Delivery review manager with AI replies

Background

Delivery reviews are not kept in one place. They sit on 배달의민족, 요기요 and 쿠팡이츠 separately, so checking them means three logins. And reading them is not the end of it. Each one needs a reply.

What we found

The real problem was not the effort of checking but what got missed. Reviews arrive all the time, but people look only now and then, and low ratings go unanswered in between. A negative review without a timely reply stays up for the next customer to read.

Replies were the same story. The same thanks for the same praise, the same apology for the same complaint, written from scratch every time. It was a heavy task for a business owner at the end of the day.

The problem

Collecting reviews in one place was not enough. Even with everything on one screen, not knowing where to start is no better than not seeing it. So we tracked the reviews that must not be missed as their own figure.

Replies called for more care. Drafting a reply is not the same as posting it. Replies go out to customers under the restaurant’s name, so nothing the owner has not read should be posted. Keeping a final human check was a design requirement from the start.

What we built

• AI reply drafts — The draft is written from the review text together with the star rating and the platform. The same words call for a different reply at 5 stars than at 2. Replies are kept under 200 characters, because long replies go unread in delivery apps. The owner reads the draft and posts it with one click, or edits it in place. Writing becomes choosing
• The button appears only on unanswered reviews — Reviews that already have a reply do not show it. The rule used to count what must not be missed applies to the button as well
• Automatic sync across three platforms — Register the accounts once, and reviews are collected
• Missed reviews as numbers — Negative reviews of 3 stars or lower, and those still unanswered, are counted separately. Reply rate against target is shown alongside, so any backlog is visible on screen
• Keyword and sentiment analysis — Frequent phrases are extracted from reviews and grouped as positive, neutral or negative. This shows the reasons that ratings alone do not
• Platform comparison — See at a glance where ratings differ by platform
• Always-on operation — Reviews are collected and drafts prepared without a PC left running

Results

Three separate platforms became one screen, and writing replies became choosing them. Unanswered negative reviews remain visible as a number. Making missed reviews visible on screen is what this system actually changed.

See more work on this topic under AI adoption.