AI drafts replies to delivery and mall reviews — you only press approve

Collecting reviews → classifying them → drafting replies can be automated; the final review and posting are done by a person. The reply goes out to customers under the store’s name, so a sentence the owner has not read must not be posted as is.

The problem is not that checking is tedious — it’s missing things. Reviews keep coming in while a person only looks occasionally, and low ratings slip by without a reply in between. So the design starts not from “writing faster” but from “counting separately what must not be missed”.

① Collecting — start with one place

Authentication methods and data formats differ per platform, so each collector is built separately. Building everything at once gets heavy, and one blocked platform would stall the whole thing — so we connect the place with the most reviews first and grow from there.

Check once before starting

Platforms define in their terms how far data from the owner-facing system may be exported. For example, Coupang Eats(쿠팡이츠) merchant terms of service clause 38.2 restricts providing or processing the statistics data of the seller-only system for third parties, and clause 44.2 of the same terms places the responsibility for managing the ID and password on the seller.

No need to overthink it. Check the merchant terms of the platform you use and start with what is allowed. They differ from platform to platform, and even within the same company, rider terms and merchant terms are different.

We do this check for you at the inquiry stage — just tell us which platform you use. The technical standard for judging is laid out in Web crawling: how far does it go?.

② Drafting — what you give the AI alongside

We don’t feed in only the review text. The rating and the platform go in with it. Even for the same text, a reply to a 5-star review and a reply to a 2-star review must be different.

Length is capped — a character limit is set in advance. How a reply looks differs per platform, so the cap is set only after actually posting one and checking it on screen.

The limits are clear — emotionally charged reviews and apology phrasings that could be misread are for a person to fix.

③ Review and post — writing becomes choosing

1. Open the review list
2. Read the AI draft
3. If it’s fine, post it with a single approve button
4. If you want changes, edit it right there

The button only shows on reviews not yet answered. What hasn’t been done is visible on the screen at a glance.

What a person decides — who presses the final approve must be settled.

④ See it in numbers — what goes on the screen

Metric	Why it’s needed
Unanswered count among low ratings	The backlog is visible at once. What counts as low is for the store to decide
Reply rate vs. target	Absolute counts alone don’t show whether you’re falling behind
Keyword and sentiment analysis	Reveals reasons that ratings alone hide
Comparison by platform	See where ratings split

The last decision — run it around the clock?

A setup that needs a PC left on and one that doesn’t are built differently. The hours customers leave reviews and the hours the owner sits at a PC don’t overlap, so without always-on operation, “missing” stays as it is.

It’s a decision that’s hard to change later, so we settle it at the start.

---

All we need is the name of the delivery app or mall you use.

You don’t need to come with counts. Tell us what you use, and we’ll start by checking that platform’s merchant terms and lay out how far the automation can go.

→ Ask about review management · Work we’ve done

---

References

• Coupang Eats(쿠팡이츠) merchant terms of service, clause 38.2 · clause 44.2 [verified — original terms, retrieved 2026-08-17]