The whole project started with a univeral question around all Martech platforms.
" How can we minimise the time and effort taken by the marketer to create a campaign?"
How it all started
When I had a conversation with Nalin, our head of product and looking at our analytics, I found 3 problems our users were facing
It takes 4-6 days to publish a campaign on average
Marketers get stuck at content which leads to delay in publishing a campaign
Marketers use AI for inspiration and A/B testing to finalise their content
I started out with research, to understand if the problem is real
After looking at almost 54 Pendo tickets and talking to marketers from Tata digital, Tira beauty, Giva, Wynk music, tata capital, Airtel, Blibli, Adda247, park+, Amazon mini_tv, alfagift and soundcloud, here’s what I found:
01
Felt their content is boring, and need inspiration for new content
03
“I rarely use AI for content creation, as it gives vague suggestions instead of clear data”
02
“After being in the industry for 5+ years, I’m not able to think of unique ideas sometimes”
04
“I always have to test my content first and only then publish it, which takes a lot of time”
Competitive research
I looked at 5 competitors (direct and indirect) to gain insights on what the market is. Direct competiors were braze and celevertap. Indirect competitors were copy.ai, grmmarly and OpenAI. The insights are as follows:
All of them have an AI-based writing tool with varying levels of sophistication. The main weaknesses for most of them are the difficulty of setting brand guidelines, the lack of intuitiveness, and the inability to handle a wide range of content. Opportunities include integrations with other tools, personalized content, and leveraging new AI models. Threats include strict content policies, a long setup time. For full SWOT analysis, click here
Design process
In addition to what I knew, I also needed a deeper understanding of:
01.
How the feature is built ?
It was build on OpenAI GPT 3.0, from MoEngage we will pass additional data wit the prompt defined by the user. This will be a major value add.
03.
The outputs the feature will generate, how they will vary, and when they will fail ?
AI is a probabilistic system, so content generated will vary for each user, It will fail of the prompt is ot structured properly
02.
What data is available to the feature? and how good and reliable it is ?
A workspaces previous campaigns details are available, it is 100% reliable data
03.
How users might react to this feature in the worst case scenario
In case we don’t generate good content, it should not be intrusive to the user to create a campaign. The entire flow should be easy to quit if needed.
“While designing a probabilistic system , where the outputs are dynamic to inputs the user provides real time. I had to predict surprises and design around them”
The design process largely followed by SAAS compaines, (James Garrett’s 5-layer model model) works well for deterministic systems. But doesn’t capture the additional elements needed for a probabilistic system ( like GenAI ) which will affect the UX considerations downstream.
User journey
I created what the current journey looks like and what expect journery will look like.
Interaction
In our school days, we would have faced different types of question formats like true or false, match the following, short answer, long answer and fill in the blanks.
I always used to feel fill in blank type questions are easier than long answer questions. As i have to structure my answers first before i write. In fill in the blanks a structure is predefined and just filling the blanks with the keywords is more than enough.
To structure a prompt with ideal inputs and starting with the blank cursor is the difficult part. So i thought why not use the same UX paradigm here. So i created a UI that mimics filling the blanks structure, the only difference being the blanks are now dropdown’s and input boxes.
After various ideations, we ended on with 3 types of user flow as each had their own benefit. I made prototypes for all 3 the flows and tested them as figma prototypes with some users.
We finalised on using a modal slider as it had the flowing advantages :
-
The flow is not intrusive but intuitive
-
If the system fails provide good output the marketer can still create the campaign
-
Clear understanding of how the preview will be before the content is pasted inside the campaign
-
Easy to ideate and tune the prompts
-
scalable to be easily reused for other campaigns
-
Conversational model, easy to use for marketers who have not use any GenAI platforms. This pattern is already available in segmentation module.
Detailed design
Detailed design in figma
As we had a design system in place, we reused most of the components, but the experience was not compromised. When i needed a new component for autocomplete which was not available in MoEngage design system, i created the and added the component to the design system.
Key highlights from the final designs
Structured Input Fields for Better Guidance
Many users don’t know how to write effective AI prompts—and they shouldn’t have to. A blank text box can also be overwhelming. Instead of forcing users to type everything manually, MerlinAI guides them through the process with structured input fields :
​
-
Drop-down menus to define the type of output (e.g., blog post, product description, UI component).
-
Auto-suggestions on keywords to include and exclude based on the previous campaign data.
-
Toggle switches to add emoji’s.
Context-aware adjustments lets users tweak specific parts of the prompt content instead of rewriting the entire prompt. Using these elements, users don’t have to struggle with the “right” way to phrase something. They just select what they need, MerlinAI does the rest.
Preset templates & keyword suggestions
Instead of making users figure things out through trial and error,
MerlinAI provides:
-
Predefined templates for common tasks (e.g., Black Friday sales,Flash sale, clearance sales etc).
-
Example prompts that adjust dynamically to different user needs based on the industry of the client.
By reducing guesswork, MerlinAI interactions are smoother and more accessible. By this way the generated output is more effective in less iterations.
Iterative adjustments & Instant feedback
One of the biggest frustrations with AI tools is having to rewrite an entire prompt just to make small changes. Instead, MerlinAI’s UI is build groundup for a iterative workflow:
Preview of the composed message with pagination. This allows the users to get creative and tune the output on each iteration
One-click regenerate with the same context for more suggestions.
This makes MerlinAI interactions into a collaborative process rather than a one-time shot.
Quick controls for generating AI variation
One of the biggest frustrations with AI tools is having to rewrite an entire prompt just to make small changes. Instead, MerlinAI’s UI is build groundup for iterative workflow:
Preview of the composed message with pagination. This allows the users to get creative and tune the output on each iteration
One-click regenerate with the same context for more suggestions.
This makes MerlinAI interactions into a collaborative process rather than a one-time shot.
A final walkthrough was held for Design and PRD with the Dev handover. GTM statergy was later decided and Merlin AI was launched in NeXt event in London.
Lets launch!!
Impact and analysis
We have surpassed the success metric we had previously defined,
Unique accounts using the feature
Target - 100 accounts
Achieved - 273 accounts
Reduction in time taken to create campaign
Target - 30% in average
Actual - 50% in average
Engagement and uplift
Target - 10% uplift
Actual - 24% uplift
Successful case study’s
Further ideations
Ideation 1
Though we were happy with the results, I could see some Merlin AI not performing in a few demographics, especially in SEA regions.
We understood from user calls it was because they needed content in native language, though it was possible using prompts we added an extra dropdown for the marketer to specify language
