Make what your organisation already knows findable by the people who need it.
Most knowledge problems are not missing knowledge. They are knowledge that exists, in three places, in two versions, owned by nobody, and faster search over that produces confident wrong answers.
What is our refund window?
+ 37 more
A query returns forty documents, which is the same as returning none if the reader cannot tell which one is authoritative.
Four ways an organisation stops being able to use what it knows.
Putting an AI answering layer over this is the most common mistake in the category. It does not resolve a contradiction; it picks one side of it, fluently, without telling anyone it chose.
The answer exists in more than one version
A policy document, a wiki page and a message thread each say something slightly different, and none is marked as the current one.
Nobody owns the content
It was written for a project that ended. It is still there, still findable, and nobody is responsible for whether it is still true.
The people who know are the interface
Requests route to individuals rather than to sources, so knowledge availability depends on who is at their desk.
Search returns everything and resolves nothing
A query returns forty documents, which is the same as returning none if the reader cannot tell which one is authoritative.
Four costs, and the last is the one that shows up in an audit.
New people take longer to become useful
Onboarding time is mostly knowledge-retrieval time, and it is paid again with every hire.
Experts are interrupted instead of working
The most knowledgeable people become a lookup service, which is the most expensive possible way to answer a repeated question.
Decisions get made from stale information
Not from no information, which is safer, but from information that used to be right.
- The one that shows up in an audit
You cannot show what your policy was
When a customer, auditor or regulator asks what the process was on a given date, an unversioned wiki cannot answer.
The workflow afterwards.
- not marked currentSeveral versions coexist, none marked current.
- not marked currentContent has no owner.
- not marked currentAnswers arrive without provenance.
- not marked currentA question outside the sources gets answered anyway.
- not marked currentNobody knows what people are failing to find.
Retired
- retired
- retired
- retired
- retired
- One source is authoritative per topic, and the others point to it or are retired.
- Each topic has a named owner and a review date, so staleness is visible rather than assumed.
- Every answer shows what it came from, so the reader can judge it.
- It returns nothing and routes to a person. Declining is a designed behaviour.
- Unanswered queries are visible, which is the most useful list for deciding what to write next.
Four things to fix before adding retrieval, in this order.
Retrieval quality is bounded by source quality, so the sequence matters more than the tooling. Each step below is cheaper than the one after it, and skipping any of them makes the later ones harder.
Ownership per topic
Why it comes first
Somebody has to be able to say which version is right
What happens if skipped
Retrieval picks arbitrarily and the contradiction becomes invisible
Retirement of superseded content
Why it comes first
Deleting the wrong answer is more valuable than ranking the right one higher
What happens if skipped
The old version keeps surfacing, and trust in the system drops permanently
Provenance on every answer
Why it comes first
A reader who can see the source can catch an error; one who cannot, cannot
What happens if skipped
Wrong answers propagate with the authority of the system that produced them
A defined way to decline
Why it comes first
The system must be able to say the sources do not cover this
What happens if skipped
It answers anyway, plausibly, and that is the failure that reaches a customer
Step five, retrieval
Only after those four does retrieval technology matter, and by then the choice is usually straightforward. Most disappointing knowledge projects bought step five and skipped one to four.
How the work runs, step by step
Start from the questions, not the content
Collect what people actually ask, from tickets, message threads and the requests that reach experts. That list defines the scope; the document library does not.
Find the contradictions
For the top questions, locate every version of the answer. The count is usually higher than expected and is the most persuasive artefact of the whole engagement.
Assign ownership and retire the rest
One authoritative source per topic with a named owner. This is organisational work, not technical, and it is where the value is.
Add retrieval with provenance and a stop condition
Answers cite their source and the system declines when the sources do not cover the question.
Track what it could not answer
The unanswered list is the content roadmap. A knowledge system that cannot tell you what it failed at cannot improve.
What this connects to, and who builds it
- The answering surfaceAI chatbot developmentA conversational interface for staff or customers is AI chatbot development.
- Source reliabilityData engineeringWhere the underlying information is scattered or contradictory, that is data engineering first.
- Keeping it honest in productionMLOps and LLMOpsEvaluating answer quality and detecting drift after launch is MLOps and LLMOps.
- Connecting to existing systemsAI integrationLanding answers inside the tools people already use is AI integration.
Questions we are actually asked
Pick a topic, or ask us directly. We answer every inbound within one business day.
Still have questions?
Talk to a senior engineer, not a bot.
Where this work has been done
Answer quality, provenance and knowing when a system should decline are the substance of this work, and they are already how Viithiisys runs AI in production.
Fitelo, evaluation and cost controls in production
Directly relevant here, because measuring whether generated output is still correct is exactly the discipline step four of the ordering above requires.
Conscious Chemist, product-question response
A documented 38% faster product-question response. A product surface answering customer questions from product information, which is the same discipline applied to customer-facing information.
Vizitor, in daily multi-site use
In production at scale: 500+ workplaces across 15+ countries.
Tell us the five questions your experts answer most often.
That list shows whether this is a knowledge problem, a documentation-ownership problem or a product problem. Two of the three need less work than people expect.