This project started with a question common to every martech platform:
"How can we reduce the time and effort it takes a marketer to create a campaign?"
How it all started
A conversation with Nalin, our Head of Product — alongside a look at our analytics — surfaced three problems our users were facing:
It takes 4–6 days on average to publish a campaign.
Marketers get stuck on content, which delays publishing.
Marketers turn to AI for inspiration and lean on A/B testing to finalise content.
Research: is the problem real?
I started with research to confirm the problem was real. I reviewed 54 Pendo tickets and spoke with marketers from Tata Digital, Tira, Giva, Wynk Music, Tata Capital, Airtel, Blibli, Adda247, Park+, Amazon miniTV, AlfaGift, and SoundCloud.
01
Their content feels boring, and they need inspiration for something new.
03
"I rarely use AI for content creation — it gives vague suggestions instead of clear, data-backed ones."
02
"After 5+ years in the industry, I sometimes still can't think of unique ideas."
04
"I always have to test my content before publishing, which takes a lot of time."
Competitive research
I studied five competitors (direct and indirect), to understand the market. Direct: Braze and CleverTap. Indirect: Copy.ai, Grammarly, and OpenAI.
All of them offer an AI writing tool, with varying levels of sophistication. The common weaknesses: brand guidelines are hard to set up, the tools aren't intuitive, and they struggle across a wide range of content types. The opportunities: integrations with other tools, personalised content, and adopting newer AI models. The threats: strict content policies and long setup times. Full SWOT analysis →
Design process
Beyond what I already knew, I needed a deeper understanding of four things:
01.
How is the feature built? Which model powers it, and what are its limitations?
It's built on OpenAI's GPT-3. From MoEngage, we pass additional data along with the user's prompt, that's the major value add.
03.
What outputs will the feature generate? How will they vary, and when will they fail?
AI is probabilistic, so the generated content varies for each user. It fails when the prompt isn't structured properly.
02.
What data is available to the feature? and how good and reliable it is ?
A workspace's previous campaign data is available, and it's fully reliable as it's the customers actual data.
03.
How might users react in the worst case?
If we generate poor content, it shouldn't get in the way of creating a campaign. The whole flow should be easy to exit at any point.
"Designing a probabilistic system where outputs change in real time based on the user's input meant, I had to anticipate surprises and design around them."
Most SaaS teams follow Jesse James Garrett's five-layer model. It works well for deterministic systems, but it doesn't account for the extra considerations a probabilistic system like GenAI introduces. Ones that affect UX decisions downstream.
User journey
I mapped the current journey against the expected one.
Interaction: the "fill in the blanks" idea
In school, we all dealt with different question formats: true/false, match the following, short answer, long answer, and fill in the blanks.
I always found fill-in-the-blank questions easier than long-answer ones. With a long answer, I have to structure my response before I write it. Fill in the blanks gives me the structure, I just have to complete it with the right keywords.
Structuring a prompt from a blank cursor is the hard part. So I applied the same paradigm here: a UI that mimics a fill-in-the-blanks structure, except the blanks are now dropdowns and input fields.
After several rounds of ideation, we landed on three user flows, each with its own benefit. I prototyped all three in Figma and tested them with users.
We chose the modal slider for a few reasons:
-
The flow is intuitive rather than intrusive.
-
If the system doesn't produce good output, the marketer can still create the campaign.
-
It gives a clear preview of the content before it's added to the campaign.
-
It's easy to ideate on and tune prompts.
-
It's scalable and easy to reuse for other campaigns.
-
It's conversational and approachable for marketers new to GenAI — a pattern already used in our segmentation module.
Detailed design
Because we already had a design system, I reused most components without compromising the experience. When I needed an autocomplete component that didn't exist in the MoEngage design system, I built it and added it to the system
Key highlights from the final designs
Structured input fields for better guidance
Most users don't know how to write effective AI prompts — and they shouldn't have to. A blank text box is overwhelming. Instead of making users type everything manually, MerlinAI guides them with structured input fields
-
Drop-downs to define the type of output (e.g. blog post, product description).
-
Keyword auto-suggestions to include or exclude, based on previous campaign data.
-
A toggle to add emojis.
​Context-aware adjustments let users tweak specific parts of the prompt instead of rewriting the whole thing. Users don't have to find the "right" way to phrase the prompt they select what they need, and MerlinAI does the rest.
Preset templates & keyword suggestions
Instead of leaving users to figure things out by trial and error, MerlinAI provides:
-
Predefined templates for common tasks (e.g. Black Friday, flash sales, clearance sales).
-
Example prompts that adapt to the client's industry.
By reducing guesswork, MerlinAI reaches an effective output in less time
Iterative adjustments & instant feedback
One of the biggest frustrations with AI tools is having to rewrite an entire prompt to make a small change. MerlinAI is built for iteration from the ground up:
-
A preview of the composed message with pagination, so users can tune the output on each pass.
-
One-click regenerate that keeps the same context for fresh suggestions.
This turns MerlinAI into a collaborative process rather than a one-shot tool.
Quick controls for generating AI variation
When a user has already written a prompt, Create AI variation quickly creates another variation of the content in a single click. This became a way for users to trust MerlinAI by running the AI variation in A/B testing.
A final walkthrough covered the design and PRD at dev handover. GTM strategy followed, and MerlinAI launched at the NeXt event in London.
Let's launch
Impact
We surpassed the success metrics we'd 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
Customer case studies
-
Glance achieved 50% faster campaign go-live times with MoEngage's Merlin AI. Case study
-
Telekom Romania (a Deutsche Telekom subsidiary) improved campaign CTR by 65% with Merlin. Case study
Further ideations
Ideation 1
Native language support. Although we were happy with the results, Merlin AI underperformed in a few regions — especially Southeast Asia. User calls revealed that marketers there needed content in their native languages. It was already possible via prompts, but we added a dedicated language dropdown to make it explicit.
Ideation 2
Upgrading the model. After GPT-4 was released, we tested it internally, saw better results, and upgraded MerlinAI to GPT-4.
Ideation 3
Scaling across channels. Once retention stabilised, we rolled MerlinAI out to Email, WhatsApp, and other channels, making the feature scalable.
Next steps
Building on the value of AI in copywriting, we're releasing more products that follow the same UX patterns. These may eventually be packaged and sold as the MerlinAI suite:
-
MerlinAI Designer (live) — generates images from user prompts; currently in beta with a few customers.
-
MerlinAI for Segmentation (in development) — describe a segment in plain language, and Merlin AI reads the existing event attributes to build it.
-
MerlinAI Labs (in progress) — a suite of martech-focused AI products (content, segmentation, personalisation, flow creation) under one roof.