AI Content Assistant
What is Symend?
Symend is a customer engagement platform that enables personalized content creation, including emails, texts, websites, and event orchestration. Enhanced by behavioral scientists, it ensures personalized messaging tailored to specific use cases.
Project Overview
The Content Assistant is broken down into three sub features, each leveraging Symend's LLM trained in partnership with a team of behavioral scientists:
1. Content Analyzer: Identifying behavioural tactics, tones of voice and other metrics.
2. Content Generator: Generating whole content text based on user input.
3. Content Improver: Modify existing text based on user selection.
Roles
Product Manager, UX Designer
Worked with
Engineers, Product Managers,
Behavioral scientists, CPO
What is a Content Builder?
A content builder allows users, regardless of technical skills, to create and manage emails, texts, and landing pages for customers. Users are able to edit content in real-time and create a set of rules as to when the content is sent and who it is sent to.
Content takes too long to create.
It has low accuracy, and there is little variety.
Behavioral tactics are tracked externally instead of attached the piece of content.
Reduce the time to create accurate content.
Better align content with intended use cases.
Enable Behavioural tactic awareness within the content builder.
Make it seamless - this feature is to assist the user, not to take their role.
Analyzed industry-leading AI-assisted content tools to leverage common behaviour and workflow patterns.
Refined common workflows from the comparative analysis to fit key use cases specific to Symend.
Created wireframes based on the user workflow and iterated with feedback from product and design peers.
Worked directly with a team of engineers to quickly develop a working prototype that connected to our LLM APIs. This was made off the platform to gather feedback on the user experience and quality of output.
Tested and interviewed users across 4 departments ranging from experienced content creators to beginers. User Interviews were conducted with experienced content builders that used the prototype, and user testing sessions were conducted with novice users.
Key Findings:
Users use the Analysis and Generate separately and saw no value in using them jointly.
There was a need for brainstorming from existing content, a way to iterate previous text.
Analysis at the document level was not enough, need to select custom amounts of text.
Analysis was often overlooked, many only thought it was a generate feature.
Actions:
Separate Generate and Analyze flows + add a new use case: Improve, by combining Analyze and Generate.
Allow for custom length analysis, at sentence level.
Present certain metrics in realtime to draw more attention to the analysis feature.
Integration & High Fidelity Prototyping.
Generate
Analyze
Improve
Tested users across 3 departments and presented to the greater Product and Design teams. Gathered feedback on the integration into the core platform and the new behaviours the Content Assistant brings to existing features and workflows.
Key Findings:
Integrate the Content Generator within the text block, not as it's own block.
Consider a more familiar approach with the Improve function.
Users did not expect for improved text to be additive.
Actions:
More Generate to block level within the rich text editor.
Ideate on Improve flow and leverage previous comparative research.
Change outcomes in user flow: Generated text is additive, Improved text gets replaced.
Final Prototypes.
Generate
Analzye
Improve
Pressure from Stakeholders.
After receiving pressure from executives and requests from clients, our timeline to deliver started to shrink. Each sub-feature was reduced to an MVP release with limited featires to streamline integration. Partnering with a team of engineers, we reduced the estimated delivery time by over 60% and got the alpha release shipped at the start of 2024.
Alpha Release.
Generate
Analyze
Improve
Learnings & Takeaways.
Learnings:
Receiving pressure from executives and clients in the end helped us deliver value faster, while maintaining similar outcomes. If this was done earlier on it would have further reduced the delivery time and in the end some features that were stripped never came back in the beta release.
Takeaways:
Practicing reducing a feature to an MVP version to see how little you can ship while delivering the same value is a good exercise. It helps highlight what the most important aspects are and what to focus on.