Reviews from three platforms gathered on one screen, and the AI writes the replies first

Background

Delivery reviews do not pile up in one place. They sit separately on 배달의민족, 요기요, and 쿠팡이츠, so checking them means visiting three sites. And checking is not the end. You have to write a reply to each one.

What we saw on site

The problem was not that checking was tedious but that things were being missed. Reviews keep coming in, but a person only looks once in a while, and low ratings slip by unanswered in the meantime. Negative reviews with slow replies stay up for the next customer to see.

Writing replies was the same. The same praise got the same thanks, the same complaint got the same apology, written from scratch every time. It was heavy work for someone running a business to sit down to at the end of the day.

The problem we defined

Gathering everything was not enough. Even with it all on screen, not knowing what to look at first is the same as not having seen it. So we decided to tally separately the things that must not be missed.

Replies needed one step more. Writing for you and posting for you are different things. The words go out to customers under the store’s name, so a sentence the owner has not read must not be posted as is. Leaving one last human review step was the premise of the design.

What we built

• AI reply draft generation — along with the review text, it takes the star rating and the platform and writes the reply first. Even for the same sentence, the reply to a 5-star and the reply to a 2-star must differ. It is kept within 200 characters — long replies do not get read in delivery apps. The owner reads it and presses confirm to post, and can edit it right there if they want. This is the point where writing turns into choosing
• The button appears only on unanswered reviews — not on ones already replied to. The same criterion used to tally the things that must not be missed is applied to the button
• Automatic sync across the three platforms — once accounts are registered, it gathers the reviews
• Missed items as numbers — negative reviews rated 3 stars or lower and, among them, the ones not yet answered are tallied separately. The reply rate against the target is shown alongside, so a backlog is visible on screen immediately
• Keyword & sentiment analysis — frequent expressions are extracted from reviews and shown split into positive, neutral, and negative. This is where reasons invisible to ratings alone surface
• Per-platform comparison — see at a glance which platform’s ratings differ
• Always-on operation — reviews keep gathering and drafts keep being prepared without leaving a PC on

Results

Visiting each of the three sites became one screen, and writing replies became choosing replies. And unanswered negative reviews remain as numbers — the fact of a miss showing up on screen is what this system actually changed.

Other work on the same topic is collected under applied AI.