AI Adoption in the Impact Sector and the Trust Gap 

Trust is the impact sector’s working capital. People hand over refugee locations, medical histories and records of violence because they believe the organisation collecting them will protect them. That trust takes years to build and can evaporate in an afternoon. 

Which is why Sally Freeman (Research Ireland)’s line from the stage at the Google.org Impact Summit in Dublin stayed with the room: innovation at the speed of trust. If the people affected cannot verify how something was built, it will not scale – and arguably it should not. 

Research from EY and Changemaker Exchange found over 90 percent of organisations in the sector are now using AI. Twelve percent have an AI policy in place. 

As Rowan Emslie of the Centre for Future Generations argued in Dublin, the impact sector runs on being believed, and AI has made it almost free to sound believable.  Grant makers and funders who once received dozens of reports a month now receive hundreds a day, with no reliable way to separate the genuinely good from the merely plausible. One funder is already operating a lottery for grants below £50,000 rather than continue sorting through applications that are all, on paper, good enough. 

The predictable response is retreat. Funders and policymakers fall back on the organisations they already know, which closes the door on precisely those that needed it open, and narrows the range of ideas the sector gets to draw on. 

The train has already left the station 

The Conduit ran three AI for Good sessions across Dublin and London this month, around the Google.org Impact Summit: two debates and a breakfast with nonprofit chief executives and corporate social impact leads. What we heard was clear – the barriers to scaling AI in this sector are structural, not technical. 

Most of that 90 percent are deploying AI for basic use cases – think email creation and simple summarisation.  Administration can swallow between 30 and 40 percent of a nonprofit’s time, and over half of those now using AI report saving five hours a week or more. The thing taking that burden away is free, immediate, and needs nobody’s approval. 

Some organisations have been deliberate about this. Most have not, and for them AI arrived without anyone deciding anything, against a backdrop of reduced funding and fewer heads to carry the workload. What separates the two is rarely appetite or ability. It is whether anyone had the time, the standing and the remit to own the question. 

The barriers 

Ownership is the first. Who is accountable when an AI-powered initiative fails to deliver what was promised? 

The practical version came up in Dublin, when the chief executive of a 23-person NGO asked who her AI lead was supposed to be. 

Operations, an external hire, a consultant? Nobody in the room had an answer. The champion employment model everyone recommends works at 150 staff members. Yet most of this sector has fewer than 50 employees, meaning the small number of AI advocates lack the guidance to drive meaningful or structured impact. 

That gap decides whether adoption is deliberate or accidental. Fast Forward’s survey of around 120 AI-powered nonprofits worldwide found 64 percent naming a champion driving the vision as the single biggest driver of adoption, ahead of any technology factor. 

Underpinning this gap is the reality that the data are not yet ready, because they are often incomplete, inconsistent, and captured differently by every organisation. Governance cannot be applied to information nobody has organised, and no amount of AI funding fixes that on its own. 

Which is why treating AI as something entirely without precedent does not help either. This sector has absorbed new technology before, and the pattern of digital transformation is familiar enough by now: donor databases, the shift to cloud systems, digital fundraising, service delivery that moved online almost overnight in 2020. In each case the technology itself turned out to be the straightforward part. What took the time was deciding who owned it, redesigning the work around it and bringing people with you – and that is precisely the work being skipped this time round. 

What it looks like when it works 

Barnardos in Ireland is the clearest example of doing this in the right order. Their AI use sits almost entirely on the administrative side – communications, policy and research papers – alongside pattern analysis across the data their own services collect, which flags where families are struggling to access a welfare payment or a local doctor so that the organisation can raise it with the provider directly.  

They drew their line before scaling anything: frontline work with families stays as human contact, because a child becoming identifiable would be catastrophic for an organisation whose relationships are its entire basis for operating. Having settled that question, they moved quickly everywhere else, taking meeting minutes from ninety minutes to fifteen and turning a ministerial budget speech into a published press release within ninety minutes of it ending. 

What needs to happen 

What is missing is practical rather than philosophical: ways of deploying AI that are bounded, observable and reversible from the first day, so that a smaller organisation can begin without betting the one thing it cannot afford to lose. 

Most of it is also cheap. Enterprise licences rather than free accounts, which would close much of the data problem on its own and is a pricing decision sitting with a handful of providers. Funding that treats AI as a line item rather than overhead, which some funders are now pushing for. Money for data foundations attached to the money for AI itself. A low-cost resource telling charity boards which questions they ought to be asking. Shared technical support, so that several hundred organisations are not solving the same problem separately. 

And champions with the time and the remit to do the job, rather than the goodwill to do it on top of everything else. 

The gap 

There was dissent worth recording. Most frontier model companies are not yet profitable, and embedding their tooling now risks locking small organisations into pricing that is unlikely to stay cheap once shareholders arrive. Why, and to what end, was a question left hanging over the room, and it deserves a better answer than the sector has managed so far. 

The most useful observation came last from someone who described themselves as an optimist before making it. Whatever else AI does in this sector, it cannot build trust. 

Which is the whole point. Adoption that is fast but undecided spends the sector’s working capital. Adoption that is deliberate protects it, and the difference between the two has very little to do with technology. It is whether organisations have the time, the money and the standing to make the decision at all. 

The goal was never faster adoption. It is a sector with a meaningful say in how this gets used. 

Katie McPhee is The Conduit’s AI for Good Expert in Residence