Where using AI and deploying it diverge

Judgement 1 · Human capital

Staff already have frontier AI at home and on their phones, and that gap, not the technology itself, is where most of the next two years of risk sits.

Confidence: High

Generative AI arrived in employees' hands as a free, general-purpose utility before most organisations had anything to offer in response, reversing the order in which earlier general-purpose technologies were adopted. Electrification took roughly four decades, from the 1880s to the 1920s, to move from a novelty retrofitted onto steam-age factories to something organisations had fully redesigned their processes around, and the pace of that adoption was something organisations controlled, through investment in machinery and retraining. Generative AI offers no such runway. Employees already have the capability. The organisation is still deciding what to do about it.

Alternative explanation: a blanket ban might seem to close the gap by removing the tool. In practice, attempts to outlaw use altogether tend to serve a leadership team's sense of control more than they reduce risk: a workforce committed to its work will use the free resource in front of it, sanctioned or not, and the people most likely to do so are often the most capable and resourceful staff in the building. The sounder response draws a line by consequence, not by tool: light governance for general drafting and summarising that a person reads before acting on it, and a far more deliberate path for anything that touches a decision about a person, moves money, or acts before a human reviews the output.

Judgement 2 · Partnerships

Access to frontier AI is a dependency, not an asset, and that exposure does not end when an export ban is lifted.

Confidence: High

In June 2026, the United States government blocked Anthropic's most capable models, Fable 5 and Mythos 5, from use by foreign nationals anywhere, allies and Anthropic's own foreign staff included. The ban was lifted a little under three weeks later: Fable 5 was restored to everyone, Mythos 5 to vetted partners inside a dedicated program. Three weeks is a short outage, and a stark lesson. Access to a capability this consequential can be withdrawn and reinstated on terms set unilaterally, by a government other than your own. The organisations and countries positioned to absorb a shock like that are the ones that already hold the three assets no vendor or government can issue or revoke: their own data, their own people, and their own capacity to deploy, execute and scale what they build.

Judgement 3 · Governance

Where responsibility for an AI system is spread across enough roles to sound universal, it resolves to nobody, and the organisations that avoid this assign a named owner before anything is deployed.

Confidence: Moderate

A case we use in briefings makes the pattern concrete: a cancer-screening model trained almost entirely on one demographic, because that was where the clean, abundant, well-funded data happened to exist. It performed brilliantly in local testing and was deployed more broadly. Accuracy then diverged sharply by demographic once it reached populations the training data never represented. Put the question "who was responsible?" to a room, and every honest answer names a different function: the engineers who built it, the team that curated the training data, the finance function that declined to pay for more of it, the sales team whose performance claims nobody had verified for the new markets, or the organisations that bought and deployed the tool without further validation. When responsibility is spread this evenly, it is not shared. It disappears, and failure eventually lands on whoever is furthest from the decisions that caused it, which is usually the executive team and the board.

A useful discipline borrowed from military doctrine is the order of battle: everyone in an operation knows, by name, what they own and what they are allowed to decide, paired with delegating authority to the lowest competent level within clear bounds. Applied to AI, that means drawing a line by consequence: general-purpose use governed lightly at the population level, and anything that touches a decision about a person, moves money, or acts without a human reading the output getting a named sponsor, a named technical lead and a named product owner before it goes anywhere near production.

Alternative explanation: Robodebt offers a contrasting case. It was automated income averaging, not AI, but it remains the most thoroughly examined automated-decision failure in Australian public administration. Its lesson was that accountability reconstructed after a failure, through a Royal Commission and years of subsequent investigation, is not the same as accountability assigned before deployment. By the time it has to be reconstructed, the organisation is no longer governing the system. It is investigating the wreckage.

Judgement 4 · Process

A governance plan updated for the new org chart after every restructure can still describe an organisation that no longer exists in practice.

Confidence: Moderate

We have watched a capability degrade this way more than once. A small team owns a system end to end, with one reporting line and direct access to users. To professionalise it, the team gets folded into a central function. Nothing breaks immediately. Specialist staff move on, the feedback loops that used to run informally stop running at all, and within a year the people notionally responsible for the system can no longer tell you what it recommended last week or whether anyone acted on it. None of this needs a bad actor. A modern organisation run by humans tends toward silos, and headcount is what lets a function exist and win funding, so every part of it responds rationally to the incentives in front of it. Updating a governance document to reflect the new org chart, as good practice requires, changes none of this. The document describes the chart, not the way the work is actually getting done.

Judgement 5 · Process

The fix for a decaying control is not a better document. It is sunlight: regular, shared evidence of what the system is actually doing.

Confidence: Moderate

The single intervention that would have caught the drift in Judgement 4 is unglamorous: a standing, shared account of what a model recommends and how often the human decision-maker agrees with it, reviewed on a regular cadence rather than produced once for an audit. That record does two things at once. It keeps the team that built the system honest about its performance, and it gives the area that depends on it the material to exercise its own judgement over individual cases. The numbers are only half of it. The people on the frontline hold risk knowledge no dashboard captures, and the discipline that actually works is talking to them regularly and treating what they say as evidence, not the quarterly status colour a board is used to seeing.