Seismisk

Drop Customer Support workload by 74% with AI

After 2 months of using this custom AI tool, the number of inbound emails customer support reps need to engage in drops 4x. The rest is done by a robot who learns.

How it works

Choose an example (or write your own)
Received email preview
6 of 6 done
1

Redacted

Before sending to the LLM, sensitive data is placed into a safe
The safe
[PERSON_1]Owen Abernathychild
[PERSON_2]Kristen Abernathyparent
[STAFF_1]Mr. Delgadoteacher
[GRADE_1]grade 4grade
Fromkabernathy.home@example.comsender, finds the family

The safe stays on this server.

The email gets redacted
Hi there, [PERSON_1] ([GRADE_1] Spanish, Tuesdays 4:30 at North) was out sick this week so he missed class. Is there a way for him to make it up? … He really likes [STAFF_1].
Thanks so much, [PERSON_2]
2

Classified

A quick LLM decides what topic is being asked, and routes it to the respective knowledge base
academic_placementattendancebilling_disputebilling_paymentsclass_operationsclass_transferenrollmentevaluationshomework_academiconboardingotherportal_technicalschedule_timezoneshipping_logisticsstaff_internalsummer_programsteacher_outreachwithdrawals_refunds
→ Knowledge base: Homework and attendance · standard reply · 92%
Missed one class through illness. Asks how to make it up and who tells the teacher.
3

Rule applied

The school rule behind the reply.
Branch: standard reply · needs CRM and calendar
P1 A missed class can be made up in any same-grade section at any location within 30 days; two make-ups per subject per semester; book through the desk, not the teacher.
FAQ used: make-ups
4

Looked up

What it checked, and what it kept.
CRMLooked up by the sender's address
ChildOwen Abernathykept here, filled in at the end
Gradegrade 4 Spanishkept here, filled in at the end
HomeNorth, Tuesday 4:30sent to the model
Make-ups0 of 2 usedsent to the model
Matched here by the address, never sent. Unknown sender: the model gets only the public policies and timetable.
CalendarOpen grade 4 Spanish sections, next 30 days, not Saturday
Mon 4:30Westgate, 2 seatssent to the model
Wed 4:30Downtown, 1 seatsent to the model
Thu 5:00North, 2 seatssent to the model
BillingNot needed for this email
PoliciesP1 Make-ups, P2 Absences
See step 3.
5

Drafted

All the model wrote. No names.
Hi [PERSON_2],
Glad [PERSON_1] is better. He can make up Tuesday's class in another [GRADE_1] Spanish section: Westgate on Monday 4:30, Downtown on Wednesday 4:30 or North on Thursday 5:00. Which one suits? The desk books it and lets [STAFF_1] know.
The Lanternwood team
6

Rehydrated

Names back in. What the parent reads.
Hi Kristen,
Glad Owen is better. He can make up Tuesday's class in another grade 4 Spanish section: Westgate on Monday 4:30, Downtown on Wednesday 4:30 or North on Thursday 5:00. Which one suits? The desk books it and lets Mr. Delgado know.
The Lanternwood team
DraftsThe draft appears in your Outlook or Gmail, ready for sending.

Nothing is stored. The model saw step 1, the public policies and the facts marked sent in step 4. Nothing else.

Built on 8,075 real email threads.

87–91%sorted right
74%workload reduced
48%sendable as-is one week after launch

A person approves every draft.

What the assistant knows.

A made-up school with four connections: calendar, CRM, billing, policies. Yours replace them.

Bring the email you were told not to paste into ChatGPT.