Five Tests Every Inclusive AI Initiative Must Pass
Prateek Shrivastava | July 13, 2026 | Reporting from AI for Good 2026 in Geneva, Switzerland.
I spent three days at the AI for Good Global Summit in Geneva, where more than 11,000 participants from 169 countries gathered alongside the first United Nations Global Dialogue on AI Governance and the World Summit on the Information Society Forum. The summit showed how rapidly artificial intelligence (AI) is advancing across health, education, agriculture, finance, science and the creative industries. It also reinforced the influence of the institutions that finance, purchase, regulate and distribute these technologies. Model capability matters, but institutions determine which applications reach people, which companies secure customers, and where the resulting economic value accumulates.
Geneva helped sharpen five tests that I now use to assess any AI initiative focused on development and inclusion.
1. Does it solve a priority problem for underserved people under real-world conditions?
Does a solution solve a priority problem for underserved people under the conditions they actually live and work in: affordable, in the right language, functional on the connectivity and power supply that exists, not the infrastructure planners wish existed?
President Paul Kagame and Marc Benioff launched the AI for Good Global Commission. It brings together political leaders, technology companies, investors and international institutions. Its mandate covers access, trust and the use of AI for sustainable development.placed the scale of the challenge in clear terms. Around 2.2 billion people remain offline. Many more live with unreliable connectivity, limited access to computing, scarce local-language data and institutions that are still building basic digital capacity.
Sangbu Kim, the World Bank’s Vice President for Digital Transformation, offered a practical direction through the concept of “Small AI.” These are purpose-built systems designed to operate with modest computing resources, intermittent connectivity and locally relevant data.
These conditions shape the design of useful AI. Products that depend on constant connectivity, expensive devices and large cloud-based models will have limited relevance in many of the markets where the need is greatest. Examples include offline tools that help clinicians prepare medical notes, WhatsApp-based tutoring, AI-supported tuberculosis screening and agricultural diagnostics that allow farmers to identify crop disease from a photograph. These systems work within services, devices and workflows that people already use.
This approach has clear implications for funders and technology providers. Language, affordability, energy, connectivity and device access should influence product design from the beginning. Investment decisions should also account for distribution, integration, training and long-term support. Reach depends on the full delivery system surrounding the technology.
2. Is there a credible route from pilot to institutional adoption and scale?
Is there a credible route from a working pilot to institutional adoption and scale: real customers, procurement processes, distribution, and capital suited to the slow sales cycles of regulated buyers?
The Innovation Factory and other summit programs showcased entrepreneurs working in payments, health, mental health, climate resilience and water management. Nearpays, an African fintech that turns ordinary smartphones into payment terminals, was among the ventures recognized in Geneva. Prizes and competitions provide visibility, validation and early funding. Sustainable growth requires institutional customers, procurement pathways, access to data and capital that reflects the realities of adoption.
Public agencies, banks, healthcare providers and development institutions often have slow procurement processes and limited capacity to evaluate emerging systems. Founders can spend years moving from a successful pilot to a repeatable contract. This weakens promising companies and delays access to potentially useful technology. Ten financial institutions exploring AI for underwriting, fraud prevention or customer service can create a more credible market by working together than by running ten unrelated experiments. This gives entrepreneurs clearer demand signals and gives investors stronger evidence about performance and scalability. Capital becomes more effective when it is connected to customers, distribution and a credible route to adoption.
Impact investors can help address this gap by aggregating demand across their portfolios and networks. Institutions facing similar challenges can define shared requirements, conduct joint diligence, establish responsible procurement standards and test solutions in comparable settings.
3. Can it be trusted, governed and challenged in practice?
Can a system be trusted, governed and challenged in practice: with evidence, human oversight, monitoring, and a working route for recourse when it gets something wrong?
Amazon CTO Werner Vogels used his keynote, “Trust is dead: Trust me,” to describe how generative AI has changed the economics of deception. Convincing synthetic text, images, audio and identities can now be produced quickly and at very low cost. Microsoft President Brad Smith addressed the same challenge from a policy perspective. His position reflected the need for rules that protect society while remaining practical enough to support innovation and adoption. The ITU’s new Focus Group on Trust and Identity for Humans and Agentic AI provides one concrete institutional response. Its work will examine how autonomous systems are identified, governed and held accountable as they begin to take actions and interact with other systems.
These questions have immediate relevance for financial services, healthcare, education and public administration. A bank using AI in underwriting must understand the data, testing methods, risk of bias and role of human judgment. A government agency must establish accountability and recourse when an automated system influences access to a public service.
Responsible AI should be visible in operating processes, procurement contracts and board oversight. Institutions should document data sources, decision rights, human controls, monitoring, escalation procedures and user recourse. Trust depends on evidence, accountability and the ability to intervene when systems fail.
4. Does it build local capability and distribute value fairly?
Does the system build local capability and distribute value fairly, to the creators, communities and local entrepreneurs whose data, knowledge and work make it possible, not only to the company that deploys it?
Björn Ulvaeus, co-founder of ABBA and president of CISAC, brought the economic rights of creators into the centre of the summit. He argued for collective licensing arrangements that allow artists and other rights holders to participate in the value generated from works used to train commercial AI systems. ABBA Voyage provided a practical illustration. Technology expanded the creative experience while preserving the artists’ consent, participation and economic rights.
The same principle extends across sectors. AI systems draw on knowledge, language, behaviour and data generated by workers, customers, communities and public institutions. Inclusive AI requires clear rules for how these inputs are collected, governed and monetized.
Investors and boards should examine the data supply chain behind every AI company. This includes consent, provenance, representation, intellectual property, community rights and the distribution of economic returns. A business that depends on poorly governed or uncompensated inputs carries legal, reputational and social risk. Fair participation strengthens legitimacy and long-term commercial resilience.
5. Are outcomes measurable and are institutions accountable for delivering them?
Are outcomes measurable, and is someone, a company, an investor, a government, a commission, a development institution, accountable for delivering them, not just for announcing them?
Amandeep Singh Gill, the UN Secretary-General’s Envoy on Technology, emphasized the need for global and evidence-based participation in AI governance. Rachel Adams reinforced the importance of meaningful representation for smaller states and lower-income economies. Nobel Peace Laureate Kailash Satyarthi brought the discussion back to human dignity and the protection of people with the least power.
These perspectives create a clear expectation for institutional accountability.
The same standard should apply across the wider AI-for-development field. Conferences, accelerators, principles and training programmes all have a role. Their contribution should connect to adoption, institutional capacity and measurable improvements in people’s lives.
The work after Geneva
Progress requires coordinated action. Governments and institutions can define priority problems and establish procurement pathways. Investors can aggregate demand and finance adoption. Technology providers can design for local conditions and demonstrate trustworthiness. Global bodies can publish standards, commitments and results. AI for good will be judged through deployment. Progress will become visible when useful systems are purchased, governed and scaled by capable institutions, and when the people they are intended to serve participate meaningfully in the benefits.
These tests guide the work of the Alliance for Inclusive AI and Accendo’s engagement with impact investors.
A lot of work needs to be done. And we’re already started running the marathon!
