Maya ReyesWhere’s my order — and can I still swap sizes?
Auto-sent · 0.96
Book a Demo AI EMAIL AGENT · CONFIDENCE-GATED
The AI email agent for customer service that drafts, replies, and resolves.
Support inboxes rot: first replies take hours, and templates answer the wrong question. LumiTalk is an AI email agent for customer service that reads each message, pulls the customer’s real context from your systems, and sends a complete, accurate reply — and when it isn’t sure, it drafts instead of sending.
Live demo line · ask it what the email agent does
(831) 216-4331
Try: “What would you do with a where's-my-order email?”
- ✓picks up instantly
- ✓answers live
- ✓same engine as email
No signup, no sales rep — just call it.
LIVE
Re: Scheduling my follow-up with Dr. Chen
1 messageSarah M. ↔ Lumi · AI receptionist
S
Sarah M. <sarah.mendez@gmail.com>
to Bright Smile Dental ▾
Hi — I saw Dr. Chen recently and she recommended I come back for a follow-up, but I left without booking it. I'd like to get that on the calendar before I forget again.
Tuesdays generally work best for me, ideally in the afternoon. Thursday afternoon could work as a backup. One more thing: I've had some cold sensitivity come back in the same tooth — nothing severe, it mostly bothers me when I drink something cold. I figured you should know before the visit.
Thanks, Sarah M.
What is an AI email agent?
An AI email agent is software that reads incoming customer emails, identifies the sender and their real context from connected business systems, composes a complete reply that addresses every question in the message, and either sends it automatically when confidence is high or routes it as a draft for human review when it is not.
Three very different tools share this name — here’s which one this page is about:
- Support-inbox automation — what Lumi does. It reads inbound customer email, resolves it against your real systems, updates the record, and auto-sends only when confident. Ongoing sequences and reminders run through our Workflows & Follow-Ups product, not the inbox agent.
- Sales-outreach email tools write and sequence cold prospecting campaigns to people who never emailed you — a different job, and not what this product does.
- Inbox-assistant drafting tools help one person write their own mail faster, but never look up an order, touch a calendar, or update a CRM.
THE CONFIDENCE GATE
Auto-send is earned per email — not assumed.
Every reply gets a confidence score before anything leaves the inbox. Lumi sends on its own only when identity, data, and policy all check out. Everything else becomes a finished draft a person approves in one click — so the failure mode of an uncertain answer is a short review, never a wrong email in a customer’s inbox.
Lumi Inbox · support@yourbrand.comCONFIDENCE-GATED
Jordan PikeRefund still missing from my statement
Draft · reviewAnika ShahMove Thursday’s appointment to next week
Auto-sent · 0.94Tom WhitfieldInvoice copy + a cancellation question
Draft · reviewFROM Maya ReyesSUBJECT Where’s my order — and can I still swap sizes?
“Hi — my order hasn’t moved in a few days and I’m starting to worry. Also, I ordered a small but I think I need a medium. Is it too late to change it?”
Matched by sender addressOrder #7241 · ShopifyExchange window · open
LUMI · DRAFTED REPLY
Hi Maya — your order left our Memphis facility this morning and is on track for Thursday. And you’re not too late: the exchange window is open until Aug 14, so I’ve included a secure link below to switch to a medium before it ships. Both answered in one place — nothing else needed from you.
CONFIDENCE 0.96 · THRESHOLD 0.90AUTO-SEND
Reply sent · 42 seconds after arrivalLogged to the thread and CRM · Jordan’s refund routed to a human draft
- ✓Sender matched to a single customer recordIdentity resolved from the email address — no “please provide your order number.”
- ✓Order located and status verifiedLive carrier and fulfillment data, fetched seconds before the reply.
- ✓Answer grounded in approved policyThe exchange-window claim traces to your policy page, not model memory.
- ✓Nothing sensitive in the reply bodyIdentifiers masked; the size swap happens behind a secure link.
- !Two customer records share this addressLumi won’t guess which account the refund belongs to.
- !Refund amount exceeds auto-approval$180 requested; your policy caps unreviewed refunds at $100.
- !References a promise Lumi can’t seeThe customer mentions “what your agent told me on the phone.”
You set the bar — and move it as trust builds
Start in draft-only mode where nothing auto-sends. Watch a week of drafts, then raise autonomy category by category: order status first, refunds last, complaints never. For the design thinking behind that escalation moment, read our guide to the AI-to-human handoff.
HOW IT WORKS
Read. Compose. Send — or route.
Three moves per email, each leaving an artifact you can audit.
Read the email, fetch the facts
Lumi parses every distinct question in the message, resolves who’s writing from the sender address, and pulls their real records — the order, the appointment, the invoice — from your connected systems.
CONTEXT ASSEMBLED
- Maya Reyes · matched by sender address
- Order #7241 · in transit · ETA Thu
- Exchange window · open until Aug 14
Compose one consolidated reply
One email that answers every point — in your tone, with identifiers masked to last-four and a secure authenticated link for anything sensitive. No reply-per-question fragments, no “please see our FAQ.”
DRAFT COMPOSED
- Answers 2 of 2 questions in thread
- Card shown as •••• 4417
- Secure link inserted for size swap
Send it — or route it
The reply is scored against your threshold. Above it: sent within seconds and logged to the thread and your CRM. Below it: a finished draft, with the uncertainty flagged, waiting for a one-click human review.
GATE DECISION
- Confidence 0.96 ≥ threshold 0.90
- Auto-sent · 42s after arrival
- Thread + CRM record updated
HOW IT COMPARES
AI email agent vs. what most inboxes run on today.
Most support email is handled by filters and canned responses, or by a shared inbox and sheer effort. Here’s where each approach actually differs — no vendor scorecards, just the mechanics.
| What the inbox needs | AI email agent | Rules, filters & canned responses | Shared inbox only |
|---|---|---|---|
| Understands intent | Parses every question in the message — even three asks buried in one email. | Matches keywords; the wrong macro fires as often as the right one. | A human reads it — whenever a human gets to it. |
| Drafts vs. resolves | Composes a complete reply from live records; the answer is the resolution. | Sends a template that acknowledges the email but rarely answers it. | Every reply is drafted by hand, from scratch. |
| Updates your systems | Writes the outcome back to the thread and your CRM automatically. | No — the reply and the customer record never meet. | Only if someone remembers to log it after sending. |
| Confidence gating | Scores every reply; below your threshold nothing auto-sends. | None — a rule fires whether or not it actually fits. | Not applicable; nothing sends itself. |
| Handoff to humans | Uncertain replies arrive as finished, flagged, one-click drafts. | Misfires surface later — as customer complaints. | The queue is the handoff, and the bottleneck. |
WHAT IT HANDLES
The email your team answers on repeat — answered.
Start with the categories below, verify quality in draft mode, and widen the gate as each one proves out. Order-status mail is usually the biggest slice — our guide on reducing WISMO tickets with AI covers the prevent-and-resolve playbook behind it.
Order status and delivery
Finds the sender’s order without asking for a number and replies with real carrier status and the actual next step — not a tracking link and a shrug.
Appointment changes
Reads the calendar, offers genuinely open slots, and confirms the move in the same thread — the reschedule is done, not “noted.”
Billing questions
Explains the specific charge with masked identifiers, and hands anything that needs payment details to a secure authenticated link.
Returns and exchanges
Checks eligibility and windows against your policy, then puts the label or secure return link directly in the reply.
Document requests
Sends the invoice, receipt, or intake form the account actually calls for — validated against your records, never guessed.
Multi-question emails
The real-world email with four asks buried in three paragraphs gets one reply that answers all four.
EMAIL-NATIVE BY DESIGN
Rules an email agent should never break.
Email has its own physics: the sender is known, the record is permanent, and anything in the body can be forwarded. Lumi is built for that — not a chatbot pasted into a mailbox.
Never asks who you are
The sender address is the identity. Lumi matches it to your records before writing a word, so customers never get the “please confirm your order number and the last four digits” opener that makes them repeat what your systems already know.
One reply, every answer
Customers write one email with several questions. They should get one email with several answers — not a partial reply that spawns a week of follow-ups. Consolidation is the default, not a feature.
Masked identifiers, always
Last-four only. Full card numbers, account numbers, and record identifiers never travel in an email body — in Lumi’s replies or in the drafts your team approves.
Secure links for the sensitive parts
Payments, document uploads, and account changes happen behind authenticated links — the email carries the answer and the doorway, never the sensitive transaction itself.
Connected customer service
Use approved knowledge, live business context,
and an AI agentic layer — not isolated chatbot answers.
LumiTalk brokers every conversation across your channels and apps,
using real-time data and workflows to take action and respond with the right context.
AIOne AI agent · Understands · Plans · Acts · Responds
AI agentic layer
Live orchestration across every conversation
◎UnderstandDetect intentReturn request detected
⌘DecideSelect toolsPolicy + Shopify selected
ϟActRun workflowExchange order created
▤RememberUpdate contextCustomer history updated
▱RespondSame channelConfirmation sent by SMS
Now
Return request detectedPolicy + Shopify selectedExchange order createdCustomer history updatedConfirmation sent by SMS
↪Channel
input
input
⠿App
actions
actions
♢
Guardrails & governance
Approved knowledgeSecurityComplianceAudit logs
▤ Real-time context engine
HONEST MEASUREMENT
Judge the inbox by what stayed answered.
No invented benchmarks here — these are the four questions your own dashboard should answer within the first month.
First-reply time
How long does a customer wait for a complete answer once Lumi holds the inbox — minutes instead of the hours a queue imposes?
Auto-send share
What portion of your email clears the confidence gate without review — and how does it grow as categories earn autonomy?
Draft acceptance
How often does your team send Lumi’s drafts unchanged? Unedited sends are the honest measure of reply quality.
Stay-closed rate
Do threads actually end after Lumi’s reply, or reopen with the follow-up questions a partial answer invites?
FROM THE CASE-STUDY SHELF
See email working inside a complete front office.
In the FormyTax implementation, email is one of five client channels feeding the same orchestration layer — including a secure document-request workflow with receipt events and missing-item follow-up, the classic inbox chase. It’s a worked implementation on a verified multi-channel foundation, documented without invented revenue or time-saved numbers — the same no-invented-benchmarks stance this page runs on.
AI EMAIL AGENT FAQ
Direct answers before you hand over an inbox.
What makes an AI email agent different from auto-replies and templates?
An auto-reply acknowledges. A template guesses. An AI email agent for customer service reads the actual question, looks up the actual customer in your systems, and writes a reply about their specific order, appointment, or account. If the email asks three things, the reply answers three things — templates can’t do that, and canned macros answer the wrong question as often as the right one.
How does LumiTalk keep answers accurate — what stops it from making things up?
Two controls. First, grounding: replies are composed from records Lumi actually fetched — the order in Shopify, the slot on the calendar, the invoice in your billing system — plus your approved policy content, never from general model knowledge about your business. Second, the confidence gate: every reply is scored before it can leave, and anything that fails a check on identity, data, or policy is demoted from auto-send to a human-reviewed draft.
When exactly does a human review a reply?
Whenever the confidence score lands below the threshold you set, and always for categories you mark as review-required — refunds over a limit, cancellations, legal or complaint language, anything touching a contract. The reviewed email arrives as a finished draft with the uncertainty flagged, so review means reading and clicking send, not writing from scratch.
Can we run it in draft-only mode first?
Yes, and most teams should. Set the threshold so nothing auto-sends, and Lumi drafts every reply for a week or two. Your team sends, edits, or rejects each one — and that record shows you exactly which categories are safe to automate before a single email goes out on its own.
Will replies sound like our brand?
Lumi writes from your tone rules and example replies — greeting style, sign-off, formality, what you never say. Draft-only mode is also a voice rehearsal: you see and correct the tone on real emails before any customer does.
What happens when one email asks three different questions?
One consolidated reply that answers all three. Lumi parses every distinct request in the message — a shipping question, a size swap, a billing doubt — resolves each against your systems, and composes a single email that closes them all, instead of answering the first question and ignoring the rest.
Does it reply in the customer’s language?
Lumi detects the language of the incoming email and can reply in kind, keeping your tone rules. You choose which languages are eligible for auto-send and which should always route to a draft for review.
How does it handle long threads and back-and-forth?
The thread is the memory. Lumi reads the full history before replying, so it never re-asks what the customer already said, and follow-ups reference earlier answers instead of starting over. Every sent reply and reviewed draft stays on the same thread and is written back to your CRM.
What does setup involve?
Connect the inbox — Gmail or any IMAP mailbox, including shared addresses like support@. Connect the systems replies should draw from: your commerce platform, CRM, calendar, or billing tool. Set the confidence threshold and review-required categories, load your tone rules, and start in draft-only mode. No change to your existing help desk is required.
How is LumiTalk priced?
There’s a free tier to start, a 14-day trial of the full platform, and a flat monthly fee after that — not per resolved email. Your cost doesn’t climb because a promotion doubled your inbox volume, and there’s no incentive misalignment where the vendor earns more when your customers write in more. Contact us for the plan shaped to your inbox volume and systems.
ONE INBOX · DRAFT-ONLY FIRST
Point Lumi at your busiest inbox and let the drafts make the case.
Connect the mailbox, set the gate to draft-everything, and watch a week of replies. Turn on auto-send only for the categories that earn it — that’s the whole rollout.
From the blogAll articles
AI ReceptionistWhat Is an AI Receptionist? A Plain-English Guide7 min read 
Real EstateSpeed to Lead: Why the First 5 Minutes Decide the Deal8 min read 
SubscriptionAI Customer Service for Subscription Ecommerce8 min read 
Property ManagementAI Assistant for Property Managers: A Buyer’s Guide8 min read 
Med SpaHow to Reduce Med Spa No-Shows Without Nagging Clients8 min read 
Estate PlanningAnswering Service for Estate Planning Attorneys8 min read