The AI Agent race begins
In early 2024, AI agents (automation via AI) was barely an emergent concept. By 2025, AI agents skyrocketed into the mainstream. Businesses began seeking ways to harness AI systems that autonomously perform work, and Relevance AI was positioned well to win this race.
I was part of an ambitious project to create an AI agent marketplace, connecting users with agent builders.
My role
Research and testing
Prototyping
Production ready UI design
My team
1 x Product designer (me)
1 x Product manager
3 x Software engineers
Initial user problems
🧑💼
Recruiters (demand side)
Businesses and individuals want to harness the capabilities of AI agents and automations. However, they struggle to find suitable use cases, and setting up these automations remain a challenge.
🧑💻
AI agencies (supply side)
AI builders and automation agencies are a growing business model incentivised by building automations for businesses. Their pain point lies in finding clients and leads.
⚠️
Problems after launch
After my designs launched in our v1, multiple usability problems started emerging.
There wasn't enough trust and clarity to give recruiters confidence in cloning templates.
Builders lacked control and customisation to market their templates.
While successful templates do get published into the marketplace, actually using them still has friction, as often templates came with a set up process.
🔮
Marketplace vision
Create a thriving agent ecosystem for recruiters and AI agencies, eliminate the complexity of adopting AI automations, and establish Relevance as the leader in agentic automation.
Business incentive
AI agencies are emerging quickly as an industry segment, and the demand for AI automation is high.
There's currently a market gap for a high-quality AI automation marketplace.
Strategically, building a marketplace will enable us to create a compounding flywheel, leading to revenue growth as both recruiters and AI agencies need to pay for Relevance AI's services.
Demand-side metrics
# of marketplace agents implemented (run once successfully) - Validates customer desirability for agents
# of active marketplace agents (run 5 times a week) - Indicates that agents are providing ongoing value
Supply-side metrics
# of approved agents
# of active builders - indicates builders see value
🔬
Hypothesis
If Relevance AI launches a marketplace that connects recruiter demand with AI agency supply, this means simplified and mass agent adoption and a compounding flywheel. This results in increased usage and paid conversions, establishing Relevance as a leader in agentic automation.
🔎
My research approach
Because of tight timelines, I went for a research approach that allowed me to move fast.
I workshopped with my product manager to prioritise features for MVP (based on existing marketplaces), and we set up a dashboard to track key metrics.
For user research, I created a Slack community of power users and AI agencies for quick feedback loops.
🎨
My design approach
To enable a fast v1 ship, I referenced other notable marketplaces as many patterns are already solved and established. Using the prioritised MVP, I aim to deliver asap.
Builder user flow
For specifically the builder's journey, I mapped out the user flow as there were multiple touch points and dependancies in order for a builder to successfully publish a template.
Wireframes
I made early wireframes exploring what key pages, such as the recruiter landing page layout, could look like. I followed a 12 column grid.
We shipped a v1 and iterated from there based on feedback. Let me share! ☺️
Iteration showcase
Template listing page
The template listing page underwent changes. My v1 design was a clean and visual modal (before).
After our first release, there was significant friction from recruiters to clone or purchase templates.
I learned that the information displayed on the modal was not nearly enough to explain what the template did or how it worked.
Although we showed the clone count, there wasn't enough trust or reputation signals that the template was worth cloning.
Lastly, by building the listing page in a modal, we did not optimise at all for SEO, which would bite us as we rebuilt an SEO site from scratch.
I redesigned the listing page to a full page, replacing the visual element with an example task preview, and added a review system (after).
Before
Agent, tool OR custom avatar, and an exploration with photos
After
Default, task preview included, image or video included
Iteration showcase
Hero section
While most 'nice to haves' were not in MVP, the landing page hero section was the exception. I wanted to make a big visual splash when users first land onto our Marketplace.
The following are some iterations.
v0 carefree exploration
This version features our Invent product and two featured templates. An entry point to invent doesn't make sense, as it takes people off the marketplace. This never got shipped.
Another exploration
This version uses 2 columns with clear, bold elements. I explored things like glassmorphic buttons on the carousel. This version never got shipped.
Shipped v1
I went with a playful search CTA. The design entails floating agent avatars and colorful grainy gradient, and I worked with front-end dev to create delightful animations. Link to post.
Shipped v2
In our new SEO marketplace, I wanted to create a more scalable hero section with featured cards. We also wanted to highlight the categories more as an entry point. FYI I designed + coded it.
Iteration showcase
Template cards
Working with my visual designer, I designed v1 cards that were clean, playful, and on brand (before).
After our first release, I learned from both recruiters and builders that these cards lacked trust and had poor information density.
For users, it's impossible to make comparisons or be incentivsed to click on a template outside of the title and integrations. For builders, creating a differentiated template were difficult.
I redesigned the cards to include the description which provided a lot more context (after). We also built a review system that will bring life and trust to the experience.
Before
After
I'm still monitoring whether this change will cause any metric improvements.
PM and I made a Posthog dashboard to track clone % after click.
I iterated a lot on this 👇
What I've shown so far is the recruiter experience. On the other side are builders, who build, manage, and submit their agents.
Below are designs 😊.
Designs, builder experience
Listings page
The builder manages all their templates in the builder platform, in the 'listings page'. Builders can see the status of their submitted templates, and make edits.
Builder profile
I designed a simple screen for builders to update the profile that gets displayed to recruiters. This is also where builders set up their Stripe account to receive payment.
Publish to marketplace
After the user finishes making the template, they can publish to marketplace via a modal. I cover all edge cases such free vs paid templates, builder profile not set up, and Stripe not set up.
Outcomes
The product is still ongoing, and we get ~920 clones per week.
4.3% of users who clone a marketplace listing upgrade to a paid plan. The total average of all sign ups is around 1.9%.
62% of users clone free template after clicking on the template
Paid templates have low purchase rate. We are doing work now on this.
Challenges and learnings
🥶
Cold start problem
The product was functional, but not thriving like we wanted.
Marketplaces depend on the network effect. As a team, we didn't put enough resources into overcoming the cold start problem, as there wasn't enough high quality templates on the supply-side.
We are now alleviating this with events and hackathons.
⏳
Time crunch
Both my product manager and I were on this project part-time, while balancing other projects.
This led to sacrifices and trade-offs in running deep research, or effort in the community and recruitment aspect of running a marketplace.
🎯
Prioritisation choices
If I redid this project, I would have pushed prioritisation differently.
We included Stripe and payments in the MVP. We later learned that paid templates were rarely cloned.
We also didn't give enough love to community building, the cold start problem, and SEO optimisation.
We later had to rebuild the marketplace as a seperate web app for SEO.
Other screens
Builder profile (carefree exploration)
All results (shipped v1)
Builder directory
In our new SEO version, I designed a simple builder directory for recruiters to browse. After designing, I coded the front-end for it as well.
