How to Build a Custom AI Assistant for Your Team in Bubble.io
A generic AI chatbot answers generic questions. A custom AI assistant trained on your company’s processes, playbooks, and knowledge answers the specific questions your team actually asks — accurately, instantly, and without bothering a senior colleague every time. This guide shows you how to build one.
vs Generic ChatGPT
| Feature | Generic ChatGPT | Custom Bubble.io AI Assistant |
|---|---|---|
| Knowledge | General world knowledge | Your company SOPs, playbooks, policies, and FAQs |
| Tone | Generic helpful AI | Your company voice and communication style |
| Context | No memory of your business | Knows your products, clients, processes |
| Privacy | Conversations processed externally | Controlled — stays in your Bubble database |
| Access control | Anyone with an account | Your team members only, with role-based access |
| Logging | Not visible to you | Every question and answer logged for review |
| Customisation | Prompt-level only | Full UI, workflows, and knowledge base customisation |
Full Build Guide
Create and organise your knowledge base
Before building anything in Bubble, gather and organise the knowledge the assistant needs. Create a Google Doc or Notion page with these sections: Company Overview (what you do, who you serve, your values), Products and Services (descriptions, pricing, positioning), Processes and SOPs (step-by-step instructions for the most common tasks), FAQs (the 20 questions your team asks most frequently), Client and Account Information (general account policies — not individual client data), and Communication Guidelines (tone of voice, escalation procedures). This document becomes the AI’s knowledge base. Aim for 1,000 to 3,000 words — comprehensive enough to answer common questions, focused enough to be relevant.
Build the Bubble.io database structure
Create these data types in Bubble. KnowledgeArticle: title (text), category (text), content (text), last_updated (date). AssistantConversation: user (link to User), created_date (date), title (text — auto-generated summary of the conversation). AssistantMessage: conversation (link to AssistantConversation), role (text: ‘user’ or ‘assistant’), content (text), created_date (date). Populate the KnowledgeArticle table by pasting your knowledge base sections as individual articles — one article per section or topic area.
Build the knowledge retrieval system
When a team member asks a question, the assistant needs to find the most relevant knowledge articles before generating a response. Build a Bubble backend workflow: receive the user’s question, search the KnowledgeArticle database for articles containing relevant keywords (use Bubble’s ‘contains keyword’ search across title and content fields), retrieve the top 3 to 5 matching articles, concatenate their content into a context string. This retrieved context is passed to Claude alongside the question — ensuring the AI answers from your company knowledge rather than general training data.
Build the AI response workflow
Create a Bubble backend workflow: API endpoint that receives the user’s question and conversation history. Retrieve the relevant knowledge articles (Step 3). Build the Claude API call: system prompt = You are [Company Name]’s internal AI assistant. Answer questions based only on the company knowledge provided. If the question is not covered by the knowledge base, say so clearly and suggest who to ask. Be concise and specific. Knowledge base: [retrieved articles]. Messages array = all previous messages in this conversation plus the new user message. Store the response as a new AssistantMessage record. Return the response to the UI.
Build the chat UI and deploy
In Bubble’s design editor, build the assistant interface: a sidebar or full-page chat UI with the conversation history displayed as a Repeating Group (same structure as the chatbot in Post 201). Add user authentication — the assistant is accessible only to logged-in team members. Add a conversation history panel showing the team member’s past conversations (useful for picking up where you left off). Add an admin panel (accessible only to admin users) showing all conversations across the team — useful for identifying gaps in the knowledge base based on questions the assistant could not answer well. Deploy and invite your team.
The Ongoing Process
Review the conversation logs weekly. Look for: questions the assistant answered incorrectly (update the relevant knowledge article), questions it could not answer because the knowledge base lacked the information (add new articles), and frequently asked questions that reveal gaps in your documentation (the questions your team asks most reveal what is under-documented).
Set a monthly knowledge base review: add any process changes from the past month, update any outdated information, and add new FAQs based on the past month’s conversation logs. An assistant whose knowledge base is actively maintained becomes more valuable over time — not less, as your team discovers they can rely on it for increasingly complex questions.
📌 Add a thumbs up / thumbs down rating to each assistant response. Team members who found the response helpful click thumbs up; those who found it unhelpful or inaccurate click thumbs down. The thumbs down responses are your highest-priority knowledge base improvement targets — review them weekly and update the relevant articles.
How do I handle sensitive information in the knowledge base?
Build role-based access to knowledge articles: add a ‘restricted’ flag to articles containing sensitive information (pricing structures, personnel policies, client-specific information). In the knowledge retrieval workflow, only include restricted articles when the querying user has the appropriate role. Regular team members get general knowledge; managers get management knowledge; executives get all knowledge. Role-based access ensures your AI assistant is comprehensive for authorised users without exposing sensitive information inappropriately.
What is the difference between this and simply giving my team access to ChatGPT?
Generic ChatGPT answers from general training data — it does not know your company’s specific processes, your clients’ names, your pricing structure, or your communication standards. Your custom assistant answers specifically from your company knowledge, in your company’s voice, with access controls appropriate to your team structure. Additionally, all conversations are logged in your system — giving you visibility into what your team is asking and where knowledge gaps exist. Generic ChatGPT provides none of this.
Want a Custom AI Team Assistant Built?
SA Solutions builds Bubble.io internal AI assistants — with your company knowledge base, role-based access, conversation logging, and admin dashboards for knowledge gap monitoring.
