AI knowledge base
Most tools sold under this name answer questions from pages someone has already written, and most are built for customer support. Knowing which of the two jobs you need saves you buying the wrong one. SecondBrain 365 does the other job: an AI agent inside your Microsoft 365 that writes the pages from email, meetings and chats.
An AI knowledge base is a knowledge base in which artificial intelligence does part of the work. Two very different jobs go by that name.
The first is answering: an assistant finds the relevant passages in pages that already exist and replies in plain language, usually citing the page. This is called retrieval, and it is what nearly every guide and product under this name describes. The second is writing: the AI drafts new pages, updates old ones and notices when a new fact contradicts something already written.
The two depend on each other. An assistant can only be as right as the pages it reads, so the question worth asking any vendor is not how well it answers, but who writes the pages it answers from — and what happens when two of them disagree.
Two questions sort every product: does the AI answer or write, and are the readers your customers or your colleagues? Search results for this term are mostly about the top-left box.
The AI answers · read by customers
A chat or search box that answers customers from the articles in your help centre. Most of what is sold as an AI knowledge base is this.
What it needs: Articles updated before every product or policy change reaches customers, and a handover to a person.
The AI answers · read by colleagues
Answers employees from what is already written in wikis, shared drives, sites and chats — within what each person is allowed to see.
What it needs: Pages that exist, that are current, and that do not contradict each other.
The AI writes · read by customers
Drafts help articles from the way agents actually solved tickets, or flags the articles that stopped helping. A person approves before anything is published.
What it needs: Enough resolved tickets to learn from, and someone who reviews the drafts.
The AI writes · read by colleagues
Writes and updates pages from the email, meetings and chats where decisions are made, and checks each new fact against what is already written.
What it needs: To be brought into those conversations. It cannot write down what it never saw.
Where SecondBrain 365 works
The boxes are not rivals: the writing side feeds the answering side. If your readers are colleagues, start from what an internal knowledge base is for; if the pages will live on SharePoint, see an AI author for your SharePoint site, where it is compared with Copilot in SharePoint.
Nothing in the answering changes. What changes is which page it happens to read.
Pages as they are
Event plan
Written after the July planning meeting
Registration closes on 30 September.
Comms sync — recap
Written after the 9 September meeting
Registration extended to 10 October; the venue has confirmed the extra seats.
Marco · Sales
When does registration close for the partner event?
AI assistant
Registration closes on 30 September.
Source: Event plan
Giulia · Marketing
Can customers still register for the partner event after 30 September?
AI assistant
Yes. Registration was extended to 10 October.
Source: Comms sync — recap
Same question, two answers. Both cite a real page.
Pages written and checked
Partner event — key dates
Updated from the 9 September meeting · change confirmed by the team
Registration closes on 10 October. It was 30 September until the team moved it on 9 September.
Marco · Sales
When does registration close for the partner event?
AI assistant
Registration closes on 10 October. It moved from 30 September on 9 September.
Source: Partner event — key dates
Giulia · Marketing
Can customers still register for the partner event after 30 September?
AI assistant
Yes, until 10 October.
Source: Partner event — key dates
Same assistant, same questions. One page, one answer.
An answering AI repeats what it retrieves. That is the point of it — it is meant to stick to your pages instead of making things up — and it is also its limit. In the ClashEval benchmark (NeurIPS 2024), six leading models were given documents with a deliberately wrong answer planted in them. They adopted the wrong content, over their own correct knowledge, more than 60% of the time.
Those were general-knowledge questions, where a model at least has something of its own to go on. For your company's dates, owners and decisions it has nothing: the page is the only source it has.
When two retrieved pages conflict, a Google Research study of conflicting sources in search-augmented models found that models often struggle to resolve the conflict appropriately; prompting them to reason about it explicitly helps, without closing the gap.
Even a model that handles it perfectly can only tell you that two pages disagree. Which one holds was decided somewhere the assistant cannot read.
Start from what is going wrong today, not from the feature list. Many teams need both, in this order: pages first, answers second.
People cannot find pages that exist
An AI that answers
Search is the bottleneck. A good assistant over good pages fixes it, and this is where most tools are strongest.
Customers ask the same questions every day
A help-centre assistant
And someone who owns the articles it reads. The assistant repeats a stale article as confidently as a fresh one.
The pages were never written, or are months behind
An AI that writes
An assistant over missing pages either says it found nothing, or answers from the nearest old page. Adding one does not create the content.
Two pages say different things
A check when the page is written
At answer time, the best an assistant can do is report the disagreement. Which version holds was settled in a meeting or an email, and that is where it has to be caught.
The writing side, on Microsoft 365
SecondBrain 365 is an agent with its own identity and company email inside your Microsoft 365. You copy it on an email, invite it to a Teams meeting, add it to a Teams group or share a document with it. From what it is shown, it writes and updates the pages of a knowledge base on a SharePoint site in your tenant, following rules you write in plain language.
Before a new fact goes in, it checks it against what is already written. When the two disagree, it asks the team on Teams which version holds — the step missing from the first half of the widget above. You can also require approval before anything is written.
The answering side stays where it is. The agent answers questions from the knowledge base it keeps, and because the pages are ordinary SharePoint pages, Microsoft 365 Copilot can answer from them too.
See the full walkthrough, from the first email in copy, a support team's version in a service desk knowledge base that grows from answered emails, and what became of the topic pages Viva Topics compiled with AI.
What it does not do
FAQ
A knowledge base in which artificial intelligence does part of the work. Usually that work is answering: an assistant retrieves the relevant passages from existing pages and replies in plain language. Some tools also write: they draft or update pages and check new facts against what is already there. Either kind can serve customers or employees.
From material it is given, yes: email threads, meeting transcripts, chats and documents can become structured pages. What it cannot do is write down knowledge nobody ever expressed. A decision that exists only in someone's head still has to be said or written somewhere first.
It reduces invention, because the AI answers from your pages instead of from general knowledge. It does not make the answer right: the AI repeats what the page says. In the ClashEval benchmark, published at NeurIPS 2024, leading models adopted a deliberately wrong answer planted in the retrieved text more than 60% of the time, over their own correct knowledge. The accuracy of the pages is the accuracy of the answers.
The readers, and where the knowledge comes from. A customer-facing one answers from help articles written for the public by a support team. An internal one answers colleagues about projects, policies and decisions, and most of that knowledge starts life in email, meetings and chats, not in articles — which is why writing it down is the harder half.
Copilot is an answering AI: it answers from the content in your Microsoft 365 tenant that each person is allowed to see. It does not decide which of two conflicting pages is right, and someone still has to write the pages it reads. SecondBrain 365 writes them into SharePoint, where Copilot can answer from them.
Not in SecondBrain 365. When something new contradicts a page, the agent alerts the workspace members or its owner on Teams and they decide what goes in. If you enable approval, people also validate each change before it is written.
