How we work
The drafts get better because your operators keep editing them, and we read those edits. Here is the loop, and an illustrative example of what one round of it changes.
When an operator opens a Sensbot draft, rewrites two sentences and sends it, that edit is the most useful piece of data in the system. It is a corrected answer, written by someone who knows the store, on a real customer case, under time pressure. Nobody had to be asked for it.
Sensbot matches each sent reply back to the draft it started from and records the difference. Once a week those differences are clustered into patterns: a knowledge gap, a missing policy, a style preference, a case that was routed to the wrong kind of answer in the first place.
That is the whole input. We do not tune on opinions about how the assistant ought to sound. We tune on what your team actually changed before hitting send.
The loop
Example, anonymized
The example below is invented for this page to show the shape of a tuning round. It is not a customer transcript, and no customer data appears anywhere on this site.
The store's return window is 30 days from delivery and items have to come back unworn. The draft committed to a refund without mentioning either, which is the kind of reply that turns into a dispute three weeks later. The operator rewrote the middle of it before sending.
The same correction appeared several times in one week, across different customers and different operators. That is a pattern, and it points at the store's policy configuration rather than at the writing style or the knowledge base. The fix was one policy stating the return window, the condition requirement, and what to say when the deadline has already passed.
Example, anonymized
Before the change went live it was checked twice: against the cases that triggered it, and against return questions that were already being answered correctly, so that a store with a longer window in one market does not inherit the wrong deadline.
Tuning is not a one-off. Stores change policies, run promotions and open markets, and an assistant that was accurate in March drifts behind them by June. The weekly loop exists because the target moves.
It is also not magic. Some patterns turn out to be a knowledge gap only you can fill, and in that case we come back to you with the specific question rather than inventing an answer. Some cases should never be answered automatically at all, and the right outcome there is a draft that says a human needs to take this one.
And we do not act on a single instance. One edit is one operator's preference on one email. We record it and wait to see whether it repeats, because a store's published policy should not be overwritten by one busy Tuesday.
Not for the weekly loop itself. It runs on the drafts and sent replies that already exist. You get involved when a pattern turns out to be a genuine knowledge gap, where we come back with a specific question rather than guessing at the answer.
Changes are verified against your own past cases first and applied deliberately, with a record of what changed and the ability to roll it back. Nothing is rewritten wholesale: a new rule is merged into what you already have, so your wording and your specific numbers survive.
It learns your store patterns from your own past support emails during setup, and from the corrections your operators make afterwards. The strictest kind of rule, the sort that must always be followed, is not promoted automatically. Those are verified before they go live.
That is what the second verification pass is for: the change is run against cases that were already correct before it is applied, and if those regress it does not ship. Applied changes keep their test cases, so later changes are re-checked against them too.
The loop runs for every Email Assistant. Hands-on tuning, where we do the work with you and keep doing it, is what the done-for-you plan is, from €390 a month. See pricing.
The loop starts the moment your operators begin editing drafts, which is the first week.
Free trial on every platform. No credit card on Shopify, Shoper or Upgates.