September 2026

Several years ago, Data Driven Districts (DDD) introduced an automated email that alerts principals when a learner has been absent for three or more consecutive days. In one case, school officials couldn’t reach a learner’s parents after the alert, so they contacted social services. A home visit found the child living with her grandmother, who didn’t know which school she attended. Social services stepped in, and the child went back to school.

That alert doesn’t use machine learning. But it is built on the principle behind every DDD Dashboard innovation that has followed: act early.

The cost of finding out late 

If a learner starts missing school in weeks one and two and drops out by week five, a termly report only reveals it around week 12. By then, the moment when a light-touch intervention would have made the biggest difference has passed.

The problem is one of timing and identification: the right nudge doesn’t reach the right person at the right time. And late information isn’t neutral. A well-resourced school can absorb a late warning. A school with one mathematics teacher across three grades can’t.

Identification is hard, too. In a large class, the top and bottom few learners stand out, while everyone in the middle blends together. Educators call this “the forgotten middle”. A learner drifting from a B to a C isn’t dramatic enough to catch a busy official’s attention. And in grades where each subject has a different teacher, no single person sees the whole learner. A child thriving in languages but quietly failing mathematics may not appear at risk to most of their teachers.

The DDD Dashboard draws on data that over 24 000 schools capture in SA-SAMS, the Department of Basic Education’s school administration system. That means it can see every subject for every learner. Over the past two to three years, we’ve been building machine learning processes into the DDD Dashboard so that officials can spot learners and schools that need extra support to either avoid underperformance or achieve excellence.

The kind of AI we use

“AI” is used for very different things, so it’s worth being clear. Our machine learning models are statistical: they learn patterns from the actual records of learners in previous years, then apply those patterns to current learners to produce forecasts. That makes them very different from generative AI , the kind of tool that writes text and can confidently invent facts that don’t exist. 

Like a weather forecast, these models can still be wrong, and they carry risks of their own, such as being less accurate for some subjects than others. We manage those risks carefully.

What we’ve built

Every tool is designed with Department officials, working from the top down: what we want to change, where (district, school, subject), who (the learner) and so what / now what (the action that follows).

• Term 4 Promotion Early Warning report (Grades R to 11) aims to reduce the number of learners who fail the year.
• Grade 6 Subject Insights and Early Warning report aims to help primary schools stay on track in terms of their targets.
• National Senior Certificate early-warning reports aim to get more Grade 12 learners ready for their final school-leaving exams and increase Bachelor passes, which qualify learners for university degree study. These are still under review and not yet live.

The first two reports to go live were each viewed by about 1 200 users in Term 2 of 2026. The Grade 12 reports are designed to go a step further: for every learner, they identify whether improving by one achievement level in one named subject could move them to a Bachelor pass.

That kind of precision is efficient, but it comes with a caution. Focusing on learners closest to the line can change outcomes quickly. If it becomes the only strategy, the learners furthest behind are left for later. Every plan built on a list of learners near a threshold needs a second list for those furthest behind.

What we’ve learned

Naming matters. We expected users to welcome “predictions”. For many people, though, a prediction sounds like a guess. The reports are now called early warnings, which is more accurate and more useful: you can’t change the weather, but you can change a learner’s outcome.

Build with users, and keep teaching. Officials shaped the designs from the start, so by the time we built them we could defend almost every decision. Training uses real examples in several formats and continues well after the first session.

People must be willing and able. An early warning is only useful if someone with the skills, time and authority acts on it. The 2026 independent review of the programme by the behavioural science organisation ideas42 found that 74% of officials said the dashboard has changed how they make decisions. But 48% reported only moderate or lower comfort using it, 21% lacked time to follow up and 29% had no or unreliable internet.

Understanding forecasts

Every forecast that the early warning reports provide carries a probability, as opposed to a certainty. A forecast can be wrong in two ways: it can flag a learner who is fine, or miss one who needs help. The second error is the more dangerous one, because it looks like reassurance. “On track” must never be read as “safe to ignore”. So we:

• say clearly, wherever it matters, that outputs are forecasts and early warnings,
• test models on data not used to build them, because anyone can forecast yesterday’s weather,
• retrain models as new data arrives, and
• don’t release insights from any model that falls below our minimum accuracy threshold.

We learned some of this the hard way. One early model looked highly accurate, until we realised it had been trained and tested on learners with three terms of data. In real use, we often have only one or two terms. We’re rebuilding it. Trust is something we have to keep earning.

Throughout, the decision stays with people. No decision about a learner is made by an algorithm. The early-warning tools are there to inform education officials.

Where to from here

We have another model in testing and are working to make every tool easier to understand and adopt. But the biggest constraint is no longer technical. It is the gap between intending to act and acting.

It will take time before officials check their dashboard forecasts daily the way they check the weather or the traffic. But more are engaging with these tools every term, and seeing how data can help prevent poor outcomes, or create excellent ones, rather than just explain them afterwards.

Which leaves a question for every education leader: what would it take for you to trust a forecast enough to change how you work?

The Data Driven Districts programme is implemented by New Leaders Foundation on behalf of the Department of Basic Education, supported by the Michael & Susan Dell Foundation.