YearBeyond: AI-assisted pathway reports for youth service partners

The problem

Year Beyond is a youth service programme delivered in partnership with the Western Cape provincial government, which co-funds and supports its operations alongside the national government and private funders. The programme places approximately 3,500 unemployed young people aged 18 to 25 into 10-month work experience placements each year across South Africa's Western Cape, serving a core public policy objective: reducing youth unemployment by moving young people out of NEET status (not in education, employment, or training) and into education, employment, training or economic activity.

The programme combines structured work experience with a weekly curriculum covering personal and professional development, emotional intelligence, and technical skills, delivered by mentors who work with groups of 20 to 25 young people. Year Beyond delivers this through approximately 30 local NGO partners on the ground, each operating in different contexts, from rural and agricultural areas to peri-urban communities and metropolitan centres with tourism or retail-driven economies.

The organisation collects outcome data on every cohort, including a tracer study that tracks what young people go on to do after the programme ends. A persistent challenge was how to feed this data back to partners in a meaningful, timely, and contextualised way. With 30 partners operating across very different local economies, each needed reporting tailored to their specific context. Producing individualised reports for each partner was a substantial undertaking, and with limited staff capacity, it typically took about 3 months. The reports also needed to go beyond summary statistics and engage with the economic realities of each partner's area to be genuinely useful for planning post-programme support for young people.

The solution

Year Beyond used Claude, a generative AI tool, to dramatically accelerate the production of individualised pathway reports for each partner organisation.

The process involved uploading the anonymised programme dataset into Claude, along with several layers of contextual information: youth tax data segmented by municipality from the SA Tax Data, prior examples of reports the organisation had produced, details of the Year Beyond programme itself, and relevant research papers on youth development and pathways into employment.

The tax data, drawn from a publicly available university-affiliated dataset, provided a picture of formal employment sectors at the municipal level. While not a perfect indicator for a population that often works informally, it served as a contextual guide, showing partners what the formal economic activity in their area looked like and where opportunities might exist for the young people they support.

After significant iteration and refinement of prompts, Claude generated reports that were substantially more detailed and contextualised than what the team had previously been able to produce manually. Rather than generic summaries, the reports walked partners through thinking pathways specific to their local economic context. For a partner in an agricultural area, the report might flag logistics as a relevant skill set. For one in a tourism-driven community, it might highlight customer service. The AI was not perfect at pinpointing local conditions, but it provided a starting point that partners could build on with their own knowledge.

All statistics, tables, and outputs in every report were manually reviewed for errors, hallucinations, or miscategorisation before being shared. Each report was also edited by the team.

The organisation was transparent throughout. The reports were clearly identified as AI-assisted, and the process was explained to directors and partners. The team framed the AI's role as a tool for synthesising and integrating data at scale, not for independently generating conclusions.

Year Beyond is now exploring how to connect its data systems more directly to Claude, so that as attendance and outcome data are recorded in Airtable (the platform the organisation uses to manage its data), that information can flow automatically into the AI tool without needing to be manually exported and uploaded each time. This would allow reports to be generated on a rolling basis, potentially every week or month, rather than only at the end of each programme cycle.

The impact

1. Report turnaround reduced from three months to approximately two weeks
What previously took a graduate intern roughly three months to produce was completed in approximately two weeks. By mid-April, shortly after the March data cut-off, individualised reports had been delivered to all 30 partners. The organisation had never achieved that kind of turnaround before.

2. Reports were more detailed and contextualised than previous versions
Instead of surface-level summaries, each partner received a report that engaged with the economic context of their specific area and suggested practical pathways for the young people in their community. Directors at the partner organisations reported that the reports were more useful than previous versions and arrived earlier, rather than on the day of their scheduled meetings.

3. Partners began collecting more targeted data in response
Because the directors understood how the AI process worked, they began identifying what additional data, such as local asset maps showing which employers and institutions they had connections to, could be fed into the model to produce even better reporting in future cycles. The process created a feedback loop that improved the quality of the data going in, not just the reports coming out.

Lessons learned

• Be transparent about AI use, even when it feels risky. There was an internal debate about whether to disclose that the reports were AI-generated, with concern that it might undermine their credibility. The team decided to be upfront with directors and partners, explaining the process and framing the AI as a synthesis tool rather than a replacement for human analysis. This transparency built trust rather than eroding it.
• AI is not a silver bullet, and learning to use it well takes time. The team invested significant time in learning how to use AI effectively, including watching tutorials, iterating on prompts, and understanding the tool's limitations. The quality of the output depended directly on the quality of the input and the prompting process. Treating AI as a third person in a meeting of two, rather than a replacement for either, was a useful frame.
• Data protection concerns require active decisions, not assumptions. All individual-level data were anonymised before uploading to Claude, but questions arose about whether it was appropriate to include partner organisation names, locations, or cohort sizes. There were no clear-cut answers, and the team had to make judgment calls about what constituted reasonable use. These questions were discussed openly rather than ignored.
• Scepticism often comes from poor first experiences with AI. Many colleagues were sceptical because their own interactions with AI had been superficial, asking broad questions and getting poor results. Demonstrating what was possible with a more structured, well-informed approach helped shift perceptions.
• Invest in the skill set, not just the subscription. The ability to use AI effectively sits at the intersection of data literacy, programme knowledge, and prompt design. As a nonprofit delivering a government-funded programme, the question of who takes charge of this work is real, and the organisation is still navigating how to build and sustain that capacity internally.

South Africa's Protection of Personal Information Act (POPIA) was a consideration throughout. All individual-level data was anonymised before uploading to Claude, and the process was discussed openly with directors and partners, including what information was and was not shared with the AI tool.