Voluntering missions easier to publish with AI

Increasing our missions volume with the same moderation team

JeVeuxAider.gouv.fr is a civic engagement platform connecting volunteers with associations. Associations publish missions for volunteers to apply on. Every mission is moderated by our internal team so we had a main challenge: how do we keep up with the increasing pace of published mission? Without having a bad impact on mission quality and delay to be published.

Our main goal was to improve the mission quality before they are sent to the moderation team, to avoid back and forth messages between moderators and associations.

What are the main paint points when moderating a mission?

The Product Manager and I went through this first step to understand what are the main reasons a mission is not published quickly. To answer this, we had several means:

• internal notes are published by moderators each time they need to ask for changes to an associations
• support tickets
• interviews with moderators

We identified these:

1. Some missions description are not detailed enough to understand what the volunteers will do
2. Some missions are not configured correctly and have inconsistencies (for example, different addresses in the content and in the specific address field)
3. Some missions titles do not fit our guidelines
4. Some missions include discriminative criteria about the volunteers (mainly about age), and we do not allow this

How do we solve this?

We had several solutions and iterations in our mind. Regarding these pain points, we audited out mission publishing form to identify what we could improve:

1. Clarify the level of information needed for the associations
2. Help users with our titles guidelines
3. Inform the association with potential issues or possible improvements BEFORE they publish the mission to be moderated

What we did

Aftet we changed wordings in the form to be as clear as possible to help the associations to understand our requirements, we developed two AI features to answer our pain points:

1. A complete content analysis on the mission, to display advices to the association, so they know what they should change or improve: lack of information, inconcistencies, discriminative criteria, etc.
2. Titles suggestions written from the mission description

While doing this, we also developed something to improve the overall quality of our missions: a feature to improve the mission description formatting and writing. It helps the association, make them earn time and it can potentially have an impact on the conversion rate to get more applications.

A word on the process

From the moderator teams feedback, I started writing a prompt to analyse a mission content. The output of the promt is a list of advice for the associations. We tested this manually at first, one mission by one mission. Then by batch with Basalt. We identified to many hallucinations so I decided to add a second prompt. So one prompt for the first analysis, and a second for the quality controle. The second prompt accept or reject the advice. After this, we decided to test it in production, internally only. With these new tests we continued to improve the prompts and then go live.

Design wise

Impact

Great user feedbacks for titles and writing. Average feedbacks for the analysis: we still have use cases that are not well handled by the prompts.

We still have to improve this. Next time it will be way easier and faster: less manuel tests, more batch tests with speciic AI tools.