AI.
Senior Product Designer • 2023-2025 • Marketing Tech • SaaS • B2B Enterprise
AI.
AI.
2023-2025 • Marketing Tech • SaaS • B2B Enterprise • Senior Product Designer
Team
Team
Myself, 2 Product Managers, 1 Product Director, 1 Design Manager, 10+ Engineers.
Myself, 2 Product Managers, 1 Product Director, 1 Design Manager, 10+ Engineers.
Approach
Approach
Design systems, Stakeholders interviews, desk research.
Design systems, Stakeholders interviews, desk research.
Tools
Tools
Figma, FigJam, Google Docs, Slack, Zoom.
Figma, FigJam, Google Docs, Slack, Zoom.
The Idea
The Idea
At Khoros—a B2B SaaS platform for enterprise social media teams—we set out to create a chat-style AI for social media pros, but they told us they needed fast, practical help right in their workflow. After rethinking our approach, we tackled technical hurdles and built an AI caption tool that’s now a daily essential for marketers, streamlining their process and saving valuable time with content creation.
At Khoros—a B2B SaaS platform for enterprise social media teams—we set out to create a chat-style AI for social media pros, but they told us they needed fast, practical help right in their workflow. After rethinking our approach, we tackled technical hurdles and built an AI caption tool that’s now a daily essential for marketers, streamlining their process and saving valuable time with content creation.
“This tool saves me 30 minutes a day. I edit the suggestions, but starting with something that fits our brand is a huge help.”
“This tool saves me 30 minutes a day. I edit the suggestions, but starting with something that fits our brand is a huge help.”
— A happy Khoros Publishing user
— A happy Khoros Publishing user
The Process.
The Process.
I made use of our internal marketing team, who shared about 90% of our customer personas, since we didn’t yet have the affordance to speak directly with real users very often. This actually worked in my favor—they cared about what was being shipped and had a say in it too. So I ended up getting multiple touchpoints with them, which helped me run fast iterations and gather feedback along the way.
I made use of our internal marketing team, who shared about 90% of our customer personas, since we didn’t yet have the affordance to speak directly with real users very often. This actually worked in my favor—they cared about what was being shipped and had a say in it too. So I ended up getting multiple touchpoints with them, which helped me run fast iterations and gather feedback along the way.
Who we are helping and some science behind it.
Who we are helping and some science behind it.
Our users are social media managers, marketers handling multiple platforms, campaigns, and approvals under tight deadlines. They need captions that align with their brand voice, but starting from a blank page slows them down. We aimed to solve this real pain point, not just add AI for hype, to keep our top clients productive and satisfied.
Our users are social media managers, marketers handling multiple platforms, campaigns, and approvals under tight deadlines. They need captions that align with their brand voice, but starting from a blank page slows them down. We aimed to solve this real pain point, not just add AI for hype, to keep our top clients productive and satisfied.
The Problem
The Problem
Writing captions wasn’t hard, but starting them was. Users said that first drafts took the longest, especially when switching between platforms and brand tones. They were pulling ideas from other tools or old files, which slowed them down and doesn't scale well with internal collaboration.
Writing captions wasn’t hard, but starting them was. Users said that first drafts took the longest, especially when switching between platforms and brand tones. They were pulling ideas from other tools or old files, which slowed them down and doesn't scale well with internal collaboration.
Our analytics revealed users were stuck at the “Needs Content” stage and rely on brand voice and tone to start with.
Our analytics revealed users were stuck at the “Needs Content” stage and rely on brand voice and tone to start with.
Why AI? Finding the right fit.
Why AI? Finding the right fit.
I kicked things off by asking the team a simple but essential question: “Do we even need AI for this?” As we explored predefined templates and autocomplete tools, we quickly realised they didn’t resonate with our users—they lacked brand personality and often required more manual tweaking than expected, offering little room for on-the-fly creativity.
On top of that, these approaches didn’t mesh well with our current system, and building around them would distract us from our core focus on publishing and writing tools.
I kicked things off by asking the team a simple but essential question: “Do we even need AI for this?” As we explored predefined templates and autocomplete tools, we quickly realised they didn’t resonate with our users—they lacked brand personality and often required more manual tweaking than expected, offering little room for on-the-fly creativity.
On top of that, these approaches didn’t mesh well with our current system, and building around them would distract us from our core focus on publishing and writing tools.
So, we ran a quick experiment to test our hypothesis. We built a lightweight method to detect AI-generated posts based on speed and engagement, inspired by tools like Gemini and GPT Detector. This MVP helped us track how fast AI-assisted posts went live and how they performed across external channels.
So, we ran a quick experiment to test our hypothesis. We built a lightweight method to detect AI-generated posts based on speed and engagement, inspired by tools like Gemini and GPT Detector. This MVP helped us track how fast AI-assisted posts went live and how they performed across external channels.
AI Generated Posts
AI Generated Posts
30%
30%
AI Assisted Posts
AI Assisted Posts
60%
60%
Manual written Posts
Manual written Posts
10%
10%
Based on previous few quarters data and industry insights.
Based on previous few quarters data and industry insights.
Our findings, backed by early stakeholder conversations, a survey, and Pendo analytics, confirmed that these AI-Assisted posts were not only quicker to publish but also matched—or even outperformed—human-written ones in terms of likes and comments.
Our findings, backed by early stakeholder conversations, a survey, and Pendo analytics, confirmed that these AI-Assisted posts were not only quicker to publish but also matched—or even outperformed—human-written ones in terms of likes and comments.
Hence the AI.
Hence the AI.
We tried a bunch of ideas to surface the AI workspace and assessed design directions to take — some didn’t land.
We tried a bunch of ideas to surface the AI workspace and assessed design directions to take — some didn’t land.
It’s super important to highlight the AI panel without messing up the current workflows. Hence we placed the AI initiation button in close proximity to the existing compose actions.
It’s super important to highlight the AI panel without messing up the current workflows. Hence we placed the AI initiation button in close proximity to the existing compose actions.
We tried popovers, modals. All these options had larger form factors and consumed too much real estate or they blocked the post preview—users’ main decision space.
We tried popovers, modals. All these options had larger form factors and consumed too much real estate or they blocked the post preview—users’ main decision space.
We explored adding a contextual AI popover near the compose box—something that could act like a real-time writing assistant. But with our tech stack using modals, APIs, and iframes - this approach wasn’t practical within our timeline. Plus, one thing that didn’t work—and something I agreed with stakeholders on—was that the form factor felt too small for the kind of AI-assisted workflow we were aiming for.
We explored adding a contextual AI popover near the compose box—something that could act like a real-time writing assistant. But with our tech stack using modals, APIs, and iframes - this approach wasn’t practical within our timeline. Plus, one thing that didn’t work—and something I agreed with stakeholders on—was that the form factor felt too small for the kind of AI-assisted workflow we were aiming for.
Instead, we were settling for creating a panel that supported a natural editorial flow.
Instead, we were settling for creating a panel that supported a natural editorial flow.
It preserved real estate, respected users’ layout expectations, and gave us room for editing tools and didn’t block preview.
This also refers to the location of the Post Assets Selection panel, making it technically possible to repurpose existing code.
It preserved real estate, respected users’ layout expectations, and gave us room for editing tools and didn’t block preview.
This also refers to the location of the Post Assets Selection panel, making it technically possible to repurpose existing code.
A Chatbot is a type of AI that answers and does tasks on command and does a little of back and forth with you.
But Co-pilot is something that lives within the regular tools in close proximity and assists you in finishing your primary tasks quickly.
The what? - Exhausting Chats vs Quick Drafts
The what? - Exhausting Chats vs Quick Drafts
A Chatbot is a type of AI that answers and does tasks on command and does a little of back and forth with you.
But Co-pilot is something that lives within the regular tools in close proximity and assists you in finishing your primary tasks quickly.
We ditched the fancy AI chat experience early.
Why? It slowed users down. Chat UIs demand focus, and often spiral into back-and-forth loops that distract from the task.
We ditched the fancy AI chat experience early.
Why? It slowed users down. Chat UIs demand focus, and often spiral into back-and-forth loops that distract from the task.
Hence we pivoted to a non-chat approach, Always visible next to the preview.
Hence we pivoted to a non-chat approach, always visible next to the preview
Straightforward prompt field.
Optional tone, audience, platform settings.
Quick preview area with edit, regenerate, accept — all in one place.
Small nudges and tips for first-time users.
Sneak peak into helix AI model,
Sneak peak into helix AI model,
I worked with our AI engineering team to ensure the captions matched brand tones. Our API had limits with user inputs, but we tested a lot and got great results.
Closely working wth AI directors and their team. Engineers shared insights on how prompt parameters affected tone and specificity, which helped in shaping what inputs to offer to get worthy AI responses.
Simple Prompt Settings - "Set it and forget it"
Simple Prompt Settings - "Set it and forget it"
Some users like tweaking everything; others want to set it once and move on. Some liked just entering some textual prompts. I designed middle ground for all. They could choose tone, audience, or platform—or add brand examples. After setting, the prompt inputs it hid away to keep things focused for responses.
Some users like tweaking everything; others want to set it once and move on. Some liked just entering some textual prompts. I designed middle ground for all. They could choose tone, audience, or platform—or add brand examples. After setting, the prompt inputs it hid away to keep things focused for responses.
Bringing it into the design system.
Bringing it into the design system.
I evolved our existing design system with introducing by some chip dropdown that don’t take up too much real estate and can easily nudge into existing workflows whenever needed for secondary scaffolding.
I evolved our existing design system with introducing by some chip dropdown that don’t take up too much real estate and can easily nudge into existing workflows whenever needed for secondary scaffolding.
Did you know?
"the sum is only as good as its parts" as Aristotle once said. Focusing on individual components attributes value to the bigger picture and I deeply follow this.
May be he is Design system engineer too back then :P Who knows, buts lets resume.
"the sum is only as good as its parts" as Aristotle once said. Focusing on individual components attributes value to the bigger picture and I deeply follow this.
May be he was a Design system engineer too back then :P Who knows, buts let's resume.
This is to also ensure UI consistency and cognitive familiarity, so nothing felt new or overwhelming.
This is to also ensure UI consistency and cognitive familiarity, so nothing felt new or overwhelming.
Response card & helpful suggestions.
Response card & helpful suggestions.
Each caption comes in an easy-to-edit card with three options. Users like having choices, so we made it simple to pick, tweak, or try again. It feels like a helpful tool, not something confusing.
Each caption comes in an easy-to-edit card with three options. Users like having choices, so we made it simple to pick, tweak, or try again. It feels like a helpful tool, not something confusing.
We made sure, the new responses are highlighted to create that “new one” effect in the growing list of options.
We made sure, the new responses are highlighted to create that “new one” effect in the growing list of options.
If a response isn’t good, users can regenerate or edit — a quick way to improve upon.
If a response isn’t good, users can regenerate or edit — a quick way to improve upon.
I Worked with our "Content Design" team to make the AI copy more personalised and delightful.
I Worked with our "Content Design" team to make the AI copy more personalised and delightful.
Due to the limitation in API with the inputs, We added clear error guidance when users input the wrong prompts, guiding them to correct it.
Due to the limitation in API with the inputs, We added clear error guidance when users input the wrong prompts, guiding them to correct it.
Added a tooltip to reveal the entire prompt, so when users are scrolling through options and making a decision, they wanted to recognise which option they picked as opposed to recall from their memory.
Added a tooltip to reveal the entire prompt, so when users are scrolling through options and making a decision, they wanted to recognise which option they picked as opposed to recall from their memory.
Welcoming the first-time User Experience.
User Education & Onboarding
Welcoming the first-time User Experience.
User Education & Onboarding
To ease adoption:
A short tooltip in the panel header
A “View Examples” link that showed what good outputs looked like
Simple nudges. Easy learning curve
To ease adoption:
A short tooltip in the panel header
A “View Examples” link that showed what good outputs looked like
Simple nudges. Easy learning curve
Did it work?
Yes! AI captions helped posts go live faster, performed just as well (or better), and were quickly picked up by busy users. We saved them time and made their work easier.
Yes! AI captions helped posts go live faster, performed just as well (or better), and were quickly picked up by busy users. We saved them time and made their work easier.
Did it work?
Some reflections.
I took over 6 weeks of cadence that included kickoffs, iterations, explorations, handoffs and internal/external reviews.
This project wasn’t smooth sailing, I had to co-evolve with the problems and letting go of ideas I loved was tough, but it taught me to zero in and took middle grounds wherever possible without disrupting the core goal.
Not every project needs to be perfect—it just needs to help and unblock. Talking to users very often showed me what worked and what didn’t. That’s where the real learning happened.
Understanding how AI models work and limitations helped in framing the right design solutions and evolve with it.
I took over 6 weeks of cadence that included kickoffs, iterations, explorations, handoffs and internal/external reviews.
This project wasn’t smooth sailing, I had to co-evolve with the problems and letting go of ideas I loved was tough, but it taught me to zero in and took middle grounds wherever possible without disrupting the core goal.
Not every project needs to be perfect—it just needs to help and unblock. Talking to users very often showed me what worked and what didn’t. That’s where the real learning happened.
Understanding how AI models work and limitations helped in framing the right design solutions and evolve with it.
Some reflections.
Maha Krishnan
© All rights reserved
Made in Chennai—a beautiful beach 🌊 city in India
Maha Krishnan
© All rights reserved
Made in Chennai—a beautiful beach 🌊 city in India