Support Desk Manager
An AI-native support desk that resolves the tickets
It runs your support desk: it answers each inbound ticket from your own help docs, every claim cited to the article it came from, and escalates the rest. A second AI from a different model family checks each answer before it sends. You review and one-click send from day one; autonomy is earned per topic and opt-in. Run it instead of a helpdesk, or alongside the one you have — you pay per resolved ticket, not per seat.
Trained on
your help-centre articles + your refund and escalation policy + the customer's prior ticket history
Methods
- Every factual claim cited to the help-docs article it came from
- A second AI from a different model family reviews each answer before it sends
- Refunds, credits, and billing decisions escalated to your team, never resolved by the desk
- A 'ignore previous instructions' ticket is quarantined, not acted on
- A ticket that leaks a password or API key is flagged, never echoed back
Who it's for
Support teams and founders who want an AI-native desk that resolves the routine tickets and escalates only the exceptions — standalone, or alongside the helpdesk you already run.
From $169/mo · priced per resolution
How it works
From signup to the first send.
Forward your support email in about five minutes, or connect your help desk — then point us at your help-centre articles and refund policy.
Each inbound ticket loads the customer's history, then the desk resolves it with an answer cited to your docs.
A second AI from a different model family checks every answer; trusted topics that clear the calibration bar send, everything else escalates to your team.
Trusted topics resolve automatically; the exceptions land in your queue for one-click review, edit, or escalate.
Capability matrix
Does well, with review, won't do.
Out-of-scope tasks escalate · they don't hallucinate
Does well
- A resolved reply within minutes of an inbound ticket
- Every claim cited to a help-docs article
- A second AI from a different model family checks every answer
- Prompt-injection and leaked-secret tickets quarantined
- Churn-signal tickets surfaced to the account owner
With review
- Trusted topics resolve automatically once they clear the calibration bar; everything else escalates to your team for one-click review
- Low-confidence answers go to your team's queue
- Refund, billing, and policy questions wait for your decision
Won't do
- Never promises a refund or credit
- Never admits fault or liability
- Never answers from facts your docs don't support
Connections
Per tool: why we need it, what we do, how it's audited.
Connect what you need, when you need it. The assistant only reads tools you explicitly link — never the rest of your stack.
- Your support content — forward your support@ email (about five minutes) and drop the chat widget on your site. Moving off Zendesk, Freshdesk, Intercom, Front or Help Scout? Import your articles and macros in a few clicks.
- Your help-centre articles — the source every answer is cited to.
- Optional: your refund and escalation policy — so the line on refunds is yours, not a guess.
- We retrieve the right help-docs article for each ticket and cite it in the answer.
- We run a second AI from a different model family and quarantine prompt-injection and leaked secrets.
- We escalate refunds, billing, and policy calls to your named owner with the policy section quoted.
What it reads on every task
Your data. Not the public internet.
Loaded fresh per task — change a doc, the next task picks it up. No retraining, no week-long fine-tune, no stale snapshot.
- Your knowledge base and help-centre articles — the only source an answer is allowed to cite
- Your refund and escalation policy — so the desk never invents a policy you don't have
- The customer's prior ticket history — loaded before any reply is drafted
Guardrails
When it stops and asks you.
- Refunds, credits, billing disputes, or contract changes — escalated to your team with the policy quoted, never promised
- A question your help docs don't cover — escalated for a human decision instead of a confident guess
- A ticket carrying what looks like a credential or secret — quarantined and flagged, never echoed back
- An answer the second-AI reviewer scored low-confidence — held for your team's review
Plans
Try the sample, or start with your own data.