The case arrives
Email, chat or a social message. Sanddo identifies the language, the case type and which store the customer belongs to.
Every case category has its own confidence score, built from your team's actual approvals and corrections. You set the threshold for when a category may send on its own — and you can withdraw it again at any time.
Four steps, every time. Sanddo does not just write a reply — it fetches the data, prepares the action and checks itself before anything leaves the building.
Email, chat or a social message. Sanddo identifies the language, the case type and which store the customer belongs to.
Order, shipment, return status, subscription and earlier contact are looked up in the connected systems — while the reply is being written.
Sanddo writes the reply and prepares the actual action: address change, return label, refund or subscription change.
If the category is above your threshold, the reply is sent and the action is executed. If it is not, the case lands with a colleague with everything prepared.
That is the right question to ask. A system allowed to refund money and change orders has to answer it — not with a promise, but with a lock.
“What if it starts refunding on its own?”
Refunds are a category like any other. It starts in Draft and can only be moved to Auto if you move it — and only when the score is above the threshold you set. Many teams leave it in Approve permanently.
“What if it answers an angry customer badly?”
Cases with doubt, warnings or escalation are held back regardless of category and mode. Autopilot covers the predictable, not the exceptions.
“What if it hands data to the wrong person?”
Customer and order details are released only when the sender can be securely matched to the order. If they cannot be, Sanddo asks for verification instead of guessing.
Every case category sits in exactly one mode at a time. You move them as trust grows — and back again if something changes.
The reply is ready as a draft with the sources visible. Your colleague edits and sends. Every correction counts towards the category's confidence score.
Both the reply and the action are ready. Your colleague reads through and approves both at once. Approvals without corrections pull the score up.
The reply is sent, the action is executed in the right system, and the case is closed. Everything lands in the audit trail, and doubtful cases are still held back.
Drag the threshold and see which categories become ready for autopilot. The numbers below are a typical picture from a Danish webshop — yours come from your own cases.
A category can only be set to Auto once its confidence score is above the threshold. Below the threshold the button is locked — for you too.
0 of 10 categories run on autopilot at the selected threshold.
You do not flip a switch and hope for the best. You start with Sanddo helping people, look at the numbers, and open one category at a time.
At setup Sanddo reads your earlier cases and proposes the categories your customers actually write about. You adjust them, merge them and name them the way you talk about them.
No category begins on autopilot. Sanddo writes drafts, your team corrects and sends. The system measures how much actually gets corrected in each individual category.
When a category has enough approvals without corrections, it clears your threshold and can be moved to Auto. The decision is still yours — the threshold just makes it documented.
Whatever the mode and threshold, there are cases Sanddo does not send. They land with a colleague with all the context gathered.
If the sender cannot be linked to the order, neither customer nor order details are released. Sanddo asks for verification.
If Sanddo meets a case it has no rule for, it goes on Hold. It does not guess and it does not send.
The same action cannot be approved and executed twice. A refund is not sent twice, whatever happens in the case.
Live chat being allowed to answer product questions does not mean it may touch orders. Every channel has its own limits.
The effect does not come from Sanddo replying faster. It comes from the case being resolved in the same moment — including the action in your systems.
We run them through live and show which category they land in, what confidence score they would get, and what it takes before exactly those cases can run on autopilot at your company.
94% confidence · 891 cases per mo.
Above the threshold · runs on autopilot
46% confidence · 167 cases per mo.
Below the threshold · held for approval