AI Knowledge Base Options for Small Businesses Compared

The right AI knowledge base for a small business is usually the one built into the office suite you already run: SharePoint with Copilot, or Google Drive with Gemini and Gemini Notebook. Add ChatGPT Business or Claude Team when answers must span several apps, and move to a wiki like Notion or Slite only when your know-how isn't written down yet.

The tool matters less than the documents behind it. Every option here answers from whatever files it can see, so a folder holding two cancellation policies, last year's price list and a scanned PDF nobody can search will produce confident wrong answers in any of them. Budget more hours for tidying than for choosing, and test with your own questions before you pay.

Follow me on Instagram@sagnikteaches

Five kinds of AI knowledge base, and the job each one does

"AI knowledge base" gets used for five quite different products. They overlap at the edges, but each starts from a different place, and picking the wrong category costs more than picking the wrong brand inside the right one. Prices below are list prices in USD for ten people on annual billing, checked in September 2026.

Connect on LinkedInSagnik Bhattacharya

If you only read one line of that table, read the last column alongside the "best when" column. The cheapest option is usually the one where your files already live, and the expensive mistakes come from buying a wiki to solve a tidying problem, or building something custom when a notebook would have done.

Subscribe on YouTube@codingliquids

Office-suite AI: answers from files you already store

This is where most small firms should start, because there's nothing to migrate and your existing sharing permissions carry over. The catch is the same thing: the AI sees whatever each person can open, including files that were shared carelessly years ago.

Microsoft 365: Copilot Chat, Copilot licences and SharePoint agents

Every Microsoft 365 business plan includes Copilot Chat at no extra cost. It now works on Outlook mail and on files you have open, which covers "summarise this procedure" but not "find the answer somewhere in our 400 documents". Searching across email, meetings and files together needs the paid Microsoft 365 Copilot Business licence at $21 per user a month on annual billing, with a promotional $18 on annual plans until 31 December 2026.

The piece small firms overlook is the SharePoint agent. You point one at a site or library, such as "Clinical protocols", and it answers only from that content. People with a Copilot licence use it as part of their licence; for everyone else it can run on pay-as-you-go billing through an Azure subscription, metered through Copilot Studio at a list rate of $0.01 per billable message. That makes it possible to give ten receptionists a scoped question box without buying ten licences.

A nine-person veterinary practice shows how this plays out. Its anaesthetic checklists, vaccination schedules and controlled-drug procedures already sit in one SharePoint library. The two practice managers take Copilot Business licences (about $42 a month at list, or $36 during the promotion) because they also want meeting and inbox help. Nurses and reception staff get a SharePoint agent scoped to the protocols library on pay-as-you-go. If staff ask around 300 questions a month and each answer uses several billable messages, the bill lands in the low tens of dollars; the practice sets an Azure budget alert so a surprise can't creep up. Before switching anything on, it spends an afternoon on cleaning up SharePoint permissions, because a payroll spreadsheet sitting in an "Everyone" folder would otherwise start turning up in answers.

Google Workspace: Gemini in Drive and Gemini Notebook

On Google Workspace, Gemini is built into the business plans rather than sold as an add-on. Business Starter (about $7 per user a month on an annual plan) gets Gemini in Gmail and the Gemini app; Business Standard (about $14) and above add Gemini across Docs, Sheets, Drive, Meet and the rest. Confirm the exact features on the Workspace pricing page for your account, because Google adjusts them by edition.

For a knowledge base, the more useful tool is Gemini Notebook, the product formerly called NotebookLM. You choose the sources, it answers only from them, and every answer links back to the passage it used, which makes checking easy. On a free personal account the limits are 100 notebooks, 50 sources per notebook and 50 chats a day; Google AI Pro raises that to 300 sources per notebook and 500 chats a day, and Workspace editions may get expanded access. Fifty sources is plenty for a staff handbook. It isn't enough for a whole shared drive, which is the point: you curate what goes in. There's a fuller walk-through in Gemini Notebook for small business teams.

ChatGPT Business and Claude Team as a knowledge layer

These make sense when the answer to a typical question is spread across several apps: part in a Slack thread, part in a Drive document, part in the CRM. Both cost $25 per seat a month on monthly billing or $20 on annual billing, with a two-seat minimum, and neither trains on business content by default.

ChatGPT Business has two relevant features. Projects hold a fixed set of files and instructions that a team can share, which suits a bounded topic like "onboarding pack". Company knowledge is the broader one: you pick it under the message box, and ChatGPT searches the apps people have connected, such as Slack, SharePoint, Google Drive, GitHub and HubSpot, then answers with citations. It respects each person's existing permissions, so it can only surface what that person could open themselves, and it only reads; it can't update records. Admins control which apps are available in the workspace settings.

Claude Team works mainly through Projects. You upload documents and write standing instructions, and on paid plans a project that outgrows Claude's working memory switches automatically to retrieval, searching the files instead of reading them all at once, which Anthropic says lets a project hold up to ten times more content. Connectors to Google Drive, Microsoft 365, Slack, Notion and others let Claude look things up live, again limited to what the signed-in person can access. Organisation owners switch connectors on for the whole team first.

One warning while you're here. If you were planning to build a "company brain" as a custom GPT, don't: OpenAI is retiring custom GPTs, which stop running on 11 December 2026. Projects are the safer home for shared context, and the difference between these approaches is covered in RAG vs custom GPT vs fine-tuning.

Wiki tools with AI built in: Notion, Slite, Confluence and Guru

A wiki earns its place when the problem isn't finding documents but the fact that nobody has written them. A wiki gives structure, templates and ownership; the AI layer then answers from pages people actually maintain.

Notion Business is $20 per member a month billed annually and includes Notion Agent (chat, drafting, autofill), AI meeting notes and Enterprise Search, which finds answers across your Notion workspace and connected tools such as Slack and GitHub. Free and Plus plans only get a limited trial of the AI features, and custom agents run on credits at $10 per 1,000. A 15-teacher language school that already keeps its teacher handbook, level descriptors and room-booking rules in Notion would get the most from this, because the pages exist and moving to Business simply adds the question box. Whether that's worth it for your team is the subject of is Notion AI worth it.

Slite is built specifically as a knowledge base rather than an all-purpose workspace. Basic is $10 per member a month on annual billing and allows 30 AI questions per seat each month; Pro is $20 and removes the question cap, searches connected tools (Slack, HubSpot, Google Drive, SharePoint and others) and adds fact-checking that suggests fixes to out-of-date docs. Its doc verification workflow, which asks an owner to confirm a page is still true, is the feature small teams end up valuing most. There's a 14-day trial with no card.

Confluence has a free plan for up to 10 users, and its paid plans bundle Atlassian's Rovo AI (search, chat and agents) with a monthly credit allowance per user. Per-user list prices start in the single digits of dollars, but check the credit allowance on Atlassian's page before assuming the AI is unlimited. It's the natural pick if your team already lives in Jira.

Guru now sells on tailored, sales-led pricing with an onboarding team rather than published per-seat prices. That suits a 60-person support operation more than a ten-person practice, so treat it as an option only if you're growing fast and want someone to design the knowledge structure with you.

When the knowledge base is for customers, not staff

A customer help centre is a different animal. The articles are public, written for outsiders, and the AI layer is a chatbot answering visitors. Help Scout has a free plan for up to five users and, on its paid plans, charges $0.75 per AI Answers resolution; Intercom's Fin charges $0.99 per outcome; Zendesk includes AI agents in its Suite plans and bills on successful automated resolutions. Read how each defines a resolution, because some count a customer going quiet as a success, and silence isn't satisfaction.

Keep internal material out. A dental practice's public help centre can explain whitening aftercare and payment plans; it must never be fed the staff rota, the complaint log or the internal note that says which fees are negotiable. If you need both, run two knowledge bases and keep the boundary strict.

Building your own retrieval system, and when it's justified

A custom build means a developer connects your documents to an AI model through its API, usually using retrieval (the model searches your files for relevant passages before answering). The model costs are often small; the build and the upkeep are not. It's justified when one of four things is true: your knowledge sits in formats off-the-shelf tools can't read, such as a practice-management database; you have strict rules about where data is stored or how long it's kept; you need answers inside your own software or website; or volumes are high enough that per-seat pricing becomes silly. Rough budgets are laid out in what a chat-with-your-documents system costs. For a team under 20 people with ordinary documents, one of the four off-the-shelf categories almost always wins.

Scoring the options on what actually matters

Feature lists all look alike. These seven criteria are the ones that decide whether staff trust the answers six months in.

Two rows deserve extra weight. Permissions, because an AI that can read everything a careless share exposed is a data problem waiting to happen. And "who keeps it current", because a knowledge base with no named owner starts drifting from the truth within a term.

A tutoring agency picks one, with the numbers

Consider an illustrative tutoring agency with 14 people: three office staff and eleven mostly part-time tutors. The office team answers the same questions every week. Can we charge for a session cancelled the night before? What's the safeguarding reporting route if a pupil discloses something? How do I log a session when the parent pays cash? What's the pay rate for a group session? Those answers live in 26 documents scattered across Google Drive, a few email threads and one office manager's memory. Everyone already has a Workspace account: the office staff on Business Standard, the tutors on Business Starter.

Here's how the realistic options compare for them at September 2026 list prices.

The agency starts with the notebook. The office manager spends a morning deleting three superseded versions of the cancellation policy, merging two pay-rate sheets and adding an "effective from" date to the top of every policy. She builds one notebook called "Tutor handbook" and shares it with the team. After six weeks the agency checks three numbers: how many tutor questions still reach the office (down from roughly 25 a week to about 9 in this illustration), how many of 20 spot-checked answers were right (19), and how long upkeep took (about 40 minutes a fortnight). The one wrong answer came from a safeguarding flowchart saved as an image, which the notebook couldn't read; she retyped it as text.

The agency would revisit the choice if tutors started needing answers that live in HubSpot or Slack, at which point ChatGPT Business for the three office seats ($60 a month on annual billing) with company knowledge becomes the next step. It would not buy seats for eleven tutors to ask five questions each.

Run a 20-question test before you pay for anything

Most free trials last 14 days. That's enough for this test, which takes about two hours per tool once the questions are ready.

1. Collect 20 real questions staff asked in the last month, copied from email, chat or memory. Include at least three about money, two about safety or compliance, and two whose answer isn't written down anywhere.
2. Write the correct answer and the document it comes from beside each one. This is your answer key.
3. Load the same documents into each tool you're testing, then ask every question with the same wording.
4. Score each answer: correct and cited; correct but uncited; wrong; or honestly "not covered". Note how long each answer took.
5. Set a pass mark in advance. A sensible one for a small firm: at least 18 of 20 correct, and zero confidently wrong answers on money or safeguarding.

Use a fixed instruction so the tools are compared fairly. This one works in Gemini Notebook, a Claude project or a ChatGPT project:

Answer using only the documents provided. After the answer, name the
document and section you used. If the documents don't cover the
question, reply "Not covered in our documents" and stop. Don't fill
gaps from general knowledge.

Question: A parent cancelled a Tuesday 4pm session at 9pm on Monday.
Do we charge for it, and how much?

An illustrative answer from a tool loaded with the agency's untidied folder:

Yes. Under the Cancellation Policy (section 3), sessions cancelled with less than 24 hours' notice are charged in full. The parent cancelled 19 hours before the session, so the full session fee applies. Source: Cancellation Policy 2024.docx, section 3.

It looks perfect, and it's wrong. The agency moved to a 48-hour rule with a 50% charge in January, but the 2024 file was still in the folder and happened to rank higher. The fix isn't a better prompt; it's deleting the old file and adding "Effective from 6 January 2026" to the new one. This is the most common failure in these tests, and it's why the source citation matters: without it, nobody would have spotted the stale file.

Where AI knowledge bases go wrong in small firms

These failures show up in every category above. Knowing them in advance saves weeks.

• Two versions of the truth. The cancellation-policy case above. Search your files for words like "old", "v2", "final" and "copy" before loading anything, and keep one current version with an effective date.
• Permissions nobody remembers setting. Office-suite AI and connected assistants show people what they can already open. A salary sheet shared with "anyone in the organisation" in 2022 becomes an answer in 2026. Audit sharing on sensitive folders before launch.
• Images and scans. Flowcharts saved as pictures, scanned forms and photographed notices are often unreadable or half-read. A language school found its placement-test marking guide returned "not covered" every time because it was a scan; retyping two pages fixed it.
• General knowledge filling the gaps. Asked about a policy you haven't written, a chat assistant may answer from what's typical elsewhere. The fixed instruction above reduces this; spot checks catch the rest.
• No owner, no review date. Policies change at term starts, price rises and staff changes. Without an owner, the knowledge base quietly drifts. Put a name and a "review by" month at the top of every document.
• Credit and message meters. Per-message agents and credit-based wiki AI are cheap at low volume and easy to forget. Set a budget alert the day you switch them on.

A 30-minute monthly upkeep routine

Whichever option you pick, this routine keeps answers trustworthy. It fits in half an hour once a month for a team under 20.

1. Ask the three most common questions from last month and check the answers against the current documents.
2. Look at anything answered "not covered". Each one is either a document you need to write or a question that shouldn't be answered by the AI.
3. Check the review dates at the top of documents and chase any that have passed.
4. Remove superseded files from the sources, not just from view.
5. Glance at the usage or billing page for anything metered.

If you'd rather build the document side first, how to build a company knowledge base AI can answer from covers writing and structuring the content itself, which is the half of the job no tool does for you.

Questions owners ask before choosing an AI knowledge base

Can we run two of these options side by side?

Yes, and many small firms do. A common pairing is office-suite AI for everyday document questions plus a public help centre for customers. What causes trouble is two internal knowledge bases holding different copies of the same policy. Pick one home for each document, link to it from anywhere else, and make sure only that copy is fed to the AI.

Will the AI provider train its models on our documents?

On business plans, no by default: ChatGPT Business, Claude Team, Microsoft 365 Copilot and Gemini in Workspace don't train on business content unless you opt in. Consumer accounts differ, and staff using personal logins can undo your careful choice. Check each wiki vendor's AI terms too, because they pass your text to model providers under their own agreements.

How many documents do we need before this is worth doing?

Fewer than people expect. If staff ask the same twenty questions every month and the answers already live in ten to thirty documents, a knowledge base pays off quickly. The trigger isn't volume of documents but volume of repeated questions landing on one or two people who know the answers.

What happens to our knowledge base if the vendor shuts down?

Keep the source documents in a format you own, such as Word, Google Docs or Markdown exports, and treat the AI layer as replaceable. Office-suite options carry the least risk because the files stay where they are. Wikis should offer a full export; test it during the trial rather than assuming it works.

Further reads

• How to Set Up Claude Projects as a Shared Team Assistant — Step-by-step setup if a Claude project becomes your team's knowledge layer.
• Turn Fixed Tickets Into a Searchable AI Knowledge Base — Turn solved support tickets into answers your knowledge base can reuse.
• Copilot Chat vs Microsoft 365 Copilot: What Do You Get Free? — What the free Copilot Chat covers before you buy licences.
• How to Catch Outdated Information in AI Answers — Spot stale answers before a customer or colleague acts on them.
• AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep your documents and prompts portable if you switch tools.
• How to Build the FAQ Your AI Chatbot Needs Before Launch — Write the FAQ a customer-facing knowledge base needs first.
• Reduce Owner Dependency: Use AI to Capture What Only You Know — Get what's in your head into a form staff can use: an interruption log, an AI interview prompt, decision rules with thresholds, and a shared assistant.
• Can AI Answer Student Questions for Your Online Course? — Sort your real student questions first, then choose between drafting replies yourself, a built-in course assistant or a separate chatbot.
• How Much Does a Custom AI Assistant Cost a Small Business? — Three ways to get an AI assistant that knows your business, what each costs up front and every month, and the upkeep that most quotes leave out.
• How to Stop an AI Chatbot Giving Customers Wrong Answers — The five causes behind almost every wrong chatbot answer, the fix for each, and a yoga studio's month-one error log with the numbers before and after.
• Slack AI for Small Teams: Catch-Ups, Search and Costs — What Slack's AI does on Pro versus Business+, what the upgrade costs a nine-person team, and how to test whether catch-ups and search save real time.
• What Is Copilot Studio and Does a Small Business Need It? — What Microsoft's agent-building tool actually does, how its credit billing adds up, and the four situations where a small firm genuinely needs it.
• What Is RAG? A Plain-English Explainer for Business Owners — RAG explained without jargon: how AI looks up your own documents before it answers, traced through a real-looking policy question.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn (Copilot pay-as-you-go meters, SharePoint agents); Google Gemini Notebook Help (plan limits); OpenAI help pages (company knowledge, admin controls for apps); Claude help pages (RAG for projects, connectors); Notion, Slite, Guru, Help Scout and Intercom pricing pages; Atlassian Confluence plan information. Checked September 2026.