The human judgement advantage.
A practical field guide for charities and humanitarian teams: what the evidence supports, how quality compounds, and where human judgement must hold. Research, worked calculations and a ninety-day operating plan.
AI adoption is easy to count. Dependable use is harder.
An organisation can buy an AI licence in an afternoon. Knowing what that licence is doing to the organisation takes more work. Which decisions does it influence? What information can it reach? Who notices when an answer is plausible, polished and wrong?
Stanford's 2026 AI Index reports organisational AI adoption at 88%, following 78% in 2024 and 55% in 2023. These are survey-based measures of use, not a census of humanitarian organisations or a measure of safe deployment. The distinction matters: being able to say “we use AI” tells a board remarkably little about whether the work has improved.
My starting point is a work decision. A donor-care colleague needs to explain an authorised project update. A field officer needs a discrepancy resolved. A programme lead needs a proposal ready for review. Each has an intended outcome, a person accountable for it and a cost when it goes wrong. Start there, and the technology becomes easier to judge.
This issue gives you a practical operating method: choose a bounded workflow; define the information and authority it needs; test it against meaningful errors; equip the people who review it; and expand only when the evidence supports the next step. The charts serve that argument. Some report external research. Others make the arithmetic or a decision framework visible. Their captions tell you which is which.
For your next leadership meeting, bring one concrete example of an AI-supported task already happening. Trace its input, output and owner. If that takes longer than finding the organisation's AI strategy, you have found a useful first piece of work.
One useful habit is to label every number in a decision paper as a measured result, a survey response, an estimate or an assumption. Add its date and denominator. These four categories deserve different weight. When somebody asks what has changed, you can then separate a new fact from a more optimistic interpretation of the same evidence.
Read a productivity claim all the way to its denominator.
A figure such as “15% more productive” is useful only after you know what was measured. More words written, more cases closed and less time spent are different outcomes. None automatically establishes better service.
Brynjolfsson, Li and Raymond studied the introduction of a conversational assistant using data from 5,172 customer-support agents. Their revised paper reports a 15% average increase in issues resolved per hour. Less experienced and lower-skilled workers improved speed and quality; the most experienced and highest-skilled workers had small speed gains and small quality declines. This is evidence from a particular support operation, not a promised improvement in donor conversion.
The operational lesson is to look beneath the average. A new donor-care colleague may benefit from finding an approved explanation quickly. An experienced colleague may lose time correcting a generic suggestion that overlooks a long relationship. Give both the same interface and the same performance target, and the average may conceal the problem you most need to see.
An illustrative calculation helps. At 20 resolved issues per hour, a 15% rise gives 23. Time per issue falls from three minutes to about 2.61 minutes: a 13% reduction, not 15%. Throughput and time are reciprocals. Neither calculation includes the quality of the resolution or the cost of a later correction.
Ask a supplier for five things beside the headline: the population studied, the task, the comparison, the quality measure and the distribution across users. For your own pilot, retain a comparison group or baseline and count reopenings, corrections and escalations. The credible claim is the one you can explain without the adjective “transformative”.
Try this in the next supplier meeting: ask for the result for new staff, experienced staff and difficult cases separately. Then ask what happened to quality when throughput rose. A trial can support a narrow purchasing decision without supporting every claim in a sales deck. Keep the narrow claim; it is often enough to justify a well-scoped experiment.
Feeling faster and being faster can diverge.
People's experience of a tool matters. It is…