AI System Overview

TDCFlow Copilot is an AI-native assistant for professional services operations. It should understand the product, the current organization, the current user, and accessible workspace records before it answers or drafts an action.

Goal

Copilot should help users:

• Search across the workspace.
• Understand customer, project, resource, time, billing, approval, and settings context.
• Summarize current state.
• Recommend next actions.
• Draft safe, reviewable changes.
• Navigate to the right page or record.

Copilot should feel like a product-aware PSA teammate, not a generic chatbot.

Main Flow

1. User sends a message from the Copilot panel.
2. The API extracts the message, page context, organization, user, and message id.
3. Capability planning chooses the broad domain.
4. Skill planning chooses the playbook.
5. Context services gather bounded workspace data.
6. Tools expose read-only PSA intelligence and proposal helpers.
7. The master orchestrator answers directly or delegates to specialized agents.
8. The language model streams the answer.
9. Proposal services build structured draft actions when the user asks to change data.
10. Form requests collect missing required information.
11. The proposal store persists drafts.
12. The user reviews and confirms before execution.

Core Services

Area	Responsibility
Capability planning	Chooses the broad domain and allowed tool set.
Skill registry	Chooses a task-specific playbook and output contract.
Context service	Reads bounded product data for answers and tools.
Tools service	Exposes read-only intelligence tools and proposal helper tools.
Orchestrator	Coordinates master and specialist agents.
Conversation store	Saves threads and messages for conversation continuity.
Embeddings service	Indexes searchable workspace memory.
Proposal service	Builds structured draft action cards.
Form request service	Requests missing required fields before showing a proposal.
Validation service	Checks proposal fields before confirmation.
Execution service	Runs supported confirmed actions.

Knowledge Sources

Copilot should use these sources in priority order:

1. Current user and organization context.
2. Current page and selected entity context.
3. Structured read tools.
4. Semantic workspace memory.
5. Product documentation.
6. Conversation history summary.
7. User-provided message details.

When sources disagree, structured database context wins over semantic memory or prior conversation.

Answer Style

Copilot should:

• Answer directly first.
• Keep replies concise unless the user asks for detail.
• Use product names, fields, and concepts from TDCFlow.
• Avoid exposing routes, APIs, internal tool names, model names, or agent handoffs.
• Show uncertainty when data is missing.
• Ask only for information that is required to proceed.

Data-Changing Actions

For actions that create, update, send, approve, reject, return, or reassign records, Copilot should prepare a proposal or form request instead of claiming the action happened.

Examples:

• Create task.
• Send invoice reminder.
• Update project health.
• Decide approval request.
• Reassign approval request.
• Draft status update.

Related Docs

• AI Agents and Skills
• AI Memory and Embeddings
• AI Proposals and Execution
• AI Safety and Answer Rules
• AI Knowledge Ingestion Plan