01
order_detective
Works out which of the customer's orders the message is about: the order number when they give one, otherwise their email address and what they wrote.
// output
order_match
// cs specialist
AI that answers customer email.
Magistry answers customers from your store's IMAP or Outlook mailbox, and the replies are sent. Every answer is grounded in the real order, cites your own policies, and has to pass the Reply Judge before it sends. Refunds and cancellations execute inside limits you set, and return requests are handled in the helpdesk.
Live Email · IMAP / Outlook Reply Judge Proactive Outbound
thread · example
from : klant@protonmail.com
subject : Bestelling 88421 niet aangekomen
lang : nl (reply in customer language)
order_match : o_88421 (order number in subject)
product : Brushed Linen Throw — Sand
ship_status : in transit, NL (carrier ping 6h ago)
supplier : on-time batch, no flags
policy_cite : cs_policy_chunks#214
"EU shipping ETA — 5–8 business days"
draft : "Hoi! Goed nieuws — je throw is
onderweg en zou tussen vrijdag en
maandag aankomen. Hier is de track-link…"
judge_score : 0.91 → APPROVED → sent// the skills
Every reply has a reasoning trail.
Specialist skills compose each reply, refund and escalation. Every step is recorded, so you can read why the agent thought it knew the answer as well as what it said.
02
catalog_xref
Checks the product the customer mentions against your catalog, so a reply about a product you have since pulled or one with a known complaint pattern says so.
// output
context_payload
03
supplier_reality
Turns fulfilment status and open supplier delays into a truthful shipping statement. If the supplier is the cause, the reply says so honestly.
// output
supplier_signal
04
policy_grounder
Retrieves the relevant sections of your shipping, refund, returns and warranty pages. A reply that states a policy must cite the section it came from.
// output
policy_citation
05
reply_drafter
Replies in the language the customer wrote in, in the tone of the replies your team already sends. Nothing leaves until the Reply Judge clears it.
// output
cs_outbound_drafts
06
refund_ev_calculator
Weighs a full refund, a partial refund, a replacement, a gift card, a polite no or a hand-off by expected value: margin, customer value, fraud risk and what your policy allows.
// output
ev_recommendation
07
action_executor
issue_refund · cancel_order · create_return · issue_replacement · change_shipping_address · issue_gift_card. Each action carries an idempotency key and is recorded.
// output
shopify_mutation
08
escalate
VIP customers, high fraud risk, legal language or a request above your refund cap go to a person, with the reasoning trail attached.
// output
human_handoff
// channels & memory
One desk, with memory.
Replies are sent from your own mailbox. In the last 30 days (measured 17 Oct 2026), 658 replies were sent from connected mailboxes, 104 of them by people.
Connect your mailbox over IMAP, or connect Outlook. Threads sync in, and replies send from your own address.
Returns
Return requests are handled in the helpdesk, alongside the rest of the conversation.
CSAT surveys
A satisfaction survey goes out after a conversation is resolved.
Your policies
Your shipping, refund, returns and warranty pages are indexed, and the agent answers from what you have written.
Resolution memory
Resolved threads become examples the agent can retrieve, so the same issue next month starts from the answer that worked.
Hard constraints
Refund caps, a fraud gate, a language gate and policy grounding are enforced in code after the model has answered, so the model cannot talk itself past them.
// the real product
The inbox, with the judge's scores visible.
A real screenshot of the dashboard, with demo store data. Threads on the left, AI drafts on the right, each scored by the Reply Judge before it sends.
// reply judge
A second pass before any reply ships.
The Reply Judge is a second model pass that scores every AI draft on policy grounding, factual grounding and brand voice. A draft it does not pass is not sent on its own; it goes to a person.
policy
Does the draft only promise things your published policy actually allows?
e.g. Promised return window matches the cited cs_policy_chunk.
fact
Does every claim trace back to the order data, supplier signal, or catalog?
e.g. Tracking ETA cited from the actual carrier event, not invented.
brand_voice
Does the draft sound like you, compared with replies your team has already sent?
e.g. Closest match to your past replies is above the floor you set.
risk
Is the request inside the trust window, or does it need a human?
e.g. Refund above your cap → escalate, whatever the draft quality.
cs_thread_evaluations · example
thread_id : t_4912
draft_id : d_1180
language : nl (matches customer ✓)
judge_score : 0.91
├─ policy : 0.94
├─ fact : 0.92
├─ brand_voice : 0.88
└─ risk : low (refund €18.50)
policy_cite : cs_policy_chunks#214
"EU returns — 30d, prepaid label"
fact_cite : order_payload#o_88421
tracking: 'in transit, NL'
verdict : APPROVE_SEND
sent_at : 2026-05-25T11:02:18Z// brand voice embeddings
Your tone, learned from your own replies.
Magistry embeds the replies your team has already sent. The drafter imitates them, and the judge measures how close each new draft comes.
Learned from sent mail
Recent replies from your mailbox are embedded as voice samples.
your sent repliesSteer every draft
reply_drafter imitates your samples for tone, sign-off, formality and length. Even a new policy sounds like you.
drafter conditioningMeasure every send
The Reply Judge scores each draft against its closest sample. Below the floor you set, the reply waits for you.
floor you set// refund_ev
The math behind every yes.
refund_ev_calculator weighs each possible answer (refund, partial refund, replacement, gift card, polite no, hand-off) on the order's margin, the customer's value, fraud risk and your policy, and picks by expected value.
// example scenario
Customer received the wrong colour throw. €89 order. Repeat customer, inside the return window, no fraud signals.
- full_refundfull cost, customer kept
- partial_refundlower cost if the item is kept
- replacementproduct cost, customer kept
- escalate_humana person decides
- deny_politeoutside policy: not here
In this example Magistry picks replacement, inside the refund cap you set.
LTV-aware
A customer likely to buy again is worth more flexibility than a one-off order.
Margin-aware
The contribution margin on the product is part of the sum.
Risk-aware
Fraud risk and a third refund request in 90 days count against an automatic yes.
Bounded
Magistry never resolves above the per-action ceiling without an escalation row.
// proactive outbound
Replies before the customer asks.
When a supplier delay holds up orders, a product is pulled for quality, or complaints cluster on one product, Magistry drafts a message to the affected customers before they write in. The first batch waits for your approval; later ones can send on their own once you allow it.
cs_outbound_drafts · example
trigger : supplier.offline on 'Brushed Linen Throw'
customer : jane.k@***.com
order : o_88421 — 'Brushed Linen Throw'
draft_subj : "Quick update on your order #88421"
draft_body : Hi Jane, your throw is held up at our
supplier and will not ship this week.
You can wait for it, or reply and we
will refund the full amount.
judge_score : 0.93 (policy ✓ fact ✓ voice 0.91)
status : PENDING_REVIEW
first batch for this signal: approval requiredAllow proactive sends in your customer service settings
// cs specialist
Reply faster than the customer can finish typing.
Connect your IMAP or Outlook mailbox and point Magistry at your policy pages. Drafts start arriving for you to approve, and you decide when replies may send on their own.
Language-gated · Brand-voice scored · Judged before send