NextBillion - A WDI Publication

In most of the international development and social impact organisations I work with as a strategic AI advisor, the actual use cases of artificial intelligence are not normally based on formal policies or other top-down guidance. Rather, AI is more commonly used via ad hoc workarounds, a practice sometimes termed “shadow AI use.”

Staff are regularly using free, publicly available AI tools to summarise evaluation reports, translate documents, draft proposals, tidy up their field notes and prepare donor updates. Some do this carefully. Others do it with little sense of how outputs should be checked, where their own human judgement must be used, and what types of training data should never be uploaded.

Leadership responses to this off-the-books AI use vary. Some leaders I’ve worked with know it is happening and look away. Some suspect it and hope the question resolves itself, or may even embrace staff experimentation with AI tools in the absence of an organisational policy. A few others have restricted or banned staff use of AI tools.

That last response deserves attention. In a sector rightly alert to technology’s risks, restricting or even banning staff use of AI tools may feel like the responsible position for many development and impact organisations. In practice, it is often the opposite. Saying no does not keep AI out of an organisation; instead, it keeps AI use invisible, and invisibility is the one condition under which none of AI’s risks can be managed.

The AI ‘Governance Vacuum’ in Global Development

This issue isn’t merely anecdotal. In a major survey by the Humanitarian Leadership Academy and Data Friendly Space, covering 2,539 humanitarian workers across 144 countries, seven in 10 reported using AI daily or weekly, while fewer than a quarter of their organisations had any formal AI policy in place. The same survey found that almost two-thirds of organisations offered staff little or no AI training, even as most staff used the tools regardless. The researchers referred to the resulting misalignment as a “governance vacuum”: individual use running well ahead of the institution’s capacity to guide it.

The timing makes this lack of guidance harder to ignore. Official development assistance from DAC members fell by 23.1% in 2025 to US $174.3 billion, the largest annual contraction on record, with core contributions to the UN system down 27% and bilateral programming cut deeply. Organisations are being asked to protect service delivery with fewer people and tighter budgets, as well as to justify every cost while doing so. Meanwhile, AI tools keep getting cheaper, more capable and easier to access. Anyone with a browser can use one, and many staff already do.

So the question of whether these organisations will adopt AI was settled some time ago, without fanfare, one prompt at a time. What remains unsettled is the governance that should underpin this growing AI use. On one side sit staff with heavier workloads and rising expectations. On the other sit boards and senior leadership teams that still tend to treat AI as a future strategy question, a procurement decision or a reputational risk to be contained. Between them sits the real-world organisation: people using AI tools informally because they help, while policy, training and accountability trail behind.

The Three Downsides to Restricting AI

None of this is an argument for rushing in, no holds barred, to adopt AI. Organisations operating in the international development and social impact sectors handle sensitive data, work with vulnerable communities and operate in charged political contexts. Badly governed AI usage can result in exposed data, fabricated evidence, biased analysis, and weak conclusions concealed by false confidence. These risks are real. But addressing them will require clear safeguards and informed oversight rather than disengagement.

That’s because AI restrictions and bans are weak forms of protection and carry costs of their own. Three stand out.

The first falls on staff. When an organisation avoids the issue, every judgement call devolves to individuals. The programme officer using AI to summarise a 90-page impact evaluation must decide alone whether the summary can be trusted. The fundraiser drafting a concept note must decide alone what AI outputs are safe to paste in. The monitoring officer experimenting with qualitative analysis must work out, unaided, whether the AI tool has flattened what respondents actually said. These are hard questions even for specialists. Organisations that refuse to engage with them are not sparing staff the burden; they are transferring that burden to them, without providing the rules, training or backup that could help staff navigate it.

There is an equity dimension here too. When organisations provide neither practical guidance nor safe routes to experiment, staff are left to make their own calculations about AI’s risks and rewards. More confident users may save time, but they may also expose sensitive information or rely on plausible-sounding but erroneous outputs; more cautious colleagues may avoid those risks, but also miss legitimate opportunities to work more effectively. The result is neither fair nor controlled. Individual confidence and risk tolerance rather than organisational priorities and consistent safeguards determine who uses AI, how they use it and who bears the consequences.

The second cost involves organisational learning. The most useful applications of AI in social impact and global development work are rarely the controversial ones. They sit in the unglamorous middle of these organisations’ work: synthesising lessons across projects, making institutional knowledge findable, drafting first versions of routine documents, comparing proposals, and pressure-testing assumptions before a decision is made. These workflows are precisely the ones many organisations struggle to resource, and precisely where AI can help without displacing expertise or automating anything sensitive.

Leadership teams that either avoid AI or respond to staff AI usage mainly by restricting tools are unlikely to discover this middle ground. They also lose the chance to build the judgement and safeguards needed for higher-stakes uses through supervised practice on low-risk work. If organisations defer that learning, they may eventually face a consequential AI decision with no institutional experience to draw on.

The third cost lands on partners and communities. All social impact organisations — from large UN agencies to grassroots NGOs — worry, rightly, about extractive or carelessly used technology. But unmanaged AI is more dangerous than visible, bounded AI. When staff paste a partner’s internal documents into public tools with no data rules, when community feedback is summarised without anyone checking for lost meaning, or when AI-polished language smooths the uncertainty out of a report, the risk is carried by the people least able to see or challenge the process. Workplace AI bans prevent none of this. They simply guarantee that nobody is watching for it.

Some organisations have grasped this. Mercy Corps developed internal generative AI chatbots grounded in its own digital library, giving staff a safer way to use organisational knowledge while reducing the risk of sensitive data being pasted into public tools. The United Nations Development Programme (UNDP) took a more specialised approach: Its Artificial Intelligence for Development Analytics platform gives staff and the public multilingual, source-linked access to evidence from nearly 7,000 evaluation reports. Launched in 2022 and substantially upgraded since, more than three quarters of users reported that it made evaluation evidence easier and quicker to use, although UNDP acknowledges that uptake remains uneven.

Guidance for Responsible AI Adoption

Responsible adoption starts from a different premise: AI is already inside the organisation, so leaders first need to understand how it is being used. Strategy, policy and procurement still matter, but they should respond to real workflows and risks rather than preclude any investigation of them.

It begins with mapping real staff use. This should be a practical exercise to establish where staff already use AI: which tools they prefer, what information they’re feeding in, and which outputs they’ve come to rely on. This exercise will highlight both where staff are successfully using AI, and where they feel out of their depth. The aim is not to shame experimentation. Rather, it is to make invisible practice visible enough to govern.

The next step is to select a small number of existing or proposed workflows for formal, supervised use. Strategy, policy and procurement should inform that choice, but they should not substitute for examining the work itself. Good starting points are low-stakes, internal tasks grounded in source material and easy for a person to verify — for example, summarising public documents, preparing internal meeting notes, compiling donor information from published sources, or checking a human-written draft against agreed-upon criteria such as factual accuracy, required content and donor rules. These are useful entry points because staff can practise source discipline, verification and human review before using AI with sensitive data or consequential decisions.

Alongside this, staff need simple data boundaries. Public information, internal operational material, personal data, partner documents and community-level data should not be treated alike. Most teams can begin with plain-language categories, concrete examples of what staff may and may not upload in common situations, and a clear route for advice when the answer is uncertain. They also need defined human review points. The higher the stakes, the stronger the requirement that a person checks sources, tests assumptions and makes the final decision. AI should be treated as a tool that can accelerate parts of a workflow while producing errors plausible enough to escape casual review — not as a colleague whose judgement can be trusted.

Finally, boards and senior leaders need enough AI literacy to govern its use. They do not need to become technologists, but they must be able to: distinguish between controlling access to tools and controlling the risks within particular workflows; understand how data protection, model error and human review requirements vary by use case; and judge whether staff have the training and approved tools to work safely. They should be asking: Do we know where AI is already used? Which tools and data are permitted? Where must a person verify sources or make the final decision? And who owns the quality, compliance and consequences of each AI-assisted workflow?

This is a manageable agenda. But it requires leaders to move from avoidance to stewardship.

The central question for impact-driven organisations is no longer whether or not AI should be used — that ship has sailed at most organisations across the sector. The challenge now is to determine where AI can responsibly improve their work, where it should be kept out, and what staff and leadership capability is needed to tell the difference. Saying no may feel prudent, given the organisational risks this technology can bring. But if prohibiting AI has the practical effect of increasing unmanaged use, fostering uneven staff practices and weakening oversight, it is not prudence. It is risk in another form, borne by the people the sector exists to serve.

Loksan Harley is the Founder and Principal of Homelands AI.

Photo credit: wildpixel