How Indeed built an AI product that users love

Eight months. That's how long we had to go from zero to a fully functioning AI product, operating at scale. 

I'm Margaret Howe, Director of Product Management at Indeed, and in January 2025, our CEO asked us to build an AI career expert in time for our annual conference. In this article, I’m going to share that journey and the lessons we picked up along the way. 

Here’s a taste of what you’ll learn:

• Why AI is more than a chatbot, and why that distinction matters
• How obsessing over quality before a launch changes everything
• Why AI works better woven into existing journeys than bolted on as a separate hub

How Career Scout came together

Career Scout launched in September 2025 with the tagline "Your personal career coach." The product sat on top of Indeed's existing job search experience, using natural language to make a complex matching system feel effortless to navigate.

We launched with three standout features: 

1. A mock interview tool 
2. A resume builder
3. A chatbot to guide users through job discovery

The launch was successful, but a lot happens in nine months, and we learned some valuable lessons along the way. Let me walk you through them.

Lesson 1: AI isn't just a chatbot

There's a tendency in senior leadership to treat AI as a synonym for ChatGPT or Claude. AI can be much more than a chatbot, and confusing the two limits what you build.

The other thing to keep in mind is the importance of staying rooted in user problems. The current AI wave is blinding a lot of people to that basic discipline. It’s tempting to build something cool that signals you're "doing AI" and keeps shareholders happy, but you still need to address users’ pain points.

Start with the problem, not the tool

We wanted to use LLMs to deliver something genuinely useful: a personalized job seeker discovery experience.

One of the biggest problems job seekers face on Indeed is that they struggle to articulate what they actually want. Many are generalists who know they need a new job but don't know what their skill set qualifies them for. We wanted LLMs to surface personalized, grouped recommendations based on what we already knew about each seeker, so they could discover options they hadn't considered. 

With the job market as tight as it is (many of the job seekers I speak to weekly have been looking for four or five months), helping people find work faster matters.

Meet users where they are

We also had to acknowledge that the chatbot interface was new for most of our users. 

The average job seeker on Indeed isn't necessarily fluent in prompting ChatGPT, so we built an experience that guided them and gave them clear ways in: conversation starters, easy-to-tap suggestions, and a confirmation flow that showed them what we already knew about them. They could still type their own queries, but the discovery felt curated and personal.

Lesson 2: Obsess over quality

This one also sounds obvious. Of course we care about quality. We launch things, we look at the metrics, we test for bugs, and we ship fixes.

The wrinkle with AI agents is that they have agency to do whatever they want. So you need the right guardrails and the right measurement processes in place to actually understand what's happening in your product.