AI Command Center

The AI Command Center at /ai is the single control surface for every AI feature on the platform. Every Claude call we make is gated by an agent config you set here — when you flip a module off, AI for that surface really stops.

The seven tabs

Tab	What it controls
Overview	Live activity snapshot, getting-started checklist, status of every wired module
Agents	Per-module mode, prompt, model, knowledge bindings, escalation, rate limits
Knowledge	Central knowledge base — items here are pulled by every wired agent at runtime
Skills	Curated action bundles you can install for an agent (Sales Assistant, Support Bot, etc.)
MCP	Model Context Protocol servers — adds extra tools to your agents
Approvals	Pending AI-drafted actions awaiting human OK (created by supervised-mode agents)
Analytics	Per-org usage: requests, tokens, latency, acceptance, cost, per-module breakdown
Settings	API key status, default model, escalation keywords, web scraper, auto-tasks, role permissions

Agent modes

Every module's agent has one of four modes:

Off — AI is fully disabled. Nothing runs. API calls return 403 AGENT_DISABLED. Background pipelines (email classifier, transcript correction) gracefully degrade.
Copilot — AI generates drafts but never sends. You review and send manually. Good starting mode.
Supervised — AI drafts land in the Approvals queue. A human approves before send.
Autonomous — AI replies and acts on its own (within rate limits and escalation rules). Use after you trust the agent.

LinkedIn agents cannot be set to Autonomous — LinkedIn ToS prohibits AI auto-DM, so the save endpoint blocks it.

Modules currently wired

When you change a module's mode, these are the AI surfaces that respect the change:

Module	What the agent controls
WhatsApp	Inbound message replies (Cloud API + Baileys). Audience-aware (employee vs customer). Tool use in autonomous mode.
Email	Reply drafting, auto-categorization (inbound classifier Lambda gates on this module — turn off and emails land uncategorized rather than mis-classified)
Instagram	DM auto-replies via Meta webhook
LinkedIn	DM drafts on inbound message (copilot/supervised only)
Reddit	Inbound reply pipeline
Live Chat	Org-level chat agent invoked in parallel with per-chatbot widget flow
Discussions	`@ai` mentions in comment threads dispatch through the agent
CRM	Prospect enrichment, prospect research, rep call coaching
Tasks	Extract-tasks-from-document/URL feature
Meetings	Ask-AI Q&A over transcripts + notetaker transcript correction Lambda
Dashboard	Sales insights generated for the Wrrkspace dashboard

Knowledge Base — centralized

All wired agents pull from the same central knowledge base at /ai/knowledge. Items can be:

Global (All Modules badge) — visible to every agent
Module-scoped — only visible to the listed modules' agents

Each agent can bind to a subset of the KB on its config page. Empty selection = use the entire org-wide KB.

Vector embeddings are generated on save (Bedrock Titan v2) and stored in the org's vector index. Retrieval is hybrid: keyword + cosine similarity, filtered by module visibility.

Two intentional carve-outs:

WhatsApp session KB: per-session text strings stored in the WhatsApp connection config. Legacy — these are not vectorized and do not participate in central retrieval. Migration to the central KB is recommended for new sessions.
Chatbot widget RAG: each public chatbot has its own embedding namespace, scoped to that bot's training documents. Intentional — protects per-bot training from leaking across bots.

Analytics — what we record

Every Claude call across the platform writes a telemetry row to the org's analytics table within milliseconds of completing. Each row records: module, type (chat / compose / summarize / analyze / agent_response / agent_escalate / error), tokensUsed, latencyMs, model, userId, orgId, and timestamp. Rows expire after 90 days.

The Analytics tab queries this table by day-bucket GSI and renders:

Estimated spend — recomputed each load from actual tokens recorded against the model's current Anthropic pricing
Total requests / tokens / avg latency — supporting metrics for the selected period
Suggestion acceptance — % of suggestions that were accepted vs. dismissed
Usage by type / module — stacked bars showing which AI surfaces are hot
Hourly volume — last 24h trend chart
Agents performance table — per-agent responses, escalations, acceptance, avg time
Insights — automated callouts (high error rate, slow responses, milestones)

Approvals queue

When any agent runs in Supervised mode, every AI-drafted action lands in /ai/approvals instead of sending. Each pending action shows:

The original incoming message / trigger
The AI's drafted response or proposed action
The agent that drafted it
Approve / Reject / Edit buttons

Approvals expire after the time set in Settings → Escalation. Expired pending actions are dropped.

Settings — General tab

API key — green badge means the Anthropic API key is configured. Required for any AI to run.
Default model — Claude Haiku 4.5 (global.anthropic.claude-haiku-4-5-20251001-v1:0). Fast, cheap, good for the vast majority of platform calls.
Temperature — 0.0 deterministic to 1.0 creative. Default 0.7.
Max tokens per request — hard cap. 4096 is a reasonable default.

Settings — Inbox tab

Controls the per-org AI Inbox Enrichment pipeline (separate from per-module agents). Three toggles:

Enable AI enrichment — master switch. When off, the inbox still shows every message, just without AI triage.
Show AI decorations — urgency dots, intent tags, sentiment indicators. Backend enrichment runs either way; this controls what your team sees in the list.
Auto-suggest reply on selection — generate a draft the moment you open a thread.

Settings — Escalation tab

Keywords and phrases that trigger handoff to a human agent. Applies to every wired agent that handles inbound messages.

Human request keywords — human, agent, representative, etc.
Negative sentiment words — escalate on detected frustration
Request phrases — exact phrases like "talk to a human"
Intent keywords — grouped triggers by intent (complaint, urgent, refund)

Settings — Web Scraper

Tuning for the crawler that runs when you import a URL into the knowledge base or use the scrape_url tool in an agent.

Settings — Auto Tasks

Per-channel rules for the auto-task-extraction pipeline. When a message qualifies, AI extracts actionable tasks and queues them on the configured board with the configured confidence threshold.

Settings — Access Control

Per-role cap on the maximum agent mode each user can configure. Default: Super Admin can set autonomous, Manager can set up to supervised, Agent can set up to copilot.

When something goes wrong

Analytics shows 0 — confirm the API key is configured at /ai/settings; confirm the relevant module's agent is enabled at /ai/agents.
Agent doesn't reply — check Approvals queue (supervised mode), check rate limits on the agent config, check escalation rules didn't fire.
403 `AGENT_DISABLED` — the module's agent at /ai/agents is off or set to mode=off. Toggle on and retry.
Email arrives uncategorized — Email agent is off. Classification gracefully degrades; the email lands in the inbox unmodified.
Transcript not corrected — Meetings agent is off. Raw ASR transcript persists; no AI correction runs.

Best practices

Start every module in Copilot mode for a week. Let your team review drafts. Promote to Supervised once drafts feel reliable.
Promote to Autonomous only after a clean week in Supervised with high acceptance and zero bad sends.
Add knowledge proactively — the AI is only as good as what it can look up.
Watch /ai/analytics weekly. High escalation rates, low acceptance, or slow latency are the early signals something needs tuning.
Re-tune escalation keywords every quarter as your customers' language and complaints shift.