Manychat's FlowBuilder is powerful — but it requires time and expertise that most new users don't have. I led the design of AI Assistant, a text-based feature that generates automations from a single prompt. Over two iterations, we reached a 36% conversion rate to publishing for new users.
FlowBuilder gives experienced users full control over automation logic — but that depth is a barrier for newcomers. Building even a simple flow requires understanding Manychat's structure, triggers, and message types.
The business question: Could we make automation accessible to users who don't yet speak Manychat's language?
The existing FlowBuilder — powerful, but complex.
We identified two distinct user segments with different value propositions:
New users
AI Assistant could remove the expertise barrier entirely, letting them build their first automation through a simple text prompt. Higher setup rate.
Experienced users
AI Assistant could save time on routine flows, reducing setup time for automations they already knew how to build. Faster time-to-publish.
This distinction shaped how we prioritized features and measured success across both iterations.
Before designing, I ran a qualitative research study to understand what users actually expected from an AI feature.
Method
Analyzed 60+ reviews from the Manychat Early Access Community on Facebook, identifying patterns across user segments.
Key finding
The majority of respondents were experienced users — agencies and educators — who didn't need simplification. Many were protective of their expertise. This told us the primary value of AI Assistant would be for new users, not the power users who were loudest in the community.
What this changed
We deprioritized advanced functionality for v1.0 and focused on making the first-time experience work.
Audience segments identified through review analysis.
Design approach
The goal of v1.0 was to validate the core concept and gather real usage data, not to ship a polished product.
User perspective
The user writes a prompt describing their business context and use case. AI generates an automation flow.
Technical perspective
We send the user's prompt alongside a JSON description of FlowBuilder logic to the model, which returns a structured automation.
Prompt → Manychat context → AI → Generated flow.
Alpha test — what the data showed
We rolled out to 1,300+ users. After one month, funnel data revealed two major drop-off points:
01 — Before writing a prompt.
02 — Before applying the generated flow.
To understand why, I ran a qualitative study with 10 users — two segments: those who abandoned before writing, and those who successfully applied a flow. I recruited via Intercom, wrote the interview scripts, conducted sessions, and synthesized findings for stakeholders.
Abandoned users
Didn't know how to describe what they wanted. They lacked both Manychat context and prompting intuition.
Successful users
Found the output too basic. They wanted more control and more complex logic than v1.0 could deliver.
Interview synthesis: abandoned vs. successful users.
What this told us
The audience for AI Assistant is new users — confirmed. But v1.0 was asking them to do something they weren't equipped to do: write a good prompt from scratch.
Design shift
Instead of a free-text prompt, we moved to a guided 4-question flow. Users select predefined answers; AI maps their responses to a template and fills in the automation content.
Why this works for new users
Removes the blank-canvas anxiety. Users make choices instead of writing from scratch.
Technical shift
Instead of generating a flow from a free prompt, the model now selects the best-matching template from a predefined library based on user answers, then fills in the content.
Classification questions → matched template → filled flow.
36% — Conversion rate to publish. New users who started the flow and shipped an automation.
20% — Retention. Users returning to AI Assistant after their first use.
The 36% CR was a meaningful signal — over a third of new users who started the flow published an automation, compared to the complex manual path before.
The 20% retention showed that AI Assistant was becoming a recurring tool, not just a novelty.
Two iterations gave us more than metrics — they gave us clarity. We confirmed that the real audience was new users, not power users. We learned that structured input outperforms open prompts for this segment. And we built a feedback loop — loyal early users, a prompt analysis dataset, and a backlog of validated features — that the next team could build on.
If I were to continue this work, I'd focus on closing the gap between what users attempt to describe and what Manychat can actually build — better expectation-setting at the entry point would likely move both CR and retention significantly.