AI Command Center

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

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:

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 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.

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

• 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.
• Re-tune escalation keywords every quarter as your customers' language and complaints shift.