Knowledge & teams

Internal knowledge assistant: stop your best people being the company search engine

By Babar Lal  ·   ·  9 min read

It is Friday afternoon. A new hire walks over to the office manager's desk. "Sorry, how do I log a warranty job again?"

It is the third time this week. The answer is in a document somewhere, but nobody knows which one. The office manager explains it again, then needs a few minutes to find her place in the payroll she was checking. By the end of the week, that adds up to another hour.

The same question two ways. On the left, staff ask the office manager how to log a warranty job on Monday, a new hire asks again on Wednesday, it is the third time by Friday, and another hour is lost. On the right, staff ask the assistant, which answers from the warranty process document and names the source, and flags a gap for the owner when no source exists.

The short answer

Give staff a chat tool that answers from your own documents and shows its source. Your senior people then answer only the questions nobody has written down yet.

QuestionShort answer
What an internal knowledge assistant isA chat tool that searches your own documents, answers from them, and names the document it used
What it is notA general chatbot that guesses, or a fix for documents nobody has written
Who it suitsClinics, trades, agencies, law and accounting offices and ecommerce teams where senior staff answer the same questions every week
Cost rangeAbout $18 to $25 a user a month for the built-in tools; a custom build has a setup fee and a monthly fee
What to do firstList the ten questions your senior staff answer most, and find where each answer is written

How much time goes on searching and asking

The most quoted study is old. McKinsey's article Capturing business value with social technologies was published in November 2012. That makes it 14 years old. I opened the original on McKinsey's own site. It found interaction workers spent 19% of an average workweek searching for and gathering information.

A newer figure comes from a vendor. Atlassian, which sells the Confluence wiki, surveyed 12,000 knowledge workers and 200 executives for its State of Teams 2025 report. It says teams "waste 25% of their time just searching for answers".

Neither study measures what I care about most here. That is the time lost by the person who gets asked. A question that takes two minutes to answer can cost far more in lost focus. I could not find an open study that measures this for small businesses, so the cost sum below uses my own assumption.

What RAG means in plain English

RAG stands for retrieval-augmented generation. The name is worse than the idea.

The assistant searches your own documents first. It finds the few passages that match the question. It writes the answer from those passages. And it shows which document it used, so staff can check.

A general chatbot works differently. It answers from what it learned in training. It has never seen your warranty process, your price list or your leave policy. Ask it anyway and it may write something that sounds right and is not. For internal questions, an answer with a source beats a confident guess.

How it works, step by step

  1. Gather the documents. Process guides, price lists, policies, templates, past job notes and help desk answers.
  2. Clean out old versions. If three copies of the price list exist, delete two. The assistant cannot tell which one is current.
  3. Connect the sources. That might be a shared drive, a wiki, a help desk or a folder of PDFs. New and changed files should flow in automatically.
  4. Staff ask where they already work. That might be Slack, Microsoft Teams or a simple web page.
  5. The answer cites its source. Every answer links to the document and section it came from.
  6. Gaps get flagged. When there is no good source, the assistant says so and logs the question. Someone then writes the missing document.
Six steps showing how the assistant answers from your documents: staff ask in Slack, Teams or a web page, it searches your own documents first, checks that the person may open each file, writes the answer from the passages it found, links the document it used, and when there is no good source it says so and logs the gap. If no source is found, it says it does not know.

That last step is the one owners forget. The questions the assistant cannot answer are a free list of the documents your business is missing.

This is the same idea as an AI website chat that answers customers from your own information, pointed inward at staff. It is a different job from pulling numbers out of invoices and receipts.

Access, privacy and accuracy

This is general guidance, not legal advice.

Who can see what. The assistant should only search files the person asking could already open. Microsoft's own guidance says Copilot "only surfaces organizational data to which individual users have at least view permissions". That sounds safe. But it also means any folder shared too widely, years ago, becomes easy to find. Microsoft's SharePoint readiness guide spends most of its length on finding overshared sites for exactly this reason.

Keep HR, salary and client-confidential files out at first. Add them only once permissions are enforced and tested with a real staff account.

What the provider does with your data. On business plans, the big providers say they do not train on it by default. Microsoft says prompts, responses and data accessed through Microsoft Graph "aren't used to train foundation LLMs". Google says Workspace content is not "used for Generative AI model training outside your domain without permission". OpenAI says it does not use ChatGPT Business data for training by default. Anthropic says it will not use chats from its commercial products for training, unless the customer opts in. Personal accounts have different terms, so keep staff on the business plan.

Saying "I don't know". A good assistant refuses to answer when it has no source. Test this on purpose. Ask it something your documents do not cover and see what it does.

Clinics need more care. Patient information must never go into a tool that is not set up for it. Our post on reducing no-shows covers the HIPAA basics, and our healthcare automation page shows what we build for clinics.

Two columns. Goes in: process guides and checklists, price lists and service menus, policies everyone may read, templates and past job notes, help desk answers, and product and supplier specs. Stays out unless permissions hold: salaries and HR files, client-confidential files, patient records, contracts under negotiation, passwords and bank details, and drafts and old versions. The rule: if a person could not open the file, the assistant must not quote it to them.

Your options compared

Prices come from each provider's own page, checked in October 2026. Google's pricing page loads its prices dynamically, so I could not read them.

OptionTypical costGood atTrade-offs
Microsoft 365 CopilotFrom $18 a user a month paid yearly, shown as a reduced price; $25.20 paid monthly; on top of a Microsoft 365 planAnswers from SharePoint, OneDrive, Teams and Outlook with each user's permissionsOnly as safe as your SharePoint sharing settings
Google Workspace with GeminiIncluded in Business plans; prices did not load for meGemini in Docs, Drive and Gmail on Business Standard and aboveBusiness Starter has only limited Gemini in the apps
Notion AIBusiness $20 a member a month, with yearly billing up to 20% cheaperAnswers across a Notion wiki, plus search across Slack and other connected toolsWorks best when your documents already live in Notion
ChatGPT Business$20 a user a month paid yearly, $25 monthly, 2 users minimumFile uploads and connectors to Microsoft 365, Google Drive and SlackSomeone has to keep the files current
Claude Team$20 a standard seat a month paid yearly, $25 monthlyProjects with your files, connectors and search across your organizationSame as above: old files mean old answers
Custom RAG buildSetup fee plus monthly fee; our builds start at $200Several sources in one place, answers in Slack or Teams, gap logging, your own access rulesOnly worth it when the built-in tools cannot reach your sources

What owners and the industry are saying

I only list pages I could open. Vendors are labelled.

WhereWhoWhat they saidWhat it tells you
Microsoft Q&A, May 2026A Microsoft 365 customerAsked what "foundation LLM" means in the privacy policy: "Is data being used to train alternative/secondary LLMs?"Owners want data terms in plain English
Microsoft Learn, updated Jul 2026Software vendor"Copilot and agents retrieve data from Microsoft Graph and respect existing permissions, sharing settings, and policies."Your sharing settings become the security model
Atlassian, State of Teams 2025Software vendorTeams "waste 25% of their time just searching for answers"A vendor survey, but the direction matches older research
McKinsey, Nov 2012Consulting firm19% of the workweek goes on searching for and gathering informationThe problem is old; the tools are new

A cost sum you can redo with your own numbers

Here is the sum for a team of 12. Two of the numbers are my assumptions. Count your own for a week.

LineHow to get itExample
Staff who ask questionsHead count, minus the people being asked12
Questions per person per weekAsk senior staff to tally them for one week5 (assumed)
Minutes lost per interruptionThe answer plus getting back into the task6 (assumed)
Senior time per week12 x 5 x 6 minutes6 hours
Hourly cost of senior staffWage plus overheads$45
Cost per month6 hours x $45 x 4.3 weeksAbout $1,160
What repeat questions cost on example numbers. Time lost: 12 staff times 5 questions each a week times 6 minutes lost per interruption is 6 hours a week. In money: 6 hours a week of senior time times 45 dollars an hour times 4.3 weeks is about 1,160 dollars a month. If the assistant answers half of them, about 580 dollars a month comes back. The 5 questions, the 6 minutes and the half answered are marked as assumed.

An assistant will not answer every question. If it handles half, about $580 a month of senior time comes back. That is my assumption too. The bigger gain is often the new hire who gets an answer at 7am without waiting for anyone.

When it backfires

Outdated documents. The assistant quotes last year's price list with total confidence. It is only as current as your files. Give every key document an owner and a review date.

Permission leaks. Someone asks about bonuses and gets a line from a salary file that was shared with "everyone" in 2019. Audit sharing before you switch it on.

Nobody maintaining the source documents. The assistant works well for three months. Then processes change and nobody updates the guides. Answers drift until staff stop asking.

Staff losing trust after one bad answer. People forgive a colleague who gets it wrong. They rarely forgive a tool. Start with the documents you know are right, and make it easy to report a wrong answer.

What we build, and when we say no

If your documents are scattered or out of date, fixing the documents comes first. An assistant on top of a mess gives you fast, confident, wrong answers. We will tell you that before we quote anything.

If you already pay for Microsoft 365 or Google Workspace, try the built-in AI first. Clean up sharing, point it at one well-kept folder, and see what staff ask. You may not need us.

When the built-in tools are not enough, we build a RAG assistant. It pulls from several sources, such as a shared drive, a wiki and a help desk. It checks access per user, answers in Slack, Teams or a web page, and links every answer to its source. Questions it cannot answer go to a list for whoever owns the documents. Our RAG knowledge base service covers the assistant. Our workflow automation service keeps the sources in sync. Our AI automation service covers the rest of the AI work around it. Agencies with years of client work in shared drives are a common fit, and our marketing agency automation page shows what that looks like.

We say no when the documents do not exist yet, when the team is small enough to ask each other, or when nobody will own the source files. New to all this? Our plain-English answers to 24 questions about AI automation is the place to start.

Questions people ask

What is an internal AI knowledge assistant? It is a chat tool for your staff that answers questions from your own documents: process guides, price lists, policies and past job notes. It shows which document each answer came from. When it cannot find a source, it says so instead of guessing.

What does RAG mean? RAG stands for retrieval-augmented generation. In plain English, the assistant searches your documents first, then writes its answer from the passages it found, and links to them. A general chatbot answers from what it learned in training, which knows nothing about your business.

Can we just upload our documents to ChatGPT or Claude? For a small team with a handful of documents, yes, that can be a good start. Use a business plan, not a personal account. It gets harder when you have hundreds of files, documents that change every month, or files that only some people may see.

Will the AI provider train its models on our documents? On business plans, Microsoft, Google, OpenAI and Anthropic all say on their own policy pages that they do not train their models on your business data by default. Read the current page for the exact plan you buy, because the terms differ between personal and business plans.

How do we stop staff seeing files they should not? The assistant should only search files the person asking could already open. Fix sharing settings before you switch it on, because loose permissions that nobody noticed become easy to find. Keep salary, HR and client-confidential files out until permissions are tested.

What happens when the assistant gives a wrong answer? Usually the source document is wrong or out of date. Because every answer names its source, staff can check it and report it. Fix the document, not just the answer. Then the next person gets the right answer too.

Find out what repeat questions really cost.

Bring a week's tally of the questions your senior staff answer to the free audit. We check where the answers are written, run the sum with you, and tell you whether your Microsoft or Google AI, a team plan or a build is the right step. If the tools you already pay for can do it, we say so.

Book the free audit →

Sources