Softgarden(ATS)
2024
I turned repetitive job ad restructuring into a one click AI workflow, increasing Structured Job Description adoption to 29.8%
Role
Team
1x Product Manager
5x Engineers
1x Designer
1x QA
Timeline
April - July 2025
Constraints
Kununu dependency: Company name had to match exactly; automatic retrieval wasn't reliable.
Existing ATS patterns: Side panels weren't ideal for a multi-step setup, so a new modal pattern was needed.
Impact
Problem area
Recruiters structured the same job description twice
Recruiters wrote their job ad once in the Standard Job Description, then manually copied and reorganised the same content into the Structured Job Description required by job boards like Indeed, Stepstone, and LinkedIn. Every time, they had to figure out where each piece belonged, copy it across, and verify the result. This daily repetition cost recruiters time and led to copy paste errors, missing information, and inconsistent job ads reaching candidates.
For the business, those inconsistencies damaged the employer's brand and weakened candidate trust in the quality of the posting. The friction also reduced recruiter efficiency and made Structured Job Descriptions harder to adopt at scale.
Image shows what recruiters had to do before publishing
Design goals
What I wanted to achive
I didn't just want to automate copy and paste. I wanted to remove the repetitive work while keeping
recruiters in control.
Easy to adopt
Fit the new experience into the workflow recruiters already know, so it doesn't add unnecessary complexity.
Why AI became the right solution?
Before jumping to AI, I wanted to try the simplest solution first. I looked into automating the restructuring with basic rules. To test the idea, I pulled real job descriptions from Customer Support and quickly saw the problem: every recruiter wrote differently. Some used one long block of text, while others created custom sections with their own headings and ordering. This made a rule based approach unreliable, and the legacy codebase made it even less realistic.
I took the finding to the Product Manager and Engineering. We explored whether an LLM could handle this variation and confirmed it was feasible. But I drew a clear line: GPT would not create or rewrite content. It would only organise what recruiters had already written.
Design decision 01
Integrated AI into the existing workflow
With 2,000+ jobs created monthly, introducing a separate AI tool would have added one more step to an already lengthy process and recruiters already relied on the existing Job Creation Wizard. Instead, I integrated AI directly into the existing workflow, making the Copy with AI action available only after recruiters enabled the Structured Job Description. This preserved recruiters' existing mental model, reduced the learning curve and introduced AI only at the moment it became relevant.
Design decision 02
Replace repetitive work with one click
Recruiters previously copied, pasted and reorganised the same content section by section from standard job ad description to structured job ad description. I replaced that repetitive workflow with a single Copy with AI action that organised the existing job description into the required structured sections. With 2,000+ jobs a month, removing that repetition across every job made a meaningful difference.
Design decision 03
Keep recruiters in control
Job ads represent the employer's brand to candidates, so a misplaced section can affect how a role appears on Indeed or Stepstone. Rather than generating or rewriting job descriptions, AI was intentionally limited to organising the recruiter's existing content into the appropriate structured sections. I made recruiters to explicitly decided when to use AI and remained free to review and edit every generated section before publishing.
These decisions ensured AI reduced repetitive work while recruiters remained in complete control of the final job advertisement.
Learning from testing
Recruiters' main hesitation was trusting AI to organise content correctly. They wanted to check before publishing, confirming that automation of conent would have been the wrong call.
"Rakesh is always trying to reach our customers through Canny, even under conditions where the response rate is significantly low. Whenever we have time to run a full design and development cycle for a project, he makes sure collaborating with the research team and get users on a meeting."
Design decision 04
Created a dedicated visual language for AI
Because this was the company's first AI-powered feature, there were no established patterns for AI interactions. I introduced a dedicated visual language using a distinct colour, iconography and button treatment so recruiters could clearly distinguish AI assisted actions from standard product interactions. These components are now reused across other AI features in the product.
Retrospective
The workflow gained adoption, and it shaped how we approached AI
Impact
Reflection
AI isn't always about doing more
The biggest lesson was that AI wasn't the solution by itself. The real design challenge was deciding what to automate and where human judgement should remain. I learned to use AI to remove repetitive work while keeping recruiters responsible for the final result.
Measure what you want to improve
We could see that recruiters were using the new workflow, but we couldn't tell how much time it actually saved or whether it reduced copy and paste errors. Next time, I would define those measures earlier and make sure we can track them from the start.
Thinking in Patterns and not Just Screens
This was Softgarden's first AI powered feature, so I wasn't just designing one experience. I also helped establish how AI could look and behave across the product. The patterns we created are now reused in other AI experiences.