A Top Dating Platform Turned 40,000 Tickets Into Support AI

The platform's data science team put an LLM support agent through a 12-week proof of concept, winning over the CEO and CPO and expanding into an embedded innovation sprint.

40,000+

Trained an LLM support agent

2–4 months

Implementation Time

Aanikh Kler

Lazer Technologies | Co-Founder Surf (acquired) | Canada's Young Entrepreneur of the Year 2023

Lazer Technologies

Connect ↓

The Challenge

One of the world's largest dating platforms, home to one of the industry's most respected data science teams, faced a high-volume support cost center and an open question: could LLMs automate first-line support without the brand damage of an AI saying something unsafe on a dating app? The bar was double-sided: answers had to be accurate, and the system had to stay composed when users asked it anything at all. The platform also wanted to understand whether AI belonged in the dating experience itself.

What They Built

In a 12-week first phase, Lazer built an AI support chatbot POC trained on more than 40,000 real support tickets, deployed in Lazer's own environment for controlled internal testing. Evaluation was designed in from the start: accuracy tracking plus explicit brand-risk metrics measuring how the model behaved under adversarial or random questioning. A customizable default prompt gave the platform direct control over the agent's voice and boundaries, a decision that particularly impressed senior stakeholders.

Phase 2 prototyped "dating copilot" concepts: conversation assistance, opener generation, and profile-browsing interaction patterns, validating the direction the platform's internal team needed to take to bring AI into the core product. On the strength of both phases, the engagement expanded into an innovation sprint with additional Lazer engineers embedded directly with the platform's data science team as the LLM experts in the room, with full code handoff to the internal trust and safety team.

Lazer scoped the work as a 12-week first phase, starting from the platform's core question: could an LLM handle first-line support without saying something unsafe on a dating app. Rather than optimize for accuracy alone, the team designed evaluation in from the outset, defining brand-risk metrics that measured how the agent behaved under adversarial or random questioning alongside standard accuracy tracking. They assembled a retrieval corpus from more than 40,000 historical support tickets and stood the chatbot up inside Lazer's own hosted environment, so testing could run internally under controlled conditions before any customer exposure. A customizable default prompt was added to give the platform direct control over the agent's voice and boundaries. Working through gradual data access in an uncertain environment shaped much of the engineering. A second phase then prototyped "dating copilot" concepts, exploring conversation assistance, opener generation, and profile-browsing patterns. As the phases progressed, additional Lazer engineers embedded with the platform's data science team, with full code handed off to the internal trust and safety team.

AI Role

An LLM-based agent answers first-line support questions using retrieval over a knowledge base built from 40,000+ historical tickets, with evaluation infrastructure measuring both answer accuracy and brand risk under adversarial prompting. A second phase applied LLMs to the dating experience itself through copilot prototypes.

AI Model

To be confirmed

Infrastructure

Lazer-hosted environment for controlled internal testing • UI and API delivered for evaluation

Integration Points

Designed for client-side integration and handoff • Code transferred to the platform's trust and safety team

Impact

40,000+ Tickets Became the Training Corpus

Years of support history were transformed into the retrieval foundation for an agent targeting a high-volume cost center.

Validated From the CEO Down

The POC won over the platform's CEO, CPO, and ML team, with the CPO pushing to get the capability into the app, the adoption signal pilots rarely achieve.

1 Engineer Became an Embedded Innovation Sprint

The engagement expanded from a single engineer to a multi-engineer sprint embedded with the data science team, spanning internal AI and new product surfaces.

Technology Utilized

Implementation Complexity

Cost Reduction

Customer Experience

Best Fit For

Consumer platforms with high-volume support operations and high brand sensitivity, where an AI failure is a screenshot away from a PR problem.

Aanikh Kler

Lazer Technologies | Co-Founder Surf (acquired) | Canada's Young Entrepreneur of the Year 2023

Lazer Technologies

Award-winning entrepreneur and tech leader helping startups and global brands like Netflix and Amazon drive growth by ethically harnessing consumer data and AI-driven innovation.

Get an intro

Industry:

Technology & Software

Media & Entertainment

Business Function:

Customer Service

Product & Engineering

Company Size:

5,001+

Project Cost:

Cost not disclosed

Ownership:

Organization Type:

AI Pattern:

Conversational AI (Chatbot / Agent)

Knowledge Management & Search (RAG)

Natural Language Processing

Value Type:

Cost Reduction

Customer Experience

AI Model:

Frequently Asked Questions

How did a dating platform build an AI support agent from its support tickets?

Lazer trained an LLM support chatbot on more than 40,000 of the platform's real historical support tickets, using retrieval so the agent could answer first-line questions. The proof of concept was built on that corpus of 40,000+ tickets.

What AI approach and tools did the support automation use?

It used a large language model with retrieval (RAG) over a knowledge base built from historical tickets, run in a Lazer-hosted evaluation environment. Evaluation tracked both answer accuracy and brand risk, and a customizable default prompt controlled the agent's voice. The specific model was not disclosed.

What results did the dating platform's support AI proof of concept achieve?

The POC turned 40,000+ tickets into the agent's training corpus and won validation from the platform's CEO, CPO, and ML team, with the CPO pushing to bring the capability into the app. The engagement then expanded from one engineer into a multi-engineer innovation sprint.

How long did the support automation proof of concept take?

The first phase ran as a 12-week proof of concept, within a two-to-four month window. A second phase then prototyped 'dating copilot' concepts before the work expanded into an embedded sprint.

Who is AI support automation best suited for?

It fits consumer platforms with high-volume support operations and high brand sensitivity, where a single AI failure could become a screenshot and a PR problem.