The last responsible handover

Imagine an ordinary Monday, not a technology showcase. A colleague asks an AI assistant to summarise field notes. The summary becomes a recommendation. The recommendation enters a funding paper. A manager approves it because the prose is clear and the deadline is close. The model never made the final decision, yet it shaped the evidence that reached the person who did.

That sequence contains several handovers: from people to records, records to a model, model output to a reviewer, and review to action. Each handover can lose context or gain authority. This issue's thesis is simple: AI governance succeeds or fails at the handover. A policy can be sound while a real service remains unsafe because nobody defined what must travel with an output, who may challenge it or who owns the next consequence.

Current sector surveys make the operating gap visible. In a January 2026 humanitarian pulse survey of 1,729 self-selected respondents across more than 120 countries, 95.3% said they had used AI at work and 75% used it daily or weekly; only 23.3% reported a formal organisational policy, 9% described wide integration and 3% expert integration. These are perceptions from respondents, not a census or a measure of impact.

Charity data points in the same direction: among 807 responses in the 2026 Charity Digital Skills research, 38% described active or strategic adoption, while 41% were exploring. The sample was weighted towards small charities and question bases vary.

The practical move is to govern the route, not admire the tool. Take one live workflow and draw every handover from source to consequence. Write beside each arrow: what evidence moves, what can be changed, who may stop it and who answers to the person affected. That is where accountable AI begins.

Handover test. Print the real workflow on one page and ask a frontline colleague to circle every moment where evidence or authority changes hands. Add any spreadsheet, message or judgement the official process omitted. Choose the handover with the largest consequence and assign one owner to make its evidence visible this month.

Evidence discipline. The visual reports external evidence with its source, date, denominator or scope limitation. It should not be read as a causal estimate outside that stated context.

A licence count is easy to report. An operating capability is harder. It includes a named purpose, authorised data, tested failure modes, trained reviewers, incident routes and an exit plan. Without those elements, adoption can grow while the organisation's ability to explain or correct the work stays still.

The 2026 charity research illustrates the movement without proving maturity. Thirty-four per cent described active use and 4% strategic use, up from 22% and 3% respectively in 2025. At the same time, reported barriers increased: skills rose from 43% to 56%, output quality from 33% to 40%, trust from 15% to 35%, and internal resistance from 20% to 34%; privacy held at 43%. Greater concern may reflect more exposure and better questions, not necessarily deteriorating tools.

For leaders, the denominator matters. The research collected 807 responses between 5 March and 20 April 2026; 570 were from small charities and 211 from large charities. Individual questions have different bases. The percentages should inform a conversation, not become a population forecast.

Build a two-column register. On the left, list where people use AI. On the right, list where the organisation has deliberately integrated it. A use moves right only when its purpose, owner, data boundary, review rule, incident route and withdrawal test are recorded. This creates a truthful portfolio: experimentation can continue, but it is not mislabelled as a controlled service.

The leadership question is not “How much AI do we have?” It is “Which AI-supported decisions can we defend today?” If the answer is unclear, reduce authority before adding features.

Handover test. Sample ten current AI uses and classify each as personal experimentation, team practice or controlled organisational service. For anything labelled…