Home/ Journal/ An AI Knowledge Base for Employees That Quotes the Exact Passage

An AI Knowledge Base for Employees That Quotes the Exact Passage

An AI knowledge base for employees is a search box over the company's own documents that answers in plain words and shows the exact passage it used. The useful ones do four things beyond answering: quote the source, say when two documents disagree, say "not in the documents" when they are silent, and withhold files from people who should not see them. We built one on sample data so you can watch each of the four.

The problem, in the owner's words

The owner's sentence is: "The answer exists somewhere in our documents, and people ask the same three colleagues every day." The prices are in a price list, the expense rule is in a policy, and the person who knows which is current is busy.

The usual fix is a chatbot over the shared drive. The risk is that it answers confidently from an old file or from a file the asker should never see. Each of the four behaviours in the demonstration exists to close one of those holes.

The job this replaces, and what it costs

Companies post for this under titles such as AI engineer or AI specialist, and the duty is to connect a model to internal documents and keep it honest. The nearest cluster we track is AI and machine learning engineer.

ColabContent's role finder read 551 US postings in that cluster between September 2 and October 6, 2026. Only 16 listed a pay range, which is fewer than 20, so treat the number as thin: the median listed range ran from $120,000 to $165,000, with a median midpoint of $145,000. Benefits were 30.0 percent of total compensation (BLS ECEC, June 2026), about 43 cents on each wage dollar. The detail is on our AI engineer page.

How the demonstration works, step by step

The knowledge desk below is a demonstration system running on invented sample data. No client uses this knowledge desk, and nothing in it is a result anyone achieved. Here is the knowledge desk recording, and below it the same path in order.

Demonstration system, sample data. About 38 seconds.
  1. A staff member asks a question in plain words. The page first says how many documents are open to that person: 24 of 25 for a regular staff member.
  2. The answer comes back marked "Answered from the documents", and quotes the exact sentence with its document and section. The after-hours call-out fee is $185, quoted from section 2 of the price list, and a footer confirms 1 of 1 quotes were found word for word in the source.
  3. When two documents disagree, it says so. The expense per day appears as $60 in the 2026 policy and $40 in the 2024 policy; the page names the 2026 file, updated February 3, 2026, as the newer one and marks the 2024 file superseded. Both passages are quoted.
  4. When the documents are silent it says "Not in the documents." and logs the question as a gap, rather than guessing.
  5. A restricted document, the salary bands, is withheld from a regular staff member. The banner names the document and who may open it, and no figure from it appears anywhere on the page. A manager asking the same question sees 25 of 25 documents and the answer.
  6. A second view for the person who maintains the documents shows what is asked most and what is missing.

What the sample data shows

In the sample history, 86 questions were asked: 70 answered, 13 with no answer and 3 stopped by permissions, and 70 + 13 + 3 = 86 is shown on screen. The most asked question, how many vacation days someone gets, was asked 14 times. The gaps list has 4 questions, including one about a drone that was asked once.

The history is seeded sample data, so those counts show what the view reports, not how any real company behaves. We did not record a separate automated check count for this system, so we state none.

Hire it, rent it or own it

The job a knowledge desk does can be done by a person, by rented software or by a system you own. This table sets hiring, renting and owning a knowledge desk side by side, using only what is on this page.

 Hire the roleRent softwareOwn a system built for you
What it costsSalary and benefits; thin data for this cluster (16 listed ranges), median midpoint $145,000A monthly fee for as long as it is usedOne fixed fee, from $10,000, quoted after the audit
Whose documents it readsWhatever the person connects firstOften a fixed set of connectors the vendor choseYour documents, with your rules on who may see what
Showing its sourceDepends on how it was builtVaries by productEvery answer quotes the passage and names the document and section
When it does not knowDepends on the buildOften answers anywaySays "Not in the documents." and logs the gap

What building it for your company involves

To build this for your company we begin with the documents people ask about most and the rules on who may see which one. Restrictions come from your existing access, not from a new list we invent.

Tell us the job you would hand over and we build a knowledge desk around your company. A working prototype on your own data comes first, in 7 to 10 days, and you pay nothing for the build until you have seen the knowledge desk work. Custom systems such as a knowledge desk start at $10,000, fixed, and everything we build is yours at handoff. If you want help with the knowledge desk afterwards, care is $997 a month.

Questions people ask

How does an AI knowledge base know which document is newer?

In the demonstration each document carries an updated date and a superseded flag. When two disagree, the answer quotes both and names the newer one. In your version those dates come from your own files.

Can it leak a restricted document?

In the demonstration it does not: a regular staff member asking about salary bands sees a restricted banner and no figure. Your version would take its permissions from your existing access rules, and we would test that case first.

What happens when it does not know?

It says "Not in the documents." and records the question as a gap, so the person who maintains the documents can fill it.

Is a company using this knowledge desk today?

No. The mechanical contractor, its people and its 25 documents are made up, so the 86 questions in the history are seeded. What is real is the behaviour you can watch: quoting, flagging disagreement, admitting silence and withholding.

What would a knowledge desk for our company cost?

It is a custom system from $10,000, fixed, and the price is set after the $499 audit tells us which documents matter. A prototype over your own files comes first, within 7 to 10 days, and the code and accounts are yours afterwards.

Keep reading

See other working demonstrations beside the knowledge desk in the prototype library, the hiring data on AI engineer pay and true cost, and AI for professional services firms. Or order the $499 AI-Ready Audit.