by JL Ecochard
A few years ago I led a design session at SOS Children’s Villages headquarters in Vienna. We were building Rafiki, an AI care assistant for parents and caregivers. It was the first NetHope IDEA Journey to put end stakeholders in the room from the start. Not just program staff. Not just funders. Alumni of SOS Children’s Villages. People who had grown up inside the system the tool was meant to serve.
Three years later I reconnected with the project. By then Rafiki had traveled from Vienna to field teams to caregivers across Africa. And a child in an SOS village in Kenya had said something that stopped the room.
“Something built for me without me is not for me.”
She was not critiquing the technology. She was naming the reason most humanitarian AI fails before the first line of code is written.
Rafiki achieved a 97% satisfaction rate in its Malawi pilot. That number did not come from better technology. It came from a decision made in Vienna to put the right people in the room at the beginning.
We keep framing this as a Global North problem. It is not. It is a headquarters problem. Secretariats build for field teams. Large INGOs build for the community-based organizations doing the actual last mile. Adults build for children. The gap compounds at every layer. By the time a tool reaches the community it was designed for, it has passed through four or five layers of people who were never asked what was needed.
This is what NetHope Connected is trying to fix. Not the global gap, the organizational one. A peer learning space built for sub-national, national, and regional nonprofits, the ones with 5 to 200 staff, no tech team, and no seat at the table where the tools get designed. Ground-truthed experience, not theory. Built with them, not for them. NetHope is the catalyst, the center of power is the community.
This matters more than most people realize.
A search engine can only surface what was already recorded. The digital record of the world reflects the people who had the means to create it, which is to say not the communities we are here to serve. Most of the world’s knowledge was never digitized. And what was digitized was not representative. And that is digitized and retrievable for AI training is not the original story, it’s often the fairytale.
AI generates. It can produce in languages and contexts that were never fully represented in its training data. Imperfectly. Inconsistently. But generatively. The space of what is possible is not bounded by what was previously recorded.
Meheret and NetHope Members did not find Amharic translation in any video-conferencing tool documentation. They found it in a meeting, by accident, when someone tried something that was not in the manual. If we never try Amharic, we never discover Amharic works.
AI is not a CRM. Not an ERP. It is more like Excel. Nobody deployed Excel. People discovered it, misused it, repurposed it, and it became the connective tissue of every NGO on earth. Not by plan. By curiosity.
The capabilities that will matter most for the communities we serve are waiting to be found. Some dead ends are the price of admission. The gems are worth it.
Go be a kid in the AI sandpit. Imagine it can do it and try. Be surprised and share when it works.
Indonesia and Kenya did not build copper wire before going digital. They skipped it entirely. Went straight to mobile and fiber. Arrived at digital parity in one infrastructure generation.
The same moment is happening right now with AI compute. And the sector is making the wrong call. LMIC organizations are being given the cheapest CPUs sufficient to do the job, decisions made by donor-oriented procurement policies. Entry-level laptops. Minimum viable devices to support a modern workplace set of tools. Nobody is framing GPU access as the new digital infrastructure. Nobody is making the leap.
CPU is copper. GPU is fiber. The organizations being given the minimum device today are being denied the infrastructure jump that could put them on par with anyone, anywhere.
NetHope was founded in 2001 to democratize connectivity. The next chapter is to democratize compute. Shared regional inference infrastructure, hotAIspots, that give local NGOs and CBOs access to GPU power they cannot afford individually. Not incrementally better laptops. The right infrastructure from the start.
The session ran fast, running out of time but not interest before lunch. Table discussions were short. But the popcorn of responses at the end named something worth holding.
The room’s aspiration was not “make AI work locally.” It was bigger than that.
Overcome poverty with AI. Not point solutions for discrete problems. Systemic change. The possibility, unproven, yet worth taking seriously, that AI is the first technology with enough generative capacity to address root causes rather than symptoms.
And the condition they named for that ambition: different AI for what is best. Not one model deployed everywhere. Not the famous Wintel model for the CPU age. Different approaches, different governance, different architectures, chosen by the people closest to the experience, not handed down from headquarters or from a vendor’s roadmap. And ideally safely run air-gapped in the places that matter by stakeholders, with full agency and self-determination.
Agency to the edge. Ownership by the community. Continuity that outlasts the funder’s grant cycle and the vendor’s business model.
Think of the last AI tool your organization deployed. Name every stakeholder group in the room when it was designed.
Now name the ones who were not.
If the people closest to the problem were not there, you did not build local. You built elsewhere.
Go find them. Start over if you have to.
The 97% is on the other side of that conversation.
JL Ecochard is Chief Innovation Officer at NetHope and co-founder of its Center for the Digital Nonprofit.
This is Part 3 of a four-part series developed from his session “When Every Hour Counts” at the Good Tech Summit, April 2026. Part 1: Fund it like roads, not meals. Part 2: Nobody Goes to Jail. Part 4: The Brain Drug Argument.
