Run it in shadow mode now. Trust it by Black Friday.
Nobody switches on an untested AI in the middle of peak week. Glassdesk drafts replies to your real tickets without sending a single one, so you read two weeks of what it would have said before a customer ever sees it. Start by 30 October and you decide with evidence, not a demo.
until Fri 30 October, the last start date that leaves two weeks of shadow drafts before peak
- Gorgias App Store
- Freshdesk Freshworks Marketplace
- Shopify order connection
- Shopify App Storein review
Test it while a wrong draft costs nothing.
Two things are true every November. Ticket volume triples for a week, mostly “where is my order”, returns and the same five policy questions. And any tool priced per resolution bills hardest in exactly that week, because it charges you for working. So test the AI on real tickets in October and go into peak already knowing what it does with your customers.
Every draft shows its sources
The help-centre passages it used, identified, and a confidence score from 0 to 1. A wrong draft traces back to a thin article you can fix, not a model you have to guess about.
Low confidence goes to a human
One threshold per store, plus a list of categories that always escalate whatever the score: refund exceptions, legal, abusive messages, clearly frustrated customers. It never guesses. That is why the published deflection rate is 65%, not 90.
“Where is my order?” from the real order
The peak-week ticket a help centre cannot answer. Connect Shopify, read-only, and drafts cite the actual order: payment status, fulfilment, tracking number, carrier link. Found by order number or the customer’s email. Live since July 2026. How it works.
Fourteen days of drafts, then a decision.
Install and point it at your help centre
Install on Gorgias or connect Freshdesk, paste your public help-centre URL, and Glassdesk crawls it into citable chunks. Before anything goes live it runs its published eval suite against your content so you see an accuracy score first. Optionally connect Shopify for real order-status answers.
Every ticket gets a private draft, with sources
Each draft lands as an internal note on the ticket, with the help-centre passages it cited and its confidence score. Read the escalations first. They are the most useful output, because each one names why it refused to answer.
Set the gate and the always-escalate list
One confidence threshold per store, plus the categories that always go to a human. Refund exceptions, legal, anything touching money: a two-minute setup. Fix the help-centre articles the citations show are thin. That is usually where wrong drafts come from.
Flip auto-send, or keep it in shadow through peak
If the drafts are good, switch auto-send on before the freeze and let it answer Tier-1 while your team handles the rest. If you want more evidence, leave it in shadow mode through Black Friday and decide in January with peak-week data nobody else will have.
No engineers. One afternoon.
Install from the Gorgias App Store
- In Gorgias, open Settings → Apps → App store and install Glassdesk.
- Gorgias shows exactly which permissions you grant and makes the connection. You never paste an API key.
- Paste your help-centre URL in Glassdesk. Shadow mode is already on.
Connect in a few minutes
- Enter your Freshdesk subdomain and API key on the connect page.
- Add the two automation rules the page gives you to paste.
- Paste your help-centre URL. Drafts appear as private notes on new tickets.
Real order status, read-only
- From the Glassdesk dashboard, click Connect Shopify and approve on Shopify’s own screen.
- “Where is my order” drafts then cite the real order: fulfilment status, carrier and tracking link.
- Read-only. Disconnect any time and access ends immediately.
Your help centre only needs to be publicly reachable. Gorgias help centres, Shopify pages and policy URLs, Zendesk-hosted centres and custom sites all crawl. A help centre rendered entirely by a JavaScript widget crawls badly; if yours is one of those, email us the URL first and we will tell you before you spend any time on a trial.
One flat price. The busy week costs the same.
No per-ticket or per-resolution fee, so the week your volume triples costs the same as the week before it. Most of this category charges $0.90 to $1.50 per resolution, a bill that grows with your best week. Put your November volume into the calculator.
Starter
- Up to 250 AI resolutions a month, then a hard cap and an upgrade prompt, never a surprise bill
- Draft and auto-send reply modes
- One knowledge-base source
- Transparency panel on every answer
Growth
Most popular- Up to 1,000 AI resolutions a month, hard cap, no overage
- Everything in Starter
- Multiple knowledge-base sources
- Eval-before-live accuracy report and shadow-mode rollout controls
Scale
- 3,000 AI resolutions a month included
- Then – per resolution, disclosed here, not discovered on an invoice
- Everything in Growth
- Multi-brand support, analytics export, priority support
FOUNDING10, in exchange for honest feedback. Ten seats, then the code stops working.
Questions stores ask before peak
Does shadow mode send anything to my customers?
Why is 30 October the last useful start date?
Does the price go up in November?
What happens to tickets it is not sure about?
Can I set a different threshold for different ticket types?
Do I need an engineer?
Two weeks of drafts beats any demo
Install today, read what it would have said on your real tickets, and go into Black Friday knowing. Flat price, nothing sent until you say so.