AI COPILOT / EMBEDDED AI ASSISTANT
PROPELOO engineers embedded AI copilots — intelligent assistants that live inside your product, understand its context, suggest next actions, generate content in context and automate routine tasks without leaving the user's workflow. The best AI copilots are the ones users forget are AI — they just make the product better.
The difference between an AI chatbot bolted onto a product and an AI copilot built into it is context. A standalone chatbot does not know the current document the user is editing, the customer record they are looking at, the code they are debugging or the data they are analysing. An embedded copilot has full product context — it sees what the user sees, knows their history with the product and can take actions inside the product directly. This context is what makes copilots genuinely useful rather than marginally helpful. PROPELOO builds AI copilots that are context-aware from the ground up: every suggestion informed by the user's current state in the product, every action executable without leaving the workflow.
Deep integration with product state — current document/record/view, recent user actions, relevant data from the product database, user permissions and role. Context injected into every LLM call.
Proactive suggestions based on current product state — next action suggestions, content completion, error corrections, optimisation recommendations. Triggered by context change, not just user query.
Generate content within the product context — draft responses in CRM context, write code completion in IDE context, suggest next steps in project management context, generate report narrative from current data.
Copilot actions that execute directly in the product — create records, update fields, run queries, send messages, schedule tasks. User approves before execution or enables autonomous execution with undo.
RAG over product documentation, help content, company knowledge base and user's own historical data — copilot answers are grounded in real product knowledge, not hallucinated.
Cross-session memory of user preferences, common tasks, style preferences and corrections — copilot improves for each user over time without explicit configuration.
The hardest part of building an AI copilot is not the AI. It is the context engineering — what product state to include, how to represent it, and what to leave out.
Every LLM has a context window limit. Including too much context (the entire database record, every conversation history) fills the context with noise and costs money. Including too little produces generic suggestions. Context engineering — selecting the right product state at the right granularity — is the primary determinant of copilot quality.
Axiom:
A suggestion the user can ignore has zero risk. An action the user cannot undo is a liability. Every copilot action must have a clear undo mechanism. Copilot actions that cannot be undone (sending an email, deleting data) should require explicit confirmation with a summary of what will happen.
Axiom:
A copilot that waits 5 seconds and then displays a complete response feels like a tool. A copilot that starts typing immediately and refines its response as it generates feels like a thinking entity. Streaming is particularly important for copilots — the generation process itself communicates intelligence.
Axiom:
Users do not trust new AI copilots to take autonomous actions immediately. Build trust gradually: suggest only → approve-then-execute → auto-execute with undo. Let users configure autonomy level per action type. A copilot that users trust to act autonomously is significantly more valuable than one they must always supervise.
Axiom:
Impact: Command palette + side panel combination: ⌘K for quick actions (familiar pattern from Notion, Linear), side panel for longer interactions. Inline for content editors. Floating chatbot only as fallback for products that cannot integrate deeper.
Impact: Current view + summarised recent history is the right default. Let users explicitly add context (highlight text + "add to copilot context") for cases where the automatic context is insufficient.
Impact: Permission-tiered is the production standard: read actions auto, low-risk writes auto with undo, high-risk writes require approval. Let users configure tiers per action type.
Impact: Claude 3.5 Sonnet for copilots with long document context — its streaming start time and 200K context window make it ideal for document-aware assistants. GPT-4o-mini for simple suggestion completion tasks.
Impact: Hybrid: sensible defaults + user corrections captured and used to adjust future suggestions. The feedback loop from corrections is the highest-quality personalisation signal.
Impact: Redact PII before sending to third-party LLMs for B2B products handling sensitive data. On-premises deployment for regulated industries (healthcare, finance, legal).
In-CRM copilot that drafts outreach emails from contact history, suggests next actions, summarises account activity and generates call prep briefs.
IDE-integrated copilot with codebase-aware suggestions, error explanation, test generation, refactoring suggestions and documentation writing.
In-editor copilot for document creation — continuation suggestions, structure recommendations, source-grounded additions and style consistency checking.
Data platform copilot that translates natural language to SQL, explains query results, suggests related analyses and generates narrative summaries from charts.
Customer support copilot that suggests responses from knowledge base, summarises ticket history, identifies similar resolved tickets and drafts escalation notes.
PM tool copilot that generates task descriptions, suggests assignees based on workload, summarises project status and drafts stakeholder updates.
Stack: Claude 3.5 Sonnet, GPT-4o, GPT-4o-mini (suggestions), Vercel AI SDK
Stack: pgvector, Pinecone, LlamaIndex, Product state API
Stack: React (panel/inline), CodeMirror (code editor), Tiptap (rich text), cmdk (command palette)
Stack: FastAPI / Node.js, Redis (session/cache), WebSocket (streaming), PostgreSQL (memory)
Stack: Product API wrappers, Tool use (function calling), Action queue with undo, Approval workflow
Stack: Copilot usage analytics, Suggestion acceptance rate, Action completion rate, User satisfaction score
Product context sent to LLM APIs may contain sensitive data. PII redaction, data processing agreements with LLM providers, and on-premises models for regulated industries.
Copilot actions must check the same permissions as the user — the copilot cannot take actions the user is not authorised to take. Never bypass authorisation for AI actions.
Content in the product (document text, customer notes, code comments) can contain prompt injection payloads. Sanitise all product data before including in LLM context.
Copilot conversation history and context logs may contain sensitive business data. Retention limits, user-controlled deletion and clear privacy policy for copilot data.
Copilot backend handles LLM API keys server-side. Never in frontend code. Per-user rate limiting prevents abuse.
In healthcare, finance or legal products: every copilot suggestion shown, every action taken and every user correction must be logged for compliance audit.
Define use cases, context sources, action capabilities, UX integration points, permission model.
Product state API, context selection logic, RAG setup for product knowledge.
LLM integration, prompt engineering per use case, structured output for actions.
Command palette, panel, inline suggestions — integrated into existing product UI.
Tool use for in-product actions, approval flows, undo mechanism.
User preference capture, correction feedback loop, memory persistence.
Suggestion acceptance rate, action completion rate, user satisfaction, quality monitoring.
A chatbot is a separate interface where users switch contexts to ask questions. A copilot is embedded in the product workflow — it sees what the user sees, knows the current state of their work and can take actions inside the product. The defining difference is context: a chatbot has none, a copilot has everything the product knows about the current user and their work.
Key metrics: suggestion acceptance rate (% of copilot suggestions the user accepts — target >40%), action completion rate (% of copilot-suggested actions completed), time-to-task-completion (do users complete tasks faster with copilot?), support ticket volume (do copilot users raise fewer support tickets?), and retention/engagement (do copilot users retain at higher rates?).
Tiered action model: read-only actions (search, summarise) are safe and run automatically. Write actions (create, update) require explicit user confirmation for new users, auto-execute for trusted users with undo. Irreversible actions (delete, send external communication) always require confirmation with clear consequences displayed.