Ideas
These are the questions rather than the answers — themes that develop across several pieces, and that I keep coming back to.
-
Does AI governance make safe experimentation easier, or drive it underground?
Boards need visibility and accountability for AI use, but governance that only restricts staff may create more shadow AI rather than less.
-
What does an AI leader actually do differently on Monday morning?
-
What should organisations actually be trying to achieve with AI?
A question about how many jobs AI might replace opened a wider debate about service redesign, workforce choices, who benefits and what counts as value.
-
If AI makes the finished artefact a weaker signal of capability, what should we assess instead?
As AI makes CVs, applications, presentations and take-home work easier to polish, assessment may need to move closer to the work itself: how someone frames a problem, uses tools, challenges output, makes decisions and stands behind the result.
-
Are our funding and accountability models pushing AI towards cost cutting rather than better services?
Short-term budget pressure can reward immediate efficiencies while making the investment and organisational change needed for better outcomes harder to justify.
-
What if the most efficient team isn't the best team?
AI may allow smaller teams to produce the same or greater output, but teams also create learning, challenge, resilience, belonging and organisational memory. Those benefits are harder to measure than productivity and easier to lose.
-
When software becomes abundant, what becomes scarce?
If AI makes producing software dramatically easier and cheaper, the important organisational constraint may move away from writing the software and towards deciding what should exist, why, and how it becomes useful.
-
What happens when AI accelerates threats faster than organisations can change their defences?
AI could reduce the time and effort needed to find and exploit vulnerabilities. The leadership question is whether security, continuity and recovery can adapt at the same pace.
-
Have we made our business cases too narrow to recognise some of AI’s most valuable benefits?
Financial returns matter, but conventional business cases can struggle to value wellbeing, judgement, service quality and the avoidance of unnecessary work.
-
When AI gives people time back, who decides what happens to it?
Productivity gains do not automatically become organisational benefits. The decisions about how freed capacity is used are managerial, cultural and sometimes political.
-
If models become interchangeable, where does durable organisational AI capability accumulate?
The durable capability around AI may sit less in the model itself and more in the context, permissions, tools, workflows, evaluations, controls and operating knowledge that allow models to work safely and repeatedly inside a real organisation.
-
What changes when it becomes cheaper to find out whether an idea works than to prove in advance that it will?
AI can reduce the cost of experimentation, rework and changing direction. That means organisations can sometimes buy certainty through small, reversible experiments rather than demanding it all before work begins.
-
As execution becomes cheaper, where do we need clarity and where can we afford uncertainty?
AI reduces the cost of implementation, experimentation and rework. That can lower the amount of certainty needed before trying something, while increasing the value of clarity about the problem, hypothesis, boundaries, learning goal and consequence of being wrong.
-
If AI does more of the doing, how do inexperienced people become experienced?
Some work that looks routine or low-value also develops judgement. If AI removes that work, organisations need to replace the learning rather than simply remove the task.
-
Does AI need human-like intelligence to become consequential, or is capability without judgement already enough?
A system can become extremely effective at pursuing an objective without understanding why the objective matters, which assumptions were implicit, what collateral effects matter, or when it should stop.
-
If building becomes cheap, when does an experiment become something the organisation is prepared to own?
AI makes it easier to build and test software, which should lower the barriers to responsible experimentation. But permission to experiment is not the same as permission to operate: services still need ownership, support, security, resilience and accountability.
-
When AI becomes infrastructure, how much of its cost and resilience is determined by architecture rather than procurement?
At enterprise scale, AI consumption is shaped by system design: model calls, context size, routing, retries, agent loops, evaluations and fallbacks all become cost and resilience decisions.
-
When does measuring AI adoption stop being useful?
AI usage can be a useful transitional signal while organisations are building capability, but the enduring measures should be what changed: service quality, time, cost, reliability, user experience, learning or some other meaningful outcome.
-
Can you have rigorous performance management without first having a rigorous understanding of performance?
Organisations need to address genuinely poor performance, but relative ranking can create an appearance of precision without resolving harder questions about outcomes, context, comparability, collaboration and what good performance actually means.
-
Before we automate work, have we understood why the work exists in the first place?
Technology can make a good operating model dramatically better, but it can also make a poor one faster. Friction and workarounds are sometimes evidence about the system that should be understood before they are automated.
-
What does expertise become when AI can increasingly perform the execution?
As AI takes on more execution, expertise does not simply disappear. More of its value moves into understanding the problem, setting direction, decomposing work, challenging outputs, recognising failure and exercising judgement about what should happen next.
-
If technology makes delivery dramatically faster, can the functions that make delivery safe keep up without becoming the bottleneck?
As AI reduces the cost and time required to build and experiment, security, data, architecture, assurance and other enabling functions need to evolve too. Good governance should make responsible action easier, not depend on friction being slow enough to control behaviour.
-
If AI changes what an organisation can feasibly do, should the strategy focus on AI or on the possibilities that have opened up?
The strategic significance of AI is not simply that organisations can now use AI. It is that activities which were previously too expensive, slow, complex or impractical may become feasible. Strategy should examine those changed constraints rather than turn use of the technology itself into the objective.
-
Why are organisations better at celebrating what they launch than what they successfully remove?
Decommissioning rarely produces a visible new feature, but it can reduce cyber exposure, failure modes, support burden, architectural complexity and technical debt while making future change easier. Technology leadership needs to recognise removal as value creation, not merely cleanup.
-
If public services are never really finished, why do we still organise so much change as though they are?
Government is often organised around a distinction between running services and changing them. Funding, business cases, programmes and spending decisions can make bounded change easier to describe and govern than permanent improvement capability. As digital services and AI make incremental change cheaper and more continuous, that distinction may become increasingly awkward.
-
Will the most useful AI agent be the one we trust enough to let in?
As agents become capable enough to act, their practical usefulness depends on context, permissions and delegated authority. Trust is part of the architecture: identity, boundaries, auditability, approvals and revocation determine what an organisation is willing to let an agent do.
-
How do people discover useful AI capabilities when they do not know what to ask for?
Technologists need to understand how capabilities work, but service owners and users also need help discovering what is possible. A blank prompt can conceal useful Copilot-style workflows. Capability discovery should begin with real problems and outcomes, not a standalone AI unit hunting for use cases.
-
Are we measuring successful transformation, or merely successful transformation programmes?
Transformation is often judged within a leader's tenure, while consequences for institutional knowledge, maintainability, apprenticeship and resilience can emerge much later. AI may widen the gap between immediate delivery output and the organisational capability needed to sustain it. Examine the incentives and governance that make capability renewal visible alongside quick wins.