The problem: questions that never get consistent answers
Teams ask the same questions every week:
- What's blocking the release?
- Who owns the API spec?
- What did we decide about pricing in Q3 planning?
- Where is the onboarding doc?
Answers scatter across Slack threads, personal notes, and stale Confluence pages. Someone pastes a CSV into ChatGPT and gets a confident wrong summary. New hires interrupt seniors. PMs become human search engines.
AI team Q&A only works when answers cite live project data and published knowledge — not model imagination.
Why generic AI chat fails for teams
Copy-paste workflows break down fast:
Stale snapshots — Export from Friday; question on Monday; tasks moved.
No permission model — Contractor questions leak finance or HR context.
Missing policy layer — Status lives in tasks; why lives in wiki — generic chat sees neither completely.
No audit trail — "The AI said we were green" with no task IDs to verify.
Teams need retrieval + reasoning over authoritative sources, scoped to what each person may see.
Framework: sources → retrieval → answer → verify
Structure team Q&A in four steps:
1. Sources — Tasks, approvals, wiki pages, meeting decisions, documents, finance (role-gated).
2. Retrieval — search_docs, wiki vector search, list_tasks with filters, get_task with history.
3. Answer — Natural language with citations: task ID, wiki title, approval status.
4. Verify — Human spot-checks high-stakes answers; empty tool results mean "not found," not invented blockers.
Train the org to ask verifiable questions:
- "List overdue tasks on Project Atlas with IDs"
- "Search wiki for change-request policy — quote title"
- "What approvals block release train 12?"
Avoid vague prompts unless tied to measurable data.
Where WKFGo fits: three grounded paths
In-app chatbot
The WKFGo chatbot (port 4000) shares PostgreSQL with the backend. It classifies intent, maintains conversation context, and uses wiki vector search for knowledge questions. Connect via REACT_APP_CHATBOT_URL — floating widget in the authenticated app shell.
Good for: quick status, wiki lookup, task guidance, finance report requests (permission-gated).
MCP for power users
Model Context Protocol connects Cursor, Claude Desktop, and compatible clients to typed tools:
list_tasks,get_task,smart_searchsearch_docs,get_wiki_pagelist_pending_approvals,get_project_reportexecutive_brief,my_day,my_queue
The model invokes tools instead of inventing rows. Same feature access as the web app.
Wiki and meeting publish
After publish_meeting_wiki, decisions become searchable. Q&A on "what did we decide?" resolves to wiki pages — not oral tradition.
Adoption ladder: read before write
| Stage | Use | Risk |
|---|---|---|
| 1 | Read tasks, search wiki, approvals | Low |
| 2 | Draft comments / descriptions | Medium — human post |
| 3 | Create tasks from templates | Medium — triage rules |
| 4 | Update status, assign | High — audit required |
Most teams should live in stages 1–2 for a quarter before stage 4.
Getting started this week
Document ten recurring team questions — status, policy, ownership, onboarding. For each, note whether the answer lives in tasks, wiki, or approvals. Configure chatbot for daily users; MCP read tools for leads in IDE. Run a drill: ask the chatbot three verifiable questions and confirm citations. Fix gaps by publishing one missing wiki page from last week's meeting. Ban paste-into-generic-chat for status questions in team norms.
Bilingual teams: maintain wiki keywords in both working languages so search_docs returns canonical pages regardless of query language.
Anti-patterns
- Trusting answers with no citations
- Service account MCP that sees everything
- Skipping wiki publish after meetings
- Using AI for HR-sensitive topics without human gate
Team questions repeat when answers live in private DMs. Route 'how do we…' to wiki pages and search_docs before chatbot. Permission-aware MCP ensures contractors never see finance answers in the same thread as engineers—feature access is enforced server-side, not by prompt politeness.
Document the top ten recurring team questions each quarter; turn answers into wiki pages. Chatbot accuracy rises when source pages exist—not when prompts get longer.
FAQ
Does the chatbot replace MCP?
No — chatbot serves in-app users; MCP serves IDE and desktop assistant workflows. Both read the same data.
Can bilingual teams query in Farsi and English?
WKFGo supports bilingual UI; search works best when wiki content matches team working language.
What if the tool returns empty?
Correct behavior: "no overdue tasks found" — not three invented blockers.
Is write access available via AI?
Yes for some MCP tools — configure policy; default to read-first adoption.
Answer from truth, not memory
Give your team grounded Q&A — chatbot for daily use, MCP for deep work, wiki for decisions that outlive Slack.
Try it now
Put these patterns on live project data—not slide decks.