Choosing What Stays Human in an AI World
The more capable AI becomes, the more I find myself appreciating distinctly human experiences, activities and ways of doing things. It might sound contradictory, given my enthusiasm for technology and the opportunities it presents, but I increasingly see these ideas as complementary rather than competing.
I am optimistic about AI and its potential to transform how we work and live. We are gaining access to capabilities that would have seemed extraordinary only a few years ago, and there is enormous potential to improve productivity, solve difficult problems and make things possible that previously weren't. Yet the pace of change is also making me think more carefully about what we value, where technology genuinely improves our lives, and where its use should be more deliberate or even constrained.
Over the summer, there was considerable discussion about AI-generated writing, and whether the growing ability to produce convincing text almost instantly diminishes its perceived value. If an article, report or personal message can be generated by AI, does that make it less meaningful? Does it matter whether a person wrote it, provided the final result is of equal quality or achieves its intended purpose? And where, if anywhere, should we insist that writing remains an entirely human activity?
I think the answer depends on what the writing is intended to achieve. A well-researched report should ultimately be judged on its accuracy, insight and usefulness, with appropriate accountability for its contents, regardless of the tools used to produce it. But writing is also a form of human expression and connection. A personal message, for example, carries meaning partly because someone has taken the time to think about what they want to say. Substantial AI involvement may change how we perceive that effort and authenticity, even when the words themselves are excellent. A 2025 study published in Nature Human Behaviour found that people judged identical expressions of empathy differently depending on whether they believed them to have come from a person or an AI system. Whilst the research concerned emotional support rather than writing more generally, it suggests that we value more than the words themselves.
There may be occasions when the human act of writing is itself fundamental to the purpose, and where we should expect someone to express themselves without delegating that responsibility to a machine. Elsewhere, the more important requirement may be that a person retains ownership, understanding and accountability for what is communicated. These distinctions will become increasingly important as the boundary between human and machine-generated work becomes less obvious.
I use AI for writing, and I also use a notebook regularly at work. Handwriting has practical benefits for me: it helps with memory, provides a simple way to capture thoughts and offers some welcome relief from screens. It is an analogue interface that works well for certain things, and I enjoy using it. I don't need to pretend it is universally superior to typing, or that writing everything by hand would somehow make me more thoughtful. It simply has a place, and I value having the choice.
Music offers another example. Streaming has given us extraordinary access to almost any music we could want, wherever and whenever we want it. I love that convenience, but I can also appreciate why vinyl continues to appeal to people. Digital audio can offer remarkable precision and consistency, while vinyl playback introduces its own physical character and imperfections. There is something distinctly tangible about a record, from the artwork and the act of handling it to the mechanical process of playing it. I have no particular desire to start collecting vinyl myself, but I can understand why people might value those qualities precisely because they offer something different from the convenience and technical precision of digital music.
These are small examples of a much broader question. As technology offers us increasing abundance, convenience and capability, how do we decide what is worth preserving? We are still physical beings, living in a world governed by nature and physics. Human connection, sensory experiences, imperfections and the simple act of doing something ourselves have value that cannot always be measured in time saved or output produced.
And I think that distinction matters just as much when we turn our attention to organisations.
Not all friction is waste
For years, digital transformation has often been framed as a process of removing friction: reducing clicks, automating handoffs, eliminating duplication and speeding things up. Much of that is sensible, but not every friction point exists by accident.
Some friction creates reflection. Some creates accountability. Some creates resilience. Some protects against error, and some creates meaning. A face-to-face conversation may be less efficient than sending a message. A handwritten note is slower than an email. A live performance is less reproducible than a recording. Those limitations can be precisely what gives them value.
The same is true inside organisations. A human approval step might be unnecessary bureaucracy, or it might be an important independent control. A manual process might be a relic that should have disappeared years ago, or it might provide resilience when everything around it fails. Good transformation depends on knowing the difference rather than assuming every manual step is simply waiting to be automated.
This is particularly important as AI expands the range of tasks and decisions that can be automated. The fact that something can be done by AI does not automatically mean that it should be. What we gain in speed or efficiency may sometimes come at the expense of understanding, accountability, reliability or human experience. Equally, retaining a manual process simply because it feels reassuring is no justification for keeping it. The challenge is to make these choices consciously, based on what we are actually trying to achieve.
Start with the work, not the technology
This thinking has also influenced how I see the next stage of FMD. There is already a crowded market of businesses promising to help organisations adopt AI, and that proposition is becoming too broad to be particularly useful. The technology is moving quickly, and capabilities that look specialist today may simply become standard features of existing software tomorrow.
The more durable challenge is not adding AI. It is redesigning how work gets done.
That starts with understanding the intended outcome. What does the organisation actually need to achieve, and what capabilities does it need to deliver that outcome? What are the functional requirements, and what standards of accuracy, reliability, safety and performance must be met?
From there, we can examine the processes, information and systems involved. Where does the data live? Who owns it, who has access to it, and can it be relied upon? How does information flow between people, teams and systems? Where are the handoffs, delays, duplicated activities and opportunities for error? It is surprising how difficult some of these seemingly basic questions can be to answer.
Only once we understand the work and its requirements can we make informed decisions about how to improve it. That might involve simplifying a process, improving data quality, changing responsibilities or redesigning how information moves through an organisation. Technology can then be applied where it genuinely improves efficiency, accuracy, capability or the experience of the people involved. This is broadly consistent with the UK Government's AI Playbook, which advises organisations to start with their goals and the needs of users, and to recognise where AI may not be the most appropriate solution.
AI is only one possible answer. Sometimes a conventional digital system will be more appropriate. Sometimes straightforward, rules-based automation will be more reliable, predictable and cost-effective. Other activities may benefit enormously from AI assistance or autonomous agents, while some should remain firmly within human control. These are not stages on a maturity ladder where the most technologically advanced option is automatically the best. They are different design choices, each with their own benefits, limitations and risks.
This is also why I am wary of simply bolting AI onto existing ways of working. Providing employees with powerful new tools can certainly generate productivity gains, but improving individual tasks is not necessarily the same as improving an end-to-end process. An employee might produce a report in minutes rather than hours, yet the organisation may still rely on the same inefficient information flows, duplicated checks and poorly defined responsibilities. Worse, the increased speed of production could overwhelm the next stage of the process or introduce new errors that take additional time to identify and correct.
A study involving more than 7,000 knowledge workers across 66 firms illustrates the distinction. Employees who used the AI tools spent approximately two fewer hours each week on email, yet researchers found no corresponding change in the overall quantity or composition of their work. The time saving was worthwhile, but it did not necessarily tell us whether the organisations themselves were operating more effectively. That requires looking beyond individual tasks to the processes and systems of which they form a part.
Without a wider view of how work is performed, organisations risk accumulating new technology without resolving the underlying problems. They may struggle to demonstrate a meaningful return on investment, maintain consistent governance or even understand how work is changing across different teams.
AI gives us an opportunity to rethink these arrangements in ways that were not previously practical. Rather than accepting existing processes as fixed, we can reconsider what work is genuinely necessary, how it should be organised and what combination of people and technology can deliver the best outcome. This is where I believe much of the lasting value of AI-enabled transformation will be found.
It will also demand better judgement. In a study of 758 management consultants, researchers found that AI assistance improved performance across a range of tasks, but reduced the likelihood of reaching the correct answer on another, more complex task. The experiment used an earlier generation of AI, but it illustrates a difficulty that organisations must take seriously: a system can be highly capable in one area whilst performing poorly in another that appears similar. We will need to become more capable not just at using AI, but at assessing when it is appropriate, what risks it introduces and when another approach would be better.
Choosing where technology stops
This becomes particularly important in regulated and safety-critical environments. As AI systems evolve from tools that generate information into systems capable of making decisions and taking actions, organisations will need to consider not just what these systems are capable of doing, but what they should actually be permitted to do.
Cybersecurity already gives us the concept of an air-gapped computer: a system deliberately isolated from wider networks. Safety engineering similarly relies on independent protection layers, while established governance practices use separation of duties and independent verification to reduce risk. These are not new ideas, but I think they deserve renewed attention as AI becomes more capable.
The National Cyber Security Centre's guidance on operational technology provides a useful example. It recommends that monitoring and analytical systems should not have direct control over the equipment they are intended to observe. The distinction is important: being able to monitor, analyse or advise on an activity does not necessarily mean a system should have the authority to carry it out.
Perhaps we should extend this thinking to air-gapped processes, applying these principles across processes and operating models so that certain actions, decisions or controls remain independent of AI systems and cannot be directly executed or bypassed by them. This would not necessarily involve literal network isolation in every case, but the deliberate design of boundaries and safeguards where independence matters.
The NCSC's advice on managing the cybersecurity risks of agentic AI, published in August 2026, is relevant here too. It recommends considering how much autonomy an AI system actually needs, restricting the environments and resources it can access, and maintaining technical safeguards alongside human oversight where the consequences of failure are significant. There is good reason to apply these established security principles more widely as AI takes on increasingly complex activities.
This could be particularly valuable in defence, healthcare, infrastructure and other environments involving safety-critical equipment or decisions with potentially serious consequences. As the capabilities of AI expand, there is a strong case for reviewing these systems and their associated processes to identify where additional separation, independent verification or physical controls could reduce risk.
That doesn't necessarily mean returning everything to paper or installing big red buttons everywhere, although there may be circumstances where physical controls are entirely appropriate. It might mean requiring independently verified human authorisation for a critical action, keeping certain control systems inaccessible to AI agents, or ensuring that particular steps require physical presence, a physical key or the movement and verification of a tangible object.
AI systems may become increasingly capable of identifying vulnerabilities and exploiting weaknesses in connected systems. But a properly isolated physical control presents a different kind of barrier. An AI agent cannot simply transmit a command across a network to turn a physical key that has no network connection. It would need some other means of influencing the physical world, which introduces additional constraints and opportunities for control.
Of course, physical safeguards are not inherently infallible. People can be deceived, keys can be stolen and poorly designed controls can be circumvented. A human approval is not much of a safeguard if that person simply accepts whatever the AI recommends. The important principle is genuine independence: ensuring that critical controls cannot be defeated by the same system, or through the same failure, that they are intended to constrain.
In some settings, the inability of a system to perform an action may itself be an important safeguard. As our technologies become more capable and interconnected, deliberately limiting their reach could be just as important as expanding their capabilities.
Choosing intentionally
None of this is an argument against technology. It is almost the opposite. The more powerful our tools become, the more deliberate we should be about where and how we use them.
AI presents extraordinary opportunities to remove unnecessary administrative work, improve services, extend human capabilities and create entirely new ways of working. It could free people to spend more time solving difficult problems, exercising judgement and connecting with one another. But those outcomes are not inevitable. Poorly considered adoption could just as easily introduce new risks, unnecessary costs, additional work or unintended harm.
For organisations, I think the more useful starting point is therefore not simply asking where AI can be introduced. It is defining what needs to be achieved, understanding the work required to get there and determining how people, information, processes and technology can best be brought together to deliver it.
This requires us to be as confident in choosing where not to use technology as we are in identifying opportunities to adopt it. It means recognising that good design is not always about removing every constraint, maximising automation or eliminating human involvement. Sometimes the best outcome depends on retaining exactly those things.
The same principle can guide our personal choices. We can embrace the convenience and possibilities of digital technology while still making room for experiences that are slower, physical, imperfect or simply more meaningful because another human being is involved.
In a world of increasingly abundant digital intelligence, knowing where to remain deliberately human may become one of the most important choices we make.