Case Study: Internal AI Knowledge Assistant

Company Brain Case Study: An AI Assistant That Knows How One Telecom Company Works

Staff at a telecom provider ask questions in plain English and get answers from the company's own procedures, with the source attached.

By Jude Horak | Updated November 15, 2026

Short answer

A company brain is an AI assistant that answers from one company's own procedures, training and past decisions, and shows the source behind each answer. For a telecom provider, I built one that searches about 3,500 reviewed records, respects who is signed in, drafts replies staff approve before sending, and flags conflicting or missing information instead of guessing.

About 3,500

searchable records, built from the wiki, training files, reviewed chat history and documents

400+

wiki pages imported as safe text, with every inventoried page given a recorded outcome

About 850

training videos, screenshots and files described or transcribed, with their limits labeled

100+

automated tests covering answers, access rules and the approve-before-send flow

How Does a Question Become a Sourced Answer?

Step 1

Inventory every source

Wiki pages, attachments, training videos, screenshots and past chat answers, each with an owner and a date.

Step 2

Clean and label

Sensitive material is held back. Every record keeps its source, date and any known conflict.

Step 3

Check who is asking

The employee signs in with their company account. Results they are not cleared for never leave the service.

Step 4

Search and read the source

The assistant finds matching records, then opens the exact source instead of trusting a snippet.

Step 5

Answer with sources, a person decides

Answers link to their sources. Drafted replies wait for the employee to edit and approve before anything is sent.

What is a company brain?

It is an internal knowledge assistant for employees. It knows where your procedures, training and past decisions live, finds the right piece for the question in front of someone, and shows where the answer came from. A useful answer from it is scoped to the person asking, cites evidence, and names the next step when the evidence runs out.

The client here is a telecom provider that bills more than 3,000 business customers every month. I am the only engineer there. Answers to staff questions were spread across a wiki, training videos, screenshots, attachments and past chat threads.

How is it different from a chatbot with uploaded documents?

Uploading a folder gives you a pile. This build tracks every source before it is searchable. Each wiki page got one recorded outcome: imported, imported with parts withheld, empty, or excluded. Training videos are labeled by what was actually captured, a full transcript or only a reviewed summary.

Every record keeps its source date, or says the date is unknown. When old and new documents disagree on a rule or a price, both stay visible with their dates, so an outdated page cannot quietly win.

Who can see what?

Staff sign in with their own company accounts. Each employee is cleared for specific collections of knowledge, and that check runs inside the service before any result, snippet or source leaves it. The same check applies when the assistant opens a full source by its ID, so nobody can go around search to read something restricted.

Sign-in sessions are short lived and renew on their own. Instructions written inside a document are treated as content, never as commands.

How does it stay accurate when procedures change?

The knowledge is rebuilt by repeatable import steps with stable IDs, so a rebuild produces the same records from the same sources and every change can be explained. A coverage report reconciles what was found, imported, withheld and still missing.

After each meaningful growth step, I replay the established questions. That is not a formality: adding correct pages once made a key answer harder to find. When a source is broken, I recover the original rather than swap in a lookalike.

What does it get wrong, and how do you find out?

Every complaint from a tester gets reproduced, traced to the layer that caused it, fixed at that layer, and checked against a nearby case that must not change. These are real problems found in this build.

More correct content buried a key answer

Loading the full set of pages pushed an important summary about conflicting instructions out of the top results for a real question. I kept the original test, then gave reviewed summaries that document a conflict a small, capped boost. Access checks still apply first.

Drafts asked for information already in the message

When someone pasted a screenshot as the answer, the conversion step dropped images and quoted replies before the AI ever saw them. Now images, files and quotes reach the model, and the assistant is told never to ask someone to resend what is already there.

Closed tasks became unreadable

The assistant could only read conversations tied to an open draft. Once a draft closed, it lost track of that history. I separated reading permission from open tasks, so past items stay readable while sending keeps its own checks.

A test passed for the wrong reason

One check looked good, but the testing session had already seen the answer elsewhere. Retrieval tests now use a fresh question, block every other source, and require the search trace, the returned source ID and an exact supporting quote.

A document in the export was cut off

One reference document arrived truncated from the wiki export, and downloading it again did not help. I found the complete original in the email where it was first sent, instead of substituting a similar guide.

One problem was outside my code: during testing, one web chat app generated sourced answers but displayed them blank. The same search worked through another supported app, so I reported it as a display issue in that app rather than calling the whole system fixed or broken.

What does it not do?

It does not change business systems. The knowledge connection only searches and reads.
It does not send a reply on its own. Every drafted message needs the employee's explicit approval of the exact text.
It does not pick a winner when two documents disagree. It shows both, with their dates, and says what is unresolved.
It does not have full transcripts of every training video. Many are covered by a reviewed summary card only, and that limit is labeled.
It does not include an outside training course that could not be retrieved. I left it out rather than fill the gap with generic vendor material.
It does not cover sensitive material that was held back on purpose.
Where a person still steps in: approving every outgoing reply, settling which of two conflicting rules is current, and supplying sources that are missing.

Frequently Asked Questions

What does a company brain cost?

It is priced per project. What moves the cost: how many sources you have and how messy they are, how much video and image material needs describing or transcribing, how detailed the access rules are, where staff will use it, whether it only answers or also drafts replies, and how often your procedures change.

Can it send messages or change records on its own?

Not in this build. Searching the knowledge is read only. Drafted replies are tied to the exact version the employee saw, and if a send fails in a way that is unclear, it stops and waits for a person instead of sending twice.

Does it learn from every edit my staff make?

No, and that is deliberate. An employee can tell it to remember a personal preference, list what it remembers and delete it. Ordinary edits do not quietly become permanent memory or company policy.

Is a bigger knowledge base always better?

No. A record count says nothing about accuracy. In this build, adding correct pages actually made one important answer harder to find until I fixed the ranking. That is why every growth step reruns the known questions.

Can it work inside the tools my team already uses?

That was the approach here. It drafts replies inside the company chat app staff already work in, and it connects privately to the AI chat tool they use, so they can ask questions there with their own sign-in.

Want an Assistant That Knows Your Company?

Book a call and tell me where your procedures live today. I will tell you what a company brain could cover and what it would need.

We use optional analytics and marketing cookies to understand visits and enquiries. Your demo works with either choice. Privacy policy