Last week, the global education community was in New York for the United Nations General Assembly (UNGA). Across the week, one word ran through the programme more than any other. Sessions asked what it takes to scale early childhood innovation, what role funders should play in scaling with government, how programmes can be locally built and designed to scale. The Global Partnership for Education made its financing case at the UN under the banner of multiplying possibility. In the corridors, over coffee and at receptions, the question was the same: what works, and how do we make it bigger?
I share the ambition. The learning crisis is too large to be solved one project at a time, and the only route to reaching most children is through public systems. But I came home with a nagging sense that we are often trying to scale the wrong thing.
We scale programmes. We should be scaling what we learn from them, and the advocacy for the systems that need to change.
Why programmes rarely travel
A programme is a set of choices that worked in one place, at one time, with one group of people, inside one set of institutions. It carries all of those conditions with it, most of them invisible. Move it to a different district, a different ministry, or a different political moment, and many of those choices stop fitting.
The clearest evidence I know of comes from Kenya. Researchers tested the same contract teacher programme twice, on identical terms, once run by a non-governmental organisation and once run by government. Under the non-governmental organisation, test scores rose. Under government, there was no measurable effect at all. The model was the same. The politics, the incentives, and the way it was implemented were not, and those turned out to matter more than the model.
Teaching at the Right Level shows what happens when that lesson is taken seriously. When Pratham and the Abdul Latif Jameel Poverty Action Lab (J-PAL) first worked with state governments in India to bring the approach into public schools, it did not work. In Bihar and Uttarakhand, teachers were trained, liked the method, and did not use it. Rather than insisting on fidelity to the original model, the teams went back to why, and redesigned for each system. In Haryana, that meant a protected hour in the school day, supported by mentors from within government. In Uttar Pradesh, it meant intensive learning camps, which produced some of the largest learning gains on record. What spread across India, and later into Africa, was less a fixed programme than a growing understanding of what it takes to make the approach work inside a particular system.
Luminos in Ethiopia has followed a similar path. Ethiopia's Ministry of Education adopted the approach as national strategy, and when regional education bureaus asked for changes, Luminos made them. I will be in Ethiopia next month for Luminos's tenth anniversary, and I suspect the story of those ten years is as much about what was learned and changed as about what was replicated.
The learning also travels much further than any one programme. What Pratham found about protected time and mentoring inside government is useful to a numeracy programme in Malawi, a teacher coaching scheme in Brazil, or an early years initiative in Bangladesh, none of which share its model. What BRAC learned about the gap between a successful trial and a model that works everywhere applies to almost every funder and implementer in the sector. A programme is tied to its context. What we learn from it can be used across very different programmes, contexts, and geographies, but only if it is shared.
The silent L
Monitoring, evaluation, and learning (MEL) is now standard language in almost every grant agreement. In practice, the monitoring and the evaluation get the budget, the indicators, and the reporting deadlines. The learning is too often silent.
Part of the reason is who the learning is for. Most evaluation is designed to answer the funder's question, which is whether the grant did what it promised. That is a fair question, and funders are right to ask it. But it is not the question a head teacher, a district education officer, or a ministry official needs answered four months into implementation, while there is still time to change course. Accountability looks backwards. Learning has to look forwards.
The other reason is timing. A review of more than a hundred impact evaluations found that the average study was published more than four years after the final data was collected. By then the next phase has usually been designed, the next grant approved, and many of the people who needed the findings have moved on. The evidence arrives, rigorous and carefully written, into a decision that has already been made.
This is not a criticism of evaluators, many of whom have been making the same argument for years. It is a description of how the system is funded. We pay for answers at the end of a grant cycle, and then wonder why so little of what we learn shapes what happens next.
There is an uncomfortable consequence. When learning is silent, bad ideas are often the easiest to scale. They tend to be simple, visible, and easy to count: devices handed out, teachers trained, schools reached. In rural Peru, One Laptop per Child raised the number of computers per pupil roughly tenfold. A large randomised evaluation found no effect on maths or language test scores. Good ideas are usually harder to scale, because they depend on people changing how they work, which is slow, uneven, and hard to photograph. If scale is the goal and counting is the measure, we will keep rewarding what spreads fastest rather than what works.
What BRAC is willing to say
There is a further reason the learning stays silent, and it is the one the sector discusses least. Very little of what goes wrong is ever shared.
BRAC has just published the sixth edition of its Failure Report, and I always look forward to it. Each year it takes a few of its own programmes that did not work and explains plainly what went wrong, what it misread, and what it changed as a result. This year's report includes BRAC International's clubs for adolescent girls and young women, which grew to 1,800 clubs across five countries after a randomised controlled trial in Uganda showed impact across whole villages. In 2019, an unannounced check of a random sample found that more than half of the clubs effectively did not exist, even though monitoring reports had looked fine for years. BRAC says openly that it read the trial as proof that the model worked everywhere, when the trial had shown that it worked under certain conditions and with the right supervision in place. It rebuilt the model, which now forms part of a wider programme aiming to reach two million adolescent girls and young women by 2033.
That single example makes the argument of this piece better than I can. A strong trial, a confident scale-up, monitoring that measured the wrong things, and a correction that only came because someone went and looked.
Very few organisations publish anything like this, and fewer funders still. Yet foundations hold more comparative learning than almost anyone in the field, across grantees, countries, and years, and most of it stays in board papers and internal reviews. Imagine if the grants that did not work, and the assumptions that did not hold, were shared as openly as the successes. The field would stop paying for the same mistakes in different countries, and grantees would find it far easier to be honest with the people who fund them.
Learning at the speed of decisions
All of this matters more now than it did a decade ago. Budgets change overnight. Aid is being cut. Technology is changing what is possible in a classroom within a single school year. Conflict and climate are closing schools with little or no warning. A learning system built around four-year answers cannot keep up.
The question that matters is not whether something worked, asked years after the fact. It is what we need to know today to make the next best decision. During my time at the LEGO Foundation, we worked on exactly this with Brink. We called it learning horizons: being clear about what you need to learn, and by when, to make the decision in front of you, rather than waiting for a final report to tell you what you should have done.
None of this is new thinking. Ten years ago, Brookings argued in Millions Learning that those running programmes rarely get the chance to pause, reflect, and adjust based on what they are learning. That remains largely true, and the reason is not a shortage of ideas. It is the way we fund.
Scaling the advocacy, too
Learning on its own is not enough. The contract teacher programme in Kenya did not fail in government because anyone lacked evidence. It failed because of politics, incentives, and institutional resistance. Teaching at the Right Level took hold where state governments changed timetables, created mentor roles, and gave teachers permission to teach differently. In each case, what decided the outcome was whether the system around the programme changed.
Those changes, to budgets, to teacher policy, to curriculum and assessment, to how ministries are held to account, do not happen because a programme works. They happen because someone makes the case for them, inside government and outside it, over years. That is advocacy, and it is the part of the work education philanthropy funds least. I have written before about education's advocacy gap. It applies here too. If we want what we learn to reach children at scale, we need to scale the advocacy for the system changes that learning points to, not just the programme that produced it.
What funders would need to do differently
I sit on the boards of grant-making foundations and advise others, and I recognise every one of these habits from the inside. Changing them does not require new money. It requires spending the same money differently.
The first change is to pay for learning during implementation, not only at the end. That means funding the people and the time it takes to collect evidence, discuss it, and act on it while a programme is running, and treating that as core cost rather than overhead.
The second is to agree, at the start of a grant, what we need to know at three, six, and twelve months, and which decisions each answer will inform. Learning horizons, set together, give grantees permission to adapt and give funders a reason to listen.
The third is to treat a change of course based on evidence as a result, not a deviation. Too many grant agreements reward fidelity to the original plan, which is another way of rewarding people for ignoring what they have learned.
The fourth is to make sure the learning reaches government first. If public systems are how children are reached at scale, then the officials running those systems need the findings more than we do, and in a form they can use.
The fifth is to fund the advocacy alongside the programme: the national coalitions, researchers, and campaigners who can turn what we learn into changes in policy and budgets.
And the last is to publish more of our own failures. Funders ask grantees for candour. We could offer some in return.
Scale what travels
The ambition in New York last week was the right one. The children who most need better learning will be reached through public systems, at scale, or not at all. But programmes will keep meeting contexts they were not designed for, and the organisations that reach the most children will be the ones that learn fastest, adapt best, make the case for the system changes that learning points to, and say openly what did not work along the way.
That means all of us, funders, implementers, researchers, and governments, getting much better at sharing what we learn, including our failures, so that what one organisation discovers in one country can shape decisions in many others.
At every grant review, one question would tell us whether our learning is real: what did we change because of what we learned? If the honest answer is nothing, the L was silent.
The programmes that scale are the ones that keep learning.