How to build an AI assistant for your team on your own data

The gap between general AI and useful AI

General-purpose AI tools are impressive until you ask them something specific. "What did the client say about the revised timeline?" "Where's the onboarding checklist for the design team?" "What was the conclusion from last quarter's post-mortem?" These are the questions that eat hours every week, and a chatbot trained on the open web cannot answer any of them.

The difference between a general AI and a useful one is context. A useful AI assistant knows your projects, your decisions, your terminology. It can pull from the Slack thread where the decision was made, the Google Doc where the scope was defined, the Notion page where the process was documented. Building that kind of assistant has traditionally required a significant engineering effort. Fabric makes it possible without writing a line of code.

What Fabric provides is a persistent memory and context layer, a vault that sits between your team's tools and whatever AI model you choose. Rather than each model seeing only what you paste into a chat window, Fabric maintains a durable, searchable understanding of your organisation's knowledge that any model can draw on across every session.

What "on your own data" means

When people talk about building an AI assistant on their own data, they are usually describing a technique called retrieval-augmented generation. Rather than fine-tuning a model on your documents (which is expensive, slow, and difficult to keep current), RAG works by searching your data at query time and passing the relevant passages to the language model as context. The model then generates an answer grounded in your actual files, with citations pointing back to the source.

This approach requires three things: a way to ingest and index your data semantically, a retrieval layer that can find the right passages for a given question, and an interface that presents the answer with its sources. Building each of these from scratch involves managing vector databases, chunking strategies, embedding models, and an orchestration layer to tie them together. For most teams, this is a project that competes with the product roadmap for engineering time.

Fabric's AI assistant handles all three layers. You connect your tools, Fabric indexes the content, and the assistant answers questions with citations drawn from your own knowledge base. The underlying infrastructure is managed for you, so there is no pipeline to build or maintain.

Why building it yourself is more expensive than it looks

The initial prototype of a RAG pipeline is deceptively simple. A weekend project can produce a chatbot that retrieves passages from a small corpus and generates reasonable answers. The gap between that prototype and a production system your team can rely on is where the cost lives.

Keeping the index fresh as documents change, handling permissions so people only see what they should, supporting dozens of file types and data sources, managing chunking so that context windows are used efficiently, dealing with latency at scale: each of these is a sustained engineering problem. And once the system is running, someone has to monitor retrieval quality, tune relevance, and update the embedding model as better ones become available.

For teams with dedicated ML infrastructure engineers, this can be a worthwhile investment. For everyone else, it means pulling people away from the work they were hired to do. The total cost of ownership, including the ongoing maintenance, tends to be far higher than the initial build suggests.

How Fabric handles it

Fabric provides a knowledge base that connects to over fifty apps and data sources, from Google Drive and Notion to Slack, Confluence, Figma, and GitHub. Once connected, Fabric indexes the content using its semantic search engine, which understands meaning rather than just matching keywords.

When someone asks the assistant a question, it retrieves the relevant passages from across all connected sources, passes them to the language model, and returns an answer with inline citations. You can click through to the original document to verify or read further. The assistant supports multiple models, including Claude, Gemini, OpenAI, DeepSeek, and Grok, so your team can choose the one that works best for the task at hand.

Fabric also lets you customise the context the assistant draws from. You can scope it to particular folders, tags, or sources, which is useful when different teams need different knowledge boundaries. A support team might want the assistant focused on product documentation and resolved tickets, while a strategy team might want it drawing from market research and board decks.

Giving the whole team access with Fabric Tag

Building an AI assistant is only half the problem. The other half is making sure people use it. If the assistant lives in a separate app that people have to switch to, adoption tends to drop off quickly. The questions that matter most come up in the middle of a conversation, not during a dedicated "go ask the AI" session.

Fabric Tag solves this by bringing the assistant into the tools your team already uses. In Slack or Discord, anyone can tag the assistant in a channel and get an answer grounded in your connected data, right in the flow of conversation. The response includes citations, so the rest of the channel can see where the answer came from.

This changes the dynamic from "someone should look that up" to an instant, sourced answer. It is particularly effective for recurring questions: the ones that experienced team members answer from memory dozens of times a month.

Automating knowledge work with agents

Beyond answering questions on demand, an AI assistant becomes more valuable when it can act on its own. Fabric's agents run scheduled or triggered tasks across your connected data. An agent might monitor a Slack channel for unanswered questions and draft responses from the knowledge base, generate a weekly summary of new documents added to a project folder, or flag inconsistencies between a product spec and its associated support documentation.

Agents operate within the same permission and context boundaries as the rest of Fabric, so they only access what they are authorised to see. Fabric's access model is deliberately scoped: you approve each agent (whether Instinct, Muse, or Claude) for limited-time access to specific tools, such as reading a particular Slack channel or checking a shared inbox, rather than granting blanket permission across everything. They extend the assistant from a reactive tool into something closer to a team member that handles the repetitive knowledge tasks no one wants to do manually.

Capturing knowledge passively with self-writing docs

One of the persistent challenges with any knowledge system is keeping it current. Documentation drifts out of date because updating it is a manual task that rarely feels urgent. Fabric's self-writing docs address this by automatically generating and updating documentation based on the data flowing through your connected sources.

As new decisions are made in Slack, as specs are updated in Notion, as meeting recordings are transcribed, self-writing docs can synthesise these into living documents that stay current without anyone having to maintain them. This means the assistant's knowledge base improves over time, even when no one is actively curating it.

Security and data handling

Connecting your internal data to an AI system raises reasonable questions about security. Fabric uses AES-256 encryption for data at rest and has achieved CASA Tier 2 certification. Your data is used to answer your team's questions, not to train models. The security architecture is designed so that connecting your tools does not mean giving up control over your information.

Permissions from your source applications are respected, so the assistant will not surface a document to someone who would not have access to it in the original tool. This is a meaningful distinction from approaches that dump everything into a single vector store without access controls.

Getting started

The path from "we want an AI assistant" to "the team is using it" is shorter than most people expect. Connect two or three of your most-used tools, let Fabric index the content, and start asking questions. You do not need to migrate data, restructure your file hierarchy, or define a taxonomy. The semantic search and memory engine handle the organisation. Fabric's cloud drive streams files on demand with zero local storage, so your team can browse and work with the full library without downloading anything to their devices.

Because Fabric is the context layer rather than the model itself, your team is not locked into a single AI provider. The assistant works through Fabric's own interface, and your Fabric content is also accessible from ChatGPT or Claude via MCP. A team member who prefers ChatGPT can query the same indexed knowledge base from within ChatGPT's own interface, pulling answers grounded in your connected tools without ever opening Fabric directly. The knowledge base stays the same regardless of which model or interface someone prefers. And through Fabric's agent access model, you can approve specific agents (Instinct, Muse, or Claude) for scoped, time-limited access to particular tools, such as reading a shared inbox or monitoring a project channel, rather than granting permanent access across everything.

From there, you can expand: add more connections, set up Fabric Tag in your team channels, configure agents for recurring tasks, and enable self-writing docs for areas where documentation tends to lag. Each step adds value without requiring the previous one to be perfect.

Frequently asked questions

Do I need an engineering team to set this up?

No. Fabric is designed so that anyone can connect their tools and start using the assistant. There are no embeddings to manage, no vector databases to host, and no retrieval pipelines to build. The setup involves authorising your apps through Fabric's connections page, which takes minutes rather than sprints.

What data sources can I connect?

Fabric supports over fifty connections, including Google Drive, Slack, Notion, Confluence, GitHub, Figma, Linear, Jira, Dropbox, OneDrive, and many more. You can also upload files directly or use MCP to connect custom sources. The full list is available on the connections page.

How does the assistant know which information is relevant?

Fabric uses semantic indexing to understand the meaning of your documents, not just the words they contain. When you ask a question, the retrieval layer finds passages that are conceptually related to your query, even if they use different terminology. This is the same principle behind retrieval-augmented generation, handled automatically.

Can different teams have different knowledge boundaries?

Yes. You can scope the assistant's context to particular sources, folders, or tags. This means a customer support team can have the assistant draw from product docs and ticket history, while a sales team can have it focused on proposals, case studies, and CRM data. The same assistant infrastructure serves both, with different context configurations.

How does Fabric Tag work in Slack?

Fabric Tag is a bot you add to your Slack workspace. When someone mentions it in a channel or thread, it reads the question, searches your connected knowledge base, and replies with a cited answer. Everyone in the channel can see the response and the sources, which makes it a shared learning moment rather than a private lookup.

Is my data used to train AI models?

No. Your data is used to generate answers for your team and is not used for model training. Fabric's security model keeps your information within your workspace boundary.

What AI models are available?

Fabric provides access to models from multiple providers under a single subscription, including Claude (Anthropic), Gemini (Google), GPT (OpenAI), DeepSeek, and Grok (xAI). You can choose the model that best fits a given task, and switch between them without separate billing or API keys.

How does this compare to building a custom RAG pipeline?

A custom pipeline gives you full control but requires significant engineering investment: choosing embedding models, managing a vector database, building an ingestion layer for each data source, handling permissions, and maintaining the system over time. Fabric provides the same end result, a grounded AI assistant over your own data, without the build or maintenance burden. For teams without dedicated ML infrastructure, it is typically the faster and more cost-effective path.

Can agents run tasks automatically?

Yes. Agents can be configured to run on a schedule or in response to triggers. Examples include generating weekly summaries, monitoring channels for questions, and keeping documentation up to date. They operate within the same permission boundaries as the rest of the system.

How long does it take to see results?

Most teams have the assistant answering questions within an hour of connecting their first data sources. The indexing runs in the background, and the assistant becomes more useful as more content is processed. There is no minimum dataset size or setup period required.

Related pages

• How to create a shared knowledge base

• How AI agents access your data

• What is retrieval-augmented generation

• How to use AI without losing your data

• How to give an AI agent access to your knowledge base

• How to stop your team asking the same questions

• How to onboard a new hire without writing a wiki