AI Knowledge Base with Sourced Answers
Your team asks in plain language and gets the answer out of your approved documents – with the passage it came from and the date beside it. If no source covers the question, the system says so instead of guessing.
The knowledge exists, just not as an answer
Every company of this size has written its knowledge down: manuals, inspection records, framework agreements, project folders, plus a few thousand emails in which the actual decision was made. All of it was written for a different purpose – for the handover, for the manufacturer, for the auditor. None of those documents is an answer to the question an engineer has at half past seven on site. So he rings the colleague who knows it by heart, and interrupts her work for the fourth time that morning.
Finding things stopped being the hard part a while ago. Semantic retrieval locates the right paragraph even when it is worded nothing like the question. The real problem is working out which paragraph still applies: three versions of the price list sit in the same folder, the 2023 rule was overturned by email, and the draft already contains the new flat rate that nobody has approved. On content alone, all four read equally convincing. They can only be told apart by what is not in the text: who owns the document, since when it applies, whether anyone signed it off.
Then comes the second hurdle. An answer you cannot verify is an answer nobody uses twice. Anyone quoting a warranty period to a customer wants to know where the figure comes from – file, section, date. So we do not build an answering machine, we build an evidence machine: every statement hangs on a passage, and where there is no passage there is no statement. That is uncomfortable, because the system then stays quiet more often than it technically would have to. It is also the only reason anyone still believes it after four weeks.
What your team actually uses it for
We start with the area where questions come up most often and the sources are clearest. Further areas follow later and share the same evidence logic.
Service and engineering in the field
The most common starting point: questions from site where the answer sits in a manual nobody carries up a ladder.
Prices, terms and contracts
Which discount is covered, what does the framework agreement say, does the payment term apply to new customers – questions where the evidence matters more than the speed.
Past projects and quotes
How did we solve this last time? The answer usually sits in a 2023 project folder that nobody carries in their head any more.
Onboarding new colleagues
The questions of the first few weeks repeat almost word for word. Sourcing them once is cheaper than answering them fifteen times.
Regulations, inspection intervals and safety
Here the summary is not the point; the reference to the valid version is – including a warning when a requirement has aged.
Answers while the customer is on the line
Internal sales need a statement within twenty seconds that still holds when somebody asks a follow-up question.
What an answer actually looks like
On the left the answer to a question from internal sales, on the right the evidence behind it – passage, version date and owner. Part of the question stays deliberately open.
Customer reports a fault on the installed unit after 14 months. What do we cover – and what do we charge for a 60 km call-out?
Our installation work carries a two-year warranty from handover, so this case falls inside it.1
The faulty part goes through the manufacturer guarantee: replacement free of charge, on-site labour billed at the service rate.2
This requires a documented maintenance record. Without maintenance the manufacturer guarantee lapses.3
-
1
Terms for installation work Clause 6.2 · page 4
Claims for defects in installation work expire two years after handover.
valid · 04/2026 · Finance lead -
2
Service manual Chapter 3.4 · page 18
Under guarantee the manufacturer supplies the part; labour is billed by the installing company.
valid · 11/2025 · Service manager -
3
Manufacturer maintenance policy Section 2
The guarantee requires annual maintenance by an approved installer.
check version · 06/2023 · over 24 months old -
Service price list 2026 Call-out zones
Draft without sign-off. It does contain a zone rule beyond 50 km, but nobody has confirmed it – so it is not quoted.
not quoted · owner unconfirmed
The last block is the one that matters: the rule for long call-outs does exist in the building, but only in a draft. The system refuses to use it because nobody approved it – and says so openly.
Evidence 3 is marked amber because its version is more than two years old: the answer uses it but flags it. You set the age threshold per source.
From source to sourced answer
-
Define the knowledge space
Together with the departments we decide which repositories belong in it, who owns each source and what deliberately stays out: personal folders, drafts, the mail archive.
-
Index with provenance
Documents are split into passages. Each passage keeps its file, chapter or page, version date, owner and the permission from the source system – without those details there is no evidence later.
-
Answer and cite
For a question the system retrieves the matching passages and forms the answer from them. What is not in those passages is not in the answer, and every statement carries its reference.
-
Hand over instead of guessing
If coverage is too thin, or two valid sources contradict each other, the system does not answer. The question goes to the owner of the topic with the thread so far – as a Teams message or a ticket.
-
Make the gaps visible
Unanswered questions, overdue sources and the most frequently cited documents go into a monthly report. That report produces the short list of documents actually worth maintaining.
What changes day to day
Today
- One person knows the answer, and she is out on site
- Three versions of the same price list, none of them definitive
- Search returns files, the question stays open
- New colleagues ask the same things every week
- Anyone who wants certainty digs out the original again
With sourced answers
- The answer appears in seconds, with chapter and page
- Only the approved version is quoted, drafts stay out
- The question produces an answer, not a list of hits
- Recurring questions are handled by the system, not the colleague
- The evidence sits next to the answer, one click from the original
Where a knowledge base does not help
Four points we raise before quoting. One of them is a reason not to run the project at all:
- If your knowledge lives in a handful of well-kept documents, this is not worth it. Five current PDFs and three people working with them are served faster and far more cheaply by a tidy repository and full-text search. The effort only pays off when many people ask often and the documents keep changing. If that is how we read it in the intro call, we will tell you so.
- The knowledge base does not replace order, it makes disorder visible. When two valid sources contradict each other, the system can report it but cannot decide which one is right. That decision needs a person with authority – and the whole project hangs on it. Without a named owner per source the clarifying work is left undone, and answer quality slides back within months.
- The language model remains a phrasing engine. It summarises retrieved passages, and with tables, drawings and forms full of cross-references the summary is sometimes worse than the original. That is why the citation is mandatory rather than decorative. For pricing, legal and safety questions the rule still stands: read the passage before you commit.
- Permissions are only as precise as the source system. We inherit the rights from SharePoint, Drive or the file share – where everyone may see everything there, the knowledge base cannot filter either. Reworking a permission model is a project in its own right with its own effort; we do not quietly price it in, we name it separately.
Fits your repositories
Scope and price
The entry price covers one knowledge space with up to three source systems, including mandatory citation, permissions and a refresh schedule.
- One knowledge space with up to three source systems
- Indexing with provenance: file, passage, version date, owner
- Permissions inherited from the source system
- Access through Microsoft Teams or as a web interface
- Mandatory citation and a defined answer limit with handover to the owner
- Refresh schedule per source plus a monthly gap report
- Documentation, handover session and 30 days of support
What increases the price
- Further source systems, or one without a usable interface
- Scanned archives without a text layer, so with OCR
- Separate knowledge spaces per department with their own rights
- Operation with a model inside your own environment
- Sources and answers in several languages
Several separate knowledge spaces, scanned archives or operation inside your own environment typically land in the range of our Workflow Enterprise package from €5,900. We quote the binding fixed price after the intro call.
All prices excl. VAT · operation and further development optionally via a support package
What you get
-
Production knowledge search
In Teams or the browser, signed off with real questions from your day-to-day work rather than demo data
-
Source register with owners
Which repository belongs in it, who owns it, how often it is refreshed – and what is deliberately excluded
-
Catalogue of verified test questions
Real questions with the answers we agreed on, so you can re-check every later change against them
-
Handover session and gap report
How your team asks, how owners respond and how you spot that a source has gone stale
Frequently asked questions about the AI knowledge base
These solutions fit alongside
AI Chatbot for Customer Support
The same sourced knowledge facing outward: customer questions around the clock, with handover to your team.
AI Inbox Assistant
Draft replies for incoming email, fed from the same approved sources.
Automatic Document Filing
The order that comes first: when every document lands properly, the knowledge base stays maintainable.
Internal Ticketing System
Where the questions go that no source answers – tracked instead of shouted down the corridor.
How often does your team ask one person the same thing?
In the free intro call we take ten real questions from your working week and check whether the answer sits in an approved source. Afterwards you know whether you are missing a knowledge base – or a tidy repository first.
Book a free intro callWhere AI knowledge base creates value in everyday work
Employees search approved documents in natural language and receive answers with specific sources, permissions and freshness status.
Three concrete operating scenarios to compare with your own process.Find policies faster
Answers point directly to the valid passage instead of a long document list.
Make project knowledge accessible
Approved decisions, manuals and handovers become searchable by role.
Detect knowledge gaps
Questions without a reliable source are collected and returned to content owners.
A strong fit when …
Recurring requests can be answered from approved knowledge while uncertain cases are handed to people with clear context.
- You handle recurring knowledge queries using repeatable rules.
- The intake, target system and accountable business role can be named clearly.
- Exceptions are allowed to remain visible and move to people deliberately.
Deliberate automation boundary
Unsupported answers, stale sources and documents outside the user’s permissions are not presented as reliable answers.
Explore the technical approach and platformsEstimate time savings with your own volume
The calculator uses 10 minutes today and 3 minutes after automation as fixed example assumptions. It does not replace process analysis.
Illustrative estimate based on the visible assumptions — not a guarantee.