AI & Automation · AI assistants
An assistant that answers everyday questions from your own sources, names the source and says when it does not know. For sales, support, onboarding and management.
01 How you recognise it
Three starting points we see regularly. One of them is enough.
Product details, pricing logic, references: questions inside the team instead of answers in the conversation.
It is spread across documents, offers, tickets and wikis.
Processes, responsibilities and rules are documented nowhere findable.
02 What you get
Whoever may not open a document does not get an answer from it either. Outdated material is marked or removed.
Every answer can be verified. Missing or uncertain knowledge is marked as such.
Test cases before launch, feedback during operation, gaps in the sources get closed.
03 What it looks like
Schematic example with fictitious data
One answer with a source, and one case in which the assistant should say that it does not know.
Question
“What notice period applies in the framework agreement with a public sector customer?”
Answer
“Three months to the end of the quarter. Basis: framework agreement public sector, version of 12 March 2026, section 8.2.” The reference leads straight to the document.
Question without a basis
“How much discount can we give this customer?”
Expected answer
“I have no approved source for that. The stored documents contain no discount rule for this contract type.” Instead of a plausible sounding number.
Permissions
If someone without access to contract documents asks the same thing, the assistant does not answer from the contract but points to the responsible team.
Limit
Source binding, tests and spot checks reduce the risk of errors considerably. They do not rule out wrong answers.
The second and the last point are the decisive ones. An assistant that never says “I do not know” is not a good assistant.
When it fits
Knowledge is scattered, and the same questions get answered again and again.
What a first step delivers
An assistant on a bounded, approved body of knowledge that answers with a source reference.
What we need from you
A selection of current documents, your permission model and typical questions from daily work.
We only need these accesses and documents once we start working together. For a first conversation, a short description of your situation is enough.
04 Definition
An AI assistant in a company is an AI system that answers everyday questions based on approved company sources and supports tasks, with a source reference, respecting access rights and able to name the unknown as unknown.
The difference to a general chat assistant lies in the grounding: answers come from your documents, your CRM and your systems, not from the internet. That makes it reliable and it requires preparation.
A knowledge assistant is an AI assistant specialised in finding and explaining company knowledge: products, processes, policies, projects, with a reference to the original document.
05 Fields of use
Product details, pricing logic, objection handling, references: answers in the conversation instead of questions inside the team.
Solutions from manuals, tickets and the knowledge base, with a source reference so the answer can be checked.
New colleagues find processes, responsibilities and rules without asking every time.
Make earlier answers, text blocks and evidence findable, for faster and more consistent offers.
Travel expenses, data protection, approvals: rules available where the questions come up, instead of as a PDF on the intranet.
Decisions, minutes and status from running projects, for everyone joining or covering for someone.
06 Building it
Which documents, systems and storage may the assistant use? What is outdated, what is confidential? Approval before connection.
Check structure, currency and duplicates in the sources. An assistant on chaotic data delivers chaotic answers.
The assistant respects the access rights of the source systems. Whoever may not open a document does not get an answer from it either.
Answers refer to the source used, and the assistant should mark missing or uncertain knowledge. Source binding, test cases and spot checks lower the risk of errors; they do not rule out wrong answers.
Where the work happens: in the chat tool, in the CRM, in the intranet or in the ticket system. Not another window nobody opens.
Update sources, evaluate feedback, close gaps. An assistant gets better with use, if somebody is responsible for it.
07 Governance
Need to know instead of full access. Permissions come from the source systems and are not duplicated.
Questions and answers are logged traceably, for quality assurance and data protection.
Personal data is only processed where the purpose requires it. Model choice and hosting follow the sensitivity of the data.
Sources have owners and review cycles. Outdated information is marked as such or removed.
08 Next
Once the assistant answers reliably, the next step is often to let it act: prepare cases, propose entries, create tasks. That is the transition to the AI agent, with the same rules and additional approvals.
09 FAQ
Through several measures that work together: the assistant answers exclusively on the basis of approved sources and names the source used. It is configured to mark missing or uncertain knowledge. Permissions limit which sources are accessible to whom. Before launch we check test cases whose correct answer we know, and in operation we evaluate feedback and close gaps in the sources.
A wrong answer cannot be ruled out entirely. That is why spot checks and a simple way to report an answer are part of operation for us, not only of the rollout.
Partly. A complete reorganisation is rarely necessary, but outdated versions, duplicates and confidential content have to be identified. That is a manageable step which pays off independently of the assistant.
That depends on the sensitivity. For many applications providers with EU processing and contractual assurances are suitable; for particularly confidential content self-operated models are an option. We advise openly and document the decision.
Next step
That is exactly where a good assistant starts. We show you what it looks like with your sources and which rules it needs.