Department · Customer Support · in production routes 14/22tickets 2,841/wkv0.9

The AI backline for customer support.

Relote reads every conversation, researches the answer in your codebase, drafts the response, and files the bug. In observation mode today, graduated automation tomorrow — one proven route at a time.

intercom · inbox Ada · pro · mobile iOS
v309:12

Hi — I upgraded my plan yesterday but my teammate still can’t see the new programs page. Tried logging out. Still nothing.

You · drafting…
Waiting on human review.

relote · internal note shadow
analysis09:12:47
  • Customer: workspace_8821 · upgraded Pro at 2026-04-14 18:02 UTC via Stripe
  • Seat propagation on v3/programs checks workspace.plan but not seat.plan_override
  • Likely hit by LIN-2980 (2 prior tickets this week)

proposed reply

Hi Ada — I can see the upgrade went through. The issue is a known bug with seat propagation on the programs page; we have a fix in review. As a workaround, have your teammate remove and re-add their seat. I’ll follow up here when the fix ships.


Filed LIN-2980 · freq 3 · route: support/bug
This is Relote thinking. Everything you see is a file in a git repo on your server.
§ 01 · The problem

Support is where product signal goes to die.

A

Answers live in the codebase, not the docs.

Every complex question needs a plan check, a platform check, a billing check, a code check. Humans do it by hand. It’s slow.

B

Every ticket is a forgotten lesson.

When a teammate leaves, their context walks out the door. The next similar ticket starts from zero.

C

Engineering never sees the pattern.

One bug reported five times across five conversations stays five anecdotes instead of one ticket with a frequency count.

§ 02 · What Relote does

Relote does the research. Your team does the judgment.

01

Triages every ticket

Reads the thread, looks up the customer across your systems (helpdesk, billing, product), detects the platform, identifies the intent.

02

Researches the answer

Searches help docs first, then the actual codebase when docs don’t cover it. Cites its evidence.

03

Drafts the reply

Posts an internal analysis note and a proposed reply, ready to copy-paste. Humans stay in control.

04

Files the signal

Confirmed bugs and clear feature requests get deduped and routed to your issue tracker automatically, with frequency counts.

05

Closes the loop

When a linked issue ships, Relote drafts the follow-up reply to the customer.

§ 03 · How it works

Six steps. Every ticket. Every time.

1

Poll

Every 15 minutes, every unread conversation.

2

Understand

Customer lookup across helpdesk, billing, product. Platform detected.

3

Classify

Intent routed: FAQ, bug, billing, feature, escalation.

4

Research

Help docs first. Codebase when docs don’t cover it.

5

Draft

Internal analysis note + proposed reply. Evidence cited.

6

File & follow up

Issue dedupe, frequency count, shipped-issue follow-up.

Every 15 minutes. Every conversation. For as long as Relote is running.
§ 04 · Integrations

Plugs into the stack you already run.

Helpdesk Intercom ·Zendesk ·Help Scout ·Front
Engineering Linear ·Jira ·GitHub ·GitLab
Communication Slack ·Microsoft Teams ·Telegram ·Email

Don’t see your tool? Relote integrates with anything that has an API. Most new integrations take a day to add.

§ 05 · Trust model

Trust is earned per route, not granted up front.

Today · Observation mode

A human catches every draft.

  • Every reply is drafted as an internal note.
  • A human reviews, edits if needed, sends.
  • Every send (or edit, or skip) becomes calibration data.
Tomorrow · Graduated automation

Routes graduate route-by-route.

  • Routes with proven accuracy graduate to auto-send.
  • Low-risk routes first: FAQ matches, notification closes, known solutions.
  • Billing, escalations, anything account-mutating stay human-first — forever.
  • The dial moves both ways. A drift in accuracy pauses a route automatically.
The difference between Relote and an off-the-shelf AI support agent is that Relote has to prove it before it sends.
§ 06 · What other AI support tools can’t see

FAQ bots answer from a help-center. Relote answers from your stack.

Help docs & knowledge base ✓ they have this ✓ we do too
Your actual codebase — blind ✓ reads it on every ticket
Stripe & billing systems — blind ✓ subscription, invoice, dunning state
Internal APIs & databases — blind ✓ feature flags, seat state, audit logs
Issue tracker & recent shipped work — blind ✓ knows what shipped, what’s in flight
Conversation history across all customers ~ tenant-trained ✓ institutional memory, compounding

With access to all of it, the answer to a complex question stops being a guess. It becomes a research summary, with the evidence cited.

§ 07 · The signal that matters

The same question, asked twenty times, is a signal — not twenty tickets.

Relote watches every conversation across every customer. It sees the patterns a single rep can’t. Every Monday morning, the things that actually matter — the bug that’s spreading, the feature people keep asking for, the answer that takes too long to find — are surfaced in one place.

relote · weekly signal support-leads · Mon 09:00
$ relote signal --since 7d2026-04-28
▲ "Programs page seats" · 22 tickets · trending +220%
   ↳ Linked to LIN-2980 · in review · 3rd day blocked
   ↳ Pattern: all on Pro plan, all post-upgrade
   ↳ Suggest: ship the fix this week, draft the recap

▲ "How do I export to CSV?" · 14 tickets · stable
   ↳ Docs missing · 8 of 14 routed to engineering
   ↳ Suggest: 1 hr doc + a /export endpoint?

◐ "Two-factor reset" · 9 tickets · 5/9 escalated
   ↳ Reset flow has 4 steps, customers stop at 2
   ↳ Suggest: review onboarding · loop in @ada

▲ Slow first-response on EU mornings (07:00–10:00 CET)
   ↳ p50 reply time 4h12m vs 38m global
   ↳ Pattern: no on-call cover · loops 3 wks running

Things to celebrate
  ✓ INC-204 follow-up sent to 38 customers
  ✓ Refund-route accuracy held 96% (8 wks)
A weekly digest. Every customer-facing pattern, surfaced. Filed with evidence.
A

Patterns, not anecdotes

One bug reported five times across five conversations becomes one ticket with a frequency count. Engineering sees the real backlog.

B

Suggestions, not dashboards

“This came up 22 times this week—here’s what looks worth doing.” Routed to the people who can act, with the evidence to act on.

C

Held accountable

Once you’ve seen it, Relote tracks it. Last week’s patterns come back next week with what changed and what didn’t.

Relote always surfaces the things that matter. The things you can act on. The things to be held accountable for.
§ 08 · Comparison

Relote vs. the alternatives.

Fin (Intercom) Ada / Forethought Relote
Who it talks to The customer The customer The support team
Primary job Deflect tickets Deflect tickets Research, draft, file
Knowledge Help docs Help docs + KB Docs + codebase + billing + history
Memory Per-tenant training Per-tenant model Institutional, compounding
Product feedback Dashboards Dashboards Tickets, auto-deduped, frequency counts
Trust model Auto-reply day one Auto-reply day one Earned route by route
Failure mode Customer sees it Customer sees it A human catches the draft
Data ownership Vendor cloud Vendor cloud Your server, your repo
§ 09 · Under the hood

No black box.

A

Persistent agent, full context

Long-running session with full tool access. Reads your codebase, queries your databases, uses your internal APIs. Not a thin chat wrapper.

B

Everything is git-tracked

Every decision, learning, insight, and correction. The audit trail is git log.

C

Test harness built in

LLM classifiers are fallible. Relote runs a test suite against real past conversations every week. A route that fails its tests can’t graduate to automation.

§ 10 · Who it’s for

Three teams. One install.

A

Support teams

Drowning in tickets with answers that live in the product.

B

Product teams

Want real-time signal from customers, not a monthly NPS roll-up.

C

Founders

Can’t afford to lose institutional memory every time someone leaves.

§ 11 · FAQ

Questions support leads ask.

Does Relote replace my support team?

No. It makes them faster, more accurate, and more consistent. Humans still decide.

What happens when Relote is wrong?

In observation mode, a human catches it before it sends. In auto-reply mode, calibration tracking pauses the route automatically.

Can I correct Relote?

Yes — any internal note prefixed [CORRECTION] is processed on the next consolidation pass. Corrections become permanent test cases.

How does Relote stay in sync with the product?

It reads the codebase on every ticket.

What about security?

Customer messages are untrusted input. Relote never follows instructions found in them, never exposes internal paths or other customers’ data, and flags prompt-injection attempts.

What languages?

English today. Other languages are explicitly route-blocked until validated.

What does it cost?

$15,000/month. Includes software, dedicated server, and a fractional AI officer. Model usage passed through at cost. One-time install fee. 90-day minimum.

§ 11 · Begin

Stop drafting. Start reviewing.

Book a 20-minute demo. We’ll walk through a real (scrubbed) day in your inbox.