationDoes the assistant make answers up?
It writes only from what it finds in your documents, and it says so when it finds nothing. No technique rules a mistake out entirely, and anyone promising you that is promising too much. That is exactly why the limits, the conversation log and the handover to a person exist wherever a wrong answer would be expensive.
What material do we need for this?
What you already hold: price lists, product sheets, terms and conditions, manuals and the replies you write by hand today. PDF, Word or your website will do. If nothing is written down, that is where we start: setting out your most common questions properly helps your team even without an assistant.
How does the assistant get onto our website?
Through one short line in the page source, whatever the site was built with. The window then appears in the bottom right, in your colours. The internal assistant sits behind the login and is not visible from outside at all. Either one can also be attached to email or to your ticketing system.
Do our documents stay in house?
That is decided by the route you choose. Through a provider’s API, the passages found for a given question go to their servers to be answered; what the provider may do with them is set out in their terms, and we go through those with you. Where that is not acceptable, an open model runs on your own server and nothing leaves your network.
Does the assistant replace someone in support?
No, and that is not how we sell it. It takes the repetitive part off the pile so your people have time for the cases where a human being genuinely matters. Anything that looks like a complaint, a contract or money we deliberately pass on rather than leave to the assistant.
Tell us the question you answer most often.
At the free, no-obligation initial consultation we look at your documents and your most frequent enquiries and tell you honestly whether an assistant is worth it. We reply within 24 hours.
Book an initial consultationJust a quick question? The AI assistant in the bottom right answers straight away.
Ready for the summit?
Tell us about your project. We reply within 24 hours with an honest assessment. Free and without obligation.
hallo@aurphi.chAurPhi MoukrimMarktstrasse 18, 8853 Lachen SZ, SwitzerlandPostal address, no walk-in service. Appointments by arrangement.