Published · September 26, 2026
By Dan Haile, Founder & CEO, AgentifyAfro.ai
“In English, leading AI speech recognition systems make errors less than 6 percent of the time. In Yoruba, that same system fails more than 60 percent of the time.” That sentence, from the Gates Foundation’s Goalkeepers 2026 report, is the clearest statement of the problem I have read this year.
On 21 September 2026, 60 initial signatories, including the Gates Foundation, announced a five-year goal to close it. In the words of the joint commitment: “Within the next five years, an estimated 3.4 billion people who speak languages currently under-represented in today’s AI models will be able to use AI tools in their own language and voice.”
At AgentifyAfro, we welcome this commitment without reservation. The problem it names is the problem our company was founded to address — and seeing it named, at this scale and by this many institutions, is a moment worth marking.
A goal the size of the problem
Roughly 7,000 languages are spoken worldwide, and today’s leading models serve a small percentage of them. The imbalance is structural: according to the Goalkeepers report, “More than 90 percent of the data used to train early, large language models came from English-language sources.” The joint commitment states the consequence plainly: “Low-resource languages are under-represented in the data, tools, and benchmarks used to build and test AI and AI-enabled systems.”
The coalition convened around this goal is global. Its publicly named signatories include African research communities and builders such as Masakhane, Lelapa AI, Data Science Nigeria, Digital Umuganda, Qhala, the Ethiopia AI Institute and the Rwanda AI Scaling Hub; model developers including Google, Microsoft, Anthropic, NVIDIA and the OpenAI Foundation; Mozilla Data Collective; and public institutions including UNICEF, the World Bank Group, IDRC Canada and the UK FCDO.
Bill Gates put the stakes in human terms in the signatory statements: “It’s important that AI tools work for everyone in the languages they actually speak … Right now, even some of the better-supported African languages have significant gaps. When AI tools can’t understand somebody’s natural language, the cost of a mistake can be serious.”
The problem we were built to solve
AgentifyAfro was founded in 2025 around exactly this gap. I started the company because I watched AI systems fail in African contexts — not because the technology was wrong, but because the training data was never built with Africa in mind.
Our founding premise has not changed: AI is advancing globally. The data infrastructure, evaluation rigour, and delivery expertise to make it trustworthy — across every language, culture, and context — is not. We built a hybrid model to close that gap, USA Strategy · Africa Execution, on a principle we hold to in every engagement: we are not outsourcing to Africa. We are building with Africa.
AgentifyAfro is a young company, and a small one beside the institutions that signed this commitment. That is what makes this moment such an opportunity: a global foundation, frontier AI developers, African research communities, multilaterals and governments are now committed to the problem we set out to address. We are grateful, and eager to contribute what we have built to a shared cause. The joint commitment says it best: “No single company, country, or foundation can do this alone.”
The workstream that decides whether the others succeed
The commitment sets out four workstreams to advance within the next year: building the open language layer; turning language data into working tools; reaching people safely, “guided throughout by responsible practices that protect privacy, consent, and data sovereignty”; and one that every other workstream depends on — “Tracking progress honestly: assessments and benchmarks that measure real gains against the global goal.”
The same point recurs in the signatory statements. Temina Madon, Co-Founder of The Agency Fund: “AI evaluations will play a critical role in making this technology work better for all of us.” XD Huang, CTO of Zoom, described the commitment as “a decision to act — with urgency, in coordination, and with a willingness to measure progress honestly.” And the Goalkeepers report is explicit about what the work requires: “investment in local language datasets and evaluations tailored to how people use language in the places these tools are meant to serve.”
Building the data is the start. Measuring whether it worked is the proof.
Enormous resources will now move into language data and models. That is exactly right. But data alone does not tell anyone whether a model now works for a farmer in her own language, a patient describing symptoms, or a citizen navigating a public service. Evaluation does. And evaluation is not a one-time audit; it is a cycle — evaluate, diagnose, improve, re-evaluate — that turns investment into verified progress.
What an average can hide
Published research makes the gap concrete. On IrokoBench — a human-translated benchmark spanning 17 African languages, published at NAACL 2025 — GPT-4o, the best-performing model tested, scored 86.9 in English but averaged 59.0 across the African languages: a gap of roughly 28 points on the same tasks.
Those results describe 2024-era models on one research benchmark, and a different task from the speech recognition figures in the Goalkeepers report. The lesson carries across both: an aggregate score can look acceptable while the people a system is meant to serve are left behind. That is why the AfroEval Scorecard™ reports results by user group rather than as one averaged number.
A number averaged across all users can hide the people the whole effort is meant to reach.
The Scorecard measures how an AI system performs for African users and in African contexts — language, culture, hallucination, bias, and safety — before it reaches production rather than after. Its findings are presented as governance-grade evidence a board, regulator or auditor can read. It is delivered as a service, and teams can request an evaluation briefing today.
Every contribution needs its counterpart
The commitment is clear that there is no single way to take part: “Partners will contribute in fundamentally different ways. Some create open datasets, some build benchmarks and scorecards that track progress, some integrate open data into models, some fund the work, some build and deploy the applications…”
Each of those contributions creates a need AgentifyAfro is built to meet:
- Dataset builders — data collection and annotation, and human-in-the-loop review, delivered with African teams.
- Labs integrating open data into models — evaluation and benchmarking, bias detection, safety testing and red-teaming, reported by user group.
- Funders and governments — governance-grade evidence of whether an investment reached the people it was meant to serve.
- Teams building and deploying applications — pre-production evaluation, and customer-experience AI through AfroCX, independently assured by AfroEval Scorecard™.
An open invitation we take seriously
The joint commitment extends an open invitation: “We invite community organizations, labs, researchers, companies, governments, and funders to join us and bring their individual resources and expertise to this work.” The coalition’s detailed structure, governance and workstreams will be developed collaboratively over the coming year.
AgentifyAfro exists for the outcome this commitment describes, and we want to contribute to it — with our evaluation practice, our data and human-in-the-loop capability, and our teams across the United States and Africa. None of it is a plan for later. It is operating now.
To every signatory, lab, researcher, builder, funder and government working toward this goal: we would welcome the conversation. Whether the need is honest measurement of progress, language data built with the communities who speak it, or independent evidence that a deployed system works for the people it serves, we are ready to work alongside you. Reach us at info@agentifyafro.ai, or connect with us directly on LinkedIn.
3.4 billion people is the right ambition. Measured honestly, by user group, in the places these tools are meant to serve, it is also an achievable one.
#GatesFoundation #AIForEveryone #RepresentativeAI #AIEvaluation #AfricanAI
