Skip to main content
Margaret is the AI assistant built into Switchbord. Open her from the sidebar icon (or the keyboard shortcut shown in the top bar) and ask anything from “summarize this conversation” to “create a template for my winback campaign.”

​
What Margaret can do today

Answer questions

Explain any Switchbord feature, what a setting does, or how to achieve a workflow.

Look up CRM data

Find a contact by name, phone, or email. Show their recent conversations, lifetime value, and notes.

Draft replies

When you’re on an inbox conversation, ask “reply asking about their preferred dates” — Margaret drafts in context.

Create templates

“Create a template for birthday wishes with the customer’s name” — Margaret proposes a draft you review.

Design journeys

Describe a flow in plain language and Margaret drafts a journey graph.

Search your knowledge base

Query workspace FAQ entries and internal notes.

​
Page-aware grounding

Margaret knows what page you’re on and what you’ve selected. On an inbox conversation, she sees the contact profile, the last six messages, and the assignee — no need to paste context. On the template editor, she sees the current draft body. On the journey editor, she sees the graph shape and your selected node.
When you ask “respond to this customer” on an inbox conversation, Margaret picks the right conversation automatically. You don’t need to paste an ID.
Pages that publish context:
  • Inbox — selected conversation, contact, last messages, assignee
  • Templates — current draft, category, language
  • Journeys — graph status, selected node, triggers/actions
  • Campaigns — composing draft, template name, audience size
  • Contacts — selected contact, filters, list size
  • Settings — current tab
Tool responses also receive this context, so Margaret’s answers stay scoped to your workspace and the artefact in view.

​
Slash commands

Type / as the first character in Margaret’s composer to open a fast-lane menu of common operator actions. The popover is keyboard-driven: use ↑ / ↓ to move between commands, Enter or Tab to pick one, and Esc to dismiss. Slash commands are a shortcut, not a restriction — you can always type the full prompt yourself. As soon as you type a space after the /, the popover dismisses so the composer treats your input as a normal message.

​
Workspace memory

Margaret has durable, workspace-level memory. Tell her a fact once and she can recall it in any future thread.

Save a fact

“Margaret, remember that our VIP cut-off is 5 000 EUR lifetime spend.”Margaret proposes an approval card with a title and body. Click Approve to save.

Recall a fact

“Do you remember what our VIP cut-off is?”Margaret searches the workspace knowledge base and cites the saved note in her reply.
Memory entries live in the same workspace_knowledge_entries table as your FAQ / policy notes (from Settings → AI). Entries written by Margaret are tagged memory so you can distinguish them at a glance.
Writing to workspace memory is admin-only and always approval-gated. Margaret describes what she’s about to save before the approval card appears, so you can review the wording.
Good things to remember:
  • Stable facts that will matter in future threads (pricing cut-offs, branch names, policy numbers).
  • Preferred wording or tone guidelines (“always refer to the Florence branch as FI”).
  • Contact- or segment-level conventions that aren’t in your CRM.
Don’t use memory for transient task state (e.g. “the user asked me to send a campaign tomorrow”). That’s what threads are for.

​
Model selection

Click the model dropdown in Margaret’s header to switch between supported models. Each provider has strengths:
  • Claude Sonnet / Opus — best for creative drafting and long reasoning
  • GPT-5 / o-series — best for complex planning
  • Gemini 2.5 — large context, fast
  • DeepSeek R1 — open, strong reasoning
  • Claude Haiku / GPT-4o — fastest, no extended thinking
Your selection persists across sessions (localStorage).

​
Thinking (reasoning) controls

Next to the model dropdown, the Thinking selector lets you decide how hard Margaret thinks before answering:
  • Auto — let the model’s own default apply. (Recommended for everyday chat.)
  • Off — disable reasoning. Fastest, lowest cost.
  • Fast — low effort. Good for quick lookups.
  • Balanced — medium effort. Default for most reasoning-capable models.
  • Deep — high effort. Use for complex analysis, multi-step plans, tricky drafting.
The dropdown is disabled (with a tooltip) when the selected model does not support extended thinking.
When Margaret thinks, you’ll see a collapsible “Thinking…” block in the transcript that streams live. It auto-collapses once the final answer begins. Click it to expand and review her reasoning.
Extended thinking uses more tokens and takes longer — reasoning tokens are billed as output tokens. The “Deep” setting can roughly 2-5× the cost of a normal response. Use it selectively.

​
Drafts & approvals

When you ask Margaret to create, update, or send something, she never commits silently. The flow is:
  1. Margaret explains what she’s about to do, in one sentence.
  2. She proposes a draft — a template body, a journey graph, a message reply.
  3. You review the draft inline in the chat.
  4. You click Approve to commit or Reject to discard.
This applies to templates, journeys, campaigns, contact notes, outbound messages, and anything else that mutates data.
Margaret will never send a WhatsApp message or commit a database change without your explicit approval. If a tool call could have a surprising consequence, she’ll ask for clarification first.

​
Good prompts to try

  • “Who is this contact and what have we talked about?” (on an inbox conversation)
  • “Draft a reply in Italian asking about their preferred departure date.”
  • “Create a utility template called document_reminder that asks for a missing document with the customer name and document type as variables.”
  • “Design a lead-capture journey that triggers on the keyword ‘preventivo’ and collects name, phone, email.”
  • “What’s our cancellation policy?” (pulls from the workspace knowledge base)
  • “Find contacts tagged vip who haven’t had an inbound in 60 days.”
Margaret can search a per-workspace knowledge base consisting of FAQ entries with titles, bodies, and tags. These entries are managed in Settings → AI → Knowledge Base. When a user asks a question that isn’t about their specific account data (e.g., “What is your cancellation policy?”), Margaret uses semantic search to find the most relevant entries and uses them to ground her answer.

​
CRM tools

Margaret has direct access to your customer data. She can:
  • Find contacts by name, phone number, or email.
  • Get contact details including recent conversations, custom fields, and tags.
  • Update contact information such as names or custom field values (requires approval).
  • Tag/Untag contacts for segmentation (requires approval).

​
ElevenLabs KB sync

If you use ElevenLabs for voice agents, Margaret can help you keep their knowledge in sync with Switchbord. You can ask her to “Push our KB to ElevenLabs” to initiate a sync (see the ElevenLabs KB guide for details).

​
Workspace grounding — knowledge base

Margaret can search a per-workspace knowledge base: FAQ entries with title, body, and tags. Seeded during setup; operators add entries in Settings → AI. The assistant uses these to ground answers in your organization’s actual policies. Margaret’s own workspace memory entries share the same table, tagged memory.

​
Limitations

  • Margaret’s cross-thread memory is opt-in via writeWorkspaceMemory — she doesn’t yet automatically remember everything you discussed in a previous thread. Full auto-condensation is on the roadmap.
  • File uploads and multimodal (images) aren’t supported yet.
  • Voice mode isn’t available.
  • Evaluations suite is in development — occasional hallucinations are possible; always review drafts before approval.

​
Observability

Every Margaret turn is logged to llm_usage (provider, model, tokens, latency). A dedicated /settings/ai-observability dashboard is tracked as a follow-up (BORD-328b). Approvals and mutations already produce audit-log rows tied to your user id.

​
Roadmap

See Linear epic BORD-323 (Margaret v2). Shipped so far:
  • BORD-324 — foundation fixes (model selector, a11y, error surfacing) ✅
  • BORD-325 — reasoning controls via OpenRouter ✅
  • BORD-326 — page context / grounding ✅
  • BORD-327 — draft artefacts + native HITL approvals ✅
  • BORD-328 — slash commands, workspace memory, eval harness ✅
In flight or planned:
  • BORD-328b — /settings/ai-observability dashboard (deferred from 328)
  • BORD-322 — ElevenLabs knowledge-base integration (uploads)