A HUMAN REVIEWS EVERY REPLY
Your queue arrives drafted. Your team still sends every word.
Overdesk sits on top of Help Scout or Zendesk and learns from the tickets your team has already resolved. New tickets show up triaged, with a draft reply ready to review. You read it, you send it.
THE MODEL NEVER SEES YOUR CONTACTS
Emails and phone numbers are swapped for tokens before any prompt leaves Overdesk.
In the ticket
jane.doe@acme.com
+1 (555) 203-8841
+1 (555) 203-8841
Sent to the model
jane.doe@acme.com[[email_3f2a]]
+1 (555) 203-8841[[phone_8b1c]]
+1 (555) 203-8841[[phone_8b1c]]
The model writes the reply against the tokens. Overdesk restores the real details in the draft your agent reads, so the reply is usable and the raw contact info never left your workspace.
Invoice shows the old seat count after downgradeBILLINGNORMAL✓ APPROVED · SENT 08:32
#4821CSV export times out on workspaces over 10k rowsEXPORTSHIGH● DRAFT READY
Hi Priya, this timeout happens when an export crosses our 10k-row streaming limit. The fix on our side shipped last Tuesday; your workspace needs the async export flag, which I’ve just enabled. Re-run the export and it will land in your email within a few minutes. If it doesn’t, reply here and I’ll dig into the job logs directly.
Angry: third failed export this month, mentions leavingURGENT◆ NEEDS A HUMAN
01 · THE LOOP
Connect it, and your queue starts arriving drafted.
STEP 1 · CONNECT
Connect your helpdesk
Link Help Scout or Zendesk in a few minutes. No migration and no downtime. Overdesk reads and writes through the tools you already use.
STEP 2 · LEARN
It learns from your history
Your resolved conversations, your docs, and your customer context become the knowledge base, so drafts match how your team actually answers instead of guessing.
STEP 3 · DRAFT
Drafts and triage appear
Every new ticket gets sorted by product, type, and urgency, then arrives with a draft written from how your team handled the last one like it.
STEP 4 · SEND · YOURS
Your team approves and sends
Review, edit if you want, send. The reply is yours. Overdesk just did the first ninety percent.
02 · WHAT IT DOES
Drafting is the first job. Overdesk also reads the whole queue and briefs the people above it.
PILLAR 1 · THE AGENT
Drafts that sound like your team, because they came from it
Every draft is retrieved from your own resolved tickets and docs, confidence-scored, and posted for review. Overnight batches mean mornings start drafted; stale tickets get follow-up nudges instead of silence.
DRAFTSwritten from your resolutions, with the context they drew on attached
CONFIDENCEscored per draft from the history it found; a thin score says look closer
FOLLOW-UPSstale conversations surface with a nudge drafted and waiting
TRIAGEproduct, type, and urgency, applied across your open queue on a schedule
PILLAR 2 · PATTERNS & COACHING
Sees the whole queue, not one ticket at a time
Spots problems spreading across conversations before they become an incident thread, and scores response quality so coaching is specific instead of vibes.
▲ 6 tickets mention export timeouts in 48h: flagged as a pattern, before ticket 7
▲ response quality on refund threads dipped this week: two examples attached for coaching
▲ response quality on refund threads dipped this week: two examples attached for coaching
PILLAR 3 · LEADERSHIP REPORTING
Briefings that are written, not exported
Morning briefings, signal digests, customer-health reports. Overdesk writes them from the queue and pushes them to Slack or ClickUp, so leadership reads prose instead of squinting at a dashboard.
“Export timeouts drove 40% of yesterday’s volume; the fix shipped, expect residuals through Thursday. Two at-risk accounts need a call, not a ticket…”
03 · SAFETY, AS ARCHITECTURE
Built for people who can’t afford an AI mistake in front of a customer.
Three mechanisms sit behind that promise. Here’s each one, and how you check it.
security & data handling: what leaves, where it goes, and what our redaction misses →
MECHANISM 1
A human sends. Always.
Drafts land unsent in your helpdesk, and sending is a person's call. There's no auto-send mode, and no code path that lets Overdesk message a customer on its own.
MECHANISM 2
Contact details become pseudonyms
Emails, phone numbers, and other contact identifiers are swapped for deterministic tokens before conversation text is stored or sent to a model. Detection is pattern-based, so it has limits: a name written in free text can still slip through.
MECHANISM 3
Every draft is on the record
Which model wrote it, what context it had, and what it cost, all stored per draft and queryable. API-key and webhook changes are audit-logged separately.
WHAT IT WILL NOT DO · ON THE RECORD
✕Send a customer reply on its own. There's no setting for it; the approval step is the workflow.
✕Hide how thin the evidence was. Every draft carries a confidence score from the history and docs it found, and a low one says so.
✕Replace your helpdesk, or talk to customers in a chat widget. It works inside the tools you already trust.
04 · WHERE IT RUNS
Works with Help Scout, Zendesk, and Intercom.
Help ScoutLIVE
Full two-way integration: conversations, customers, tags, and internal notes.
ZendeskLIVE
Drafting, triage, and follow-ups on top of your existing Zendesk queue.
IntercomLIVE
Drafting, triage, and follow-ups on top of your existing Intercom inbox.
Slack
Briefings, pattern alerts, and review notifications, plus Ask Overdesk, the assistant that answers questions about your live queue right in Slack.
ClickUp
Scheduled reports and briefings delivered to your ClickUp chat channels.
THE SUCCESS TAX
Everyone else bills per resolution. Your bill doesn’t move.
Per-resolution pricing grows your invoice as your AI improves. Overdesk is priced per seat, so the number stays put.
See the math →per-resolutionOverdesk
05 · THE RECORD
The record
0
ways to send a reply without a human
No auto-send setting today, no autonomous mode, and no background job that can reach a customer: the drafting workers can only leave a draft unsent in your helpdesk. A person approves every send, or nothing goes out. We didn’t build the alternative.
EVERY DRAFTcarries a confidence score built from the history and docs it found
ON THE RECORDthe model that wrote it, the context it had, and what it cost
YOUR QUEUEtriaged by product, type, and urgency on a schedule you control
See this afternoon’s queue arrive drafted.
Connect your helpdesk, let Overdesk read your resolved history, and review your first drafts today. Approve the good ones. That’s the pitch.
Does it ever send without approval?
No. Drafts are placed in your helpdesk unsent, and a person approves every send. There is no auto-send setting to switch on today.
Does it replace my helpdesk?
No. It sits on top of Help Scout or Zendesk. Your queue, your workflows, your tool, with drafts, triage, and reports layered on top.
What happens to customer data?
Drafts run on your own Anthropic API key, under your account. Your customers' contact details are swapped for tokens before anything is stored or sent to a model, and every draft keeps its own record: the model that wrote it, the context it had, and the cost.