# Cosmic Latte: AI Support Replies Reviewed by Humans

[Gleb Archipov](https://www.strv.com/blog/authors/gleb-archipov)  
Engineering Director

---

Cosmic Latte builds dating apps for the LGBTQ+ community. Its two platforms, Zoe and Surge, connect more than 15 million users around the world and facilitate roughly 1.5 million matches a month, with an average app store rating of 4.5.

As the apps grew, so did the support queue. Messages came in through chat and email, in multiple languages, and every one needed an answer that followed Cosmic Latte's support manuals. The team knew where the time went: the prep work before anyone wrote a word.

### The Challenge

Every reply started with the same groundwork. Admins reread the conversation to understand what the user needed, tracked down the right section of the right manual and adapted the answer to the channel's tone. When a message arrived in a language the admin didn't speak, translation came on top of that.

For the first phase, Cosmic Latte wanted to cut that groundwork while keeping people in charge of customer conversations. Every reply and every support action had to stay with an admin.

### What We Built

We built the AI Support Reply Assistant, powered by Anthropic's Claude. It prepares a tailored draft for every incoming message and puts it in front of an admin for review.

Here's how it works. When a message comes in, the assistant reads it alongside recent messages from the same conversation and picks out questions, reported issues, feedback and relevant details. It matches the request to the right chat or email support manual and drafts a reply that fits. For email, drafts appear as internal notes in Groove, so admins stay in the tool they already use.

Each draft is written in the user's language, with an English version next to it. That means any admin can review any reply, whatever languages they speak. Chat and email each keep their own defined tone.

### Why Claude

We chose Claude because support work is hard to script. Threads get messy, users describe problems in their own words and the right answer often depends on context spread across several messages. Claude follows that context well and stays grounded in the manual it's given, which keeps drafts aligned with Cosmic Latte's policies. It also writes natural replies across languages and holds a consistent tone.

That combination is where the time savings come from. By the time an admin opens a ticket, a manual-backed draft is already waiting, and their job becomes reviewing and refining.

### Designed for Trust

We spent as much time on what the assistant shows admins as on what it writes. Every draft comes with an estimated confidence score, the manual sections it matched and a short note on recommended checks or actions.

We also built it to know its limits. When the manual doesn't support a clear answer, the assistant flags the missing information and can return guidance without a draft. In customer care, a confident wrong answer does more damage than a short delay, so we treated uncertainty as a feature.

Admins stay responsible for reviewing, editing and sending every reply, along with any support actions a request requires.

### The Impact

- **144ms** average time to generate a suggestion  
- **57%** of suggestions scored above 90% confidence  
- **45%** more requests processed per admin  
- **33%** of drafts sent with minimal or no edits  
- **18** languages handled

### Their Words, Not Ours

> *"I am extremely excited to have AI introduced into our daily workflow. It saves time and helps us provide stellar customer service."*

-- Monika Ondruj, Head of Customer Support, Cosmic Latte

---

Don't miss anything