Ataski
Customer

Customer · Support Reply Drafter

A ready-to-send support reply, quoting your own help docs.

Every inbound ticket comes back as a drafted reply grounded in a specific article from your help docs — never an invented answer. A second AI from a different vendor re-reads each draft before you see it. You approve every send. It will not authorize a refund, admit fault, or promise anything your help docs do not say — refund, cancel, legal, and angry tickets go straight to a human.

Try a sample reply See pricing Help-doc grounded · 2-vendor AI review · from $199/mo, 14-day trial

From inbound ticket to a draft in your approval inbox

01

Ticket lands

From Postmark, Zendesk, or Intercom.

02

Classify & ground

Tagged, then matched to a help-doc article.

03

Draft & review

A second AI re-reads every draft.

04

You approve

One click to send, edit, reject, or escalate.

No help-doc article matches the question? The drafter refuses rather than invents an answer — the ticket lands in your inbox with a one-line reason. Refund, cancel, legal, and angry tickets never reach a draft; they go straight to a human.

The contract

What it runs on its own, what it checks with you, what it won't touch.

Runs on its own

  • Classifies every ticket — topic, urgency, sentiment — as it reads it.
  • Searches your help-doc library and grounds the reply in the closest-matching article.
  • Refuses to draft when no article answers the question — never stitches together a half-grounded reply.
  • Quotes the verbatim help-doc sentence behind every factual claim, so each one is checkable.
  • Runs every draft through a second, independent AI from a different vendor before you see it.
  • Detects and redacts leaked PII — emails, account IDs, payment data — before the draft reaches your inbox.

Checks with you first

  • Every draft lands in your approval inbox — one click to send, edit, reject, or escalate. Nothing sends on its own.
  • Billing questions only when the worker is confident; ambiguous billing tickets are escalated to a human instead of guessed.
  • Suggests a help-doc update when the article you used was missing the answer — but never edits your help docs itself.
  • Summarizes a long ticket thread for a teammate taking it over, marked as an internal-only note.

Won't touch

  • Authorize a refund, credit, chargeback, or any account-level write — those always escalate to a human.
  • Admit liability or legal fault, or give legal or regulatory advice — GDPR, deletion, DPA, liability all route to a human.
  • Promise a pricing change, discount, SLA, or a feature that is not already documented in your help docs.
  • Refund, cancel, legal, and angry or distressed tickets are hard-escalated straight to a human — they never reach a draft.
Full capability matrix from the role registry

PROFICIENT

  • Classify an inbound ticket by topic and urgency. Uses your taxonomy; flags low-confidence classifications for review.
  • Draft a first-response reply grounded in a knowledge-base article. Cites the article it pulled from; refuses to answer when no article matches.
  • Ask the customer clarifying questions when the request is ambiguous. Frames questions one at a time, in plain language.
  • Detect and route urgent or sensitive tickets to a human. Outage reports, threats, and emotional language always route to a human.

ASSISTED

  • Suggest a knowledge-base update when a recurring gap shows up. Owner reviews the suggestion; the AI does not edit the KB directly.
  • Summarize a long ticket thread for the human who is taking it over. Summary is posted as an internal note, not sent to the customer.

REFUSED

  • Issue refunds, credits, or any account-level changes. Always escalates to the human owner with the proposed action.
  • Make commercial commitments on behalf of the company. Pricing exceptions, SLAs, and feature promises route to a human.
  • Provide legal or regulatory advice. GDPR / data-deletion / liability questions go to counsel.

What you get

This is what one card in your approval inbox looks like.

Every draft carries the matched help-doc article, the verbatim sentence each claim is grounded in, and the second-AI verdict. Nothing sends until you click.

Approval inbox · ticket #4,521

2nd-AI review: approved

Topic

billing.invoice

Urgency · sentiment

Medium · neutral

Grounded in

art_283 — “Where to find past invoices”

Confidence

High

Subject

Re: Downloading my last invoice

Hi Priya — you can download any past invoice yourself: open Settings → Billing → Invoice history and click the date of the invoice you need. Each one downloads as a PDF with the full payment breakdown. Let me know if a specific month isn't showing and I'll take a closer look.

Claim grounded in art_283: “Past invoices are available under Settings → Billing → Invoice history and download as a PDF.”

Send Edit then send Reject Escalate Nothing sends until you click.

What the second AI checked

Cleared

A different vendor, different blind spots. One vendor cannot reliably grade whether its own output stayed grounded.

  • Every factual sentence is supported by the cited help-doc article — no hallucination.
  • The quoted help-doc sentence is a verbatim match — no fabricated citation.
  • No refund, pricing, or commercial commitment leaked into the reply body.
  • No customer PII echoed in the subject or body.
  • Tone matches a calm, neutral ticket — not over-formal, not dismissive.

A draft that fails any check lands flagged with the reason. Never silently dropped, never silently sent. A draft that quotes the help docs incorrectly is forced back to “no match” before it ever reaches you.

Pricing

Four tiers by daily ticket volume.

Every tier gets the same drafting engine — help-doc-grounded replies, the second-vendor AI review, PII redaction, and the hard refusal on refund, cancel, and legal. 14-day trial on every tier; no credit card to start.

What it replaces

Doing this in-house means a full salary, benefits, and months of ramp — for one function. Ataski ships the output from day one at a flat, predictable monthly price.

Enterprise-grade quality for a fraction of a hire. Scale up or pause anytime — you only pay for the months you use.

Solo

$129 / month

50 tickets / day · hard cap

≈ $0.09 per ticket

  • Help-doc-grounded drafts — no article, no draft
  • Second, independent AI review from a different vendor
  • Hard refusal on refund / cancel / legal

Founder plus one support generalist — the first paid plan above the trial.

Choose Solo
Most picked

Team

$1,299 / month

200 tickets / day · hard cap

≈ $0.22 per ticket

  • Everything in Solo · higher daily cap
  • PII redaction on every draft
  • Approval inbox, every send

Seed-stage SaaS with a first dedicated support hire.

Choose Team

Scale

$2,999 / month

500 tickets / day · hard cap

≈ $0.20 per ticket

  • Everything in Team
  • Multi-mailbox routing
  • Audit log on every classify / draft / review

Series A with a dedicated customer-support team.

Choose Scale

Business · $5,999/mo · 2,000 tickets / day · Enterprise — above 2,000 tickets/day

For volumes past the Business cap, or when procurement wants per-resolution pricing instead of a daily cap. Custom SLAs, SAML, bring-your-own LLM keys, regional data residency, and a dedicated customer-success manager. Sold inbound.

  • · Everything in Scale
  • · SSO · audit-log export
  • · DPA on request
Talk to sales

Daily caps are hard. When your account hits the daily ticket cap, the next inbound ticket is held until the counter resets at 00:00 UTC. Upgrade a tier or contact sales for headroom.

Setup

About fifteen minutes from signup to your first drafted reply.

  1. 01

    Connect your helpdesk. Postmark inbound works out of the box; for Zendesk or Intercom, paste the inbound-webhook secret.

  2. 02

    Load your help docs. One article at a time, a CSV, a public help-doc URL, or a Notion Personal Integration Token.

  3. 03

    Turn the drafter on. A single toggle. Every inbound ticket then runs the pipeline.

  4. 04

    Review your inbox. About a minute after a ticket arrives, the draft is ready — send, edit, reject, or escalate.

Under the hood

Every model and service we use, named.

Drafts the reply
A drafting AI classifies the ticket and drafts from the matched help-doc article — a lighter model on entry tiers
Reviews the draft
A second, independent AI from a different vendor — on purpose; it re-reads every draft for hallucination, citation mismatch, scope leaks, and tone
Grounds the answer
OpenAI text-embedding-3-small — your articles in a pgvector index isolated to your workspace; no match, no draft
Loads help docs
Single-article form, CSV upload, public URL, or a Notion Personal Integration Token — Notion syncs on a button
Safety
Customer text wrapped as untrusted content, a verbatim-citation post-validator, and a two-layer PII pass before any draft reaches you
Privacy
Every classify / draft / review call in an append-only audit log with cost and latency; 30-day data deletion on offboarding

Support replies your team can send as-is — grounded, reviewed, safe by default.

About fifteen minutes to wire up your helpdesk and help docs. Every draft is quoted from your own articles and re-read by a second AI before it reaches you. You approve every send.