Assistance, judgement and the human in the loop
Where AI accelerates creative and technical work, and where it should not decide.
The wrong question
The interesting question about AI in creative and technical work is not whether it will replace the people doing that work. It is what changes about how the work gets done, now that a fast, capable assistant is available at almost every stage of it. That question has a practical answer rather than a philosophical one, and it is worth answering properly rather than settling for either the hype or the panic.
What genuinely accelerates
Four stages of a project speed up more than any other: research, exploration, drafting and testing.
Research used to mean a day spent reading around a problem before writing anything useful. Now, a first pass through a regulatory framework, a competitor landscape or an unfamiliar codebase takes an hour, not because the reading itself matters less, but because the summarising, cross referencing and pattern spotting happen far faster.
Exploration benefits even more. A visual identity, a page layout, an interaction flow: these used to be expensive to prototype, so fewer options got tried before one was chosen. Now several genuine, working versions of an idea can exist before lunch, which changes the decision from whether one thing works to which of several actually serves the brief.
Drafting is the most obvious case. A first version of copy, a first pass at a component, a first cut of an architecture diagram: AI is fast at producing a plausible draft, and a plausible draft is a genuinely useful thing to react to, even when most of it gets rewritten before anyone else sees it.
Testing and implementation gain too, quietly. Generating test cases, checking an interface across viewports, tracing through an edge case before it reaches a user: this is work that used to be skipped under time pressure and now gets done, because it costs so much less.
None of this changes what good means. It changes how quickly something can be put in front of the part of the process that judges good, which is where the real work still lives.
A problem that resists automation
Solution architecture is a useful example, because it is the part of my own work AI helps with least. The job is to take a business problem that a client has not yet fully articulated, often because it sits across several departments, systems and incentives at once, and work out how the pieces of software, data, infrastructure and process should actually fit together so the idea works end to end rather than just on a whiteboard.
That translation step, between what a business is actually trying to achieve and what a workable system looks like, depends on context an AI does not have: the politics of a stakeholder group, the real reason a previous project stalled, which stated requirement is a genuine constraint and which is a preference dressed up as one. AI can help document the options once that context has been gathered and understood. It cannot gather or understand it, and pretending otherwise produces architecture that is technically coherent and practically wrong.
A second example: finding a voice
Copywriting shows the same pattern from a different angle. AI is genuinely good at producing fluent, correctly structured sentences on almost any subject, and that fluency can be mistaken for the job being done. It is not. Fluent is not the same as true to how a particular business actually talks, or judged correctly for a particular audience's patience, or honest about a claim the business cannot really back up.
Getting a brand's voice right is a matter of taste built from context: what this client's customers already believe, what tone earns trust in this specific sector, where confidence tips into overclaiming. AI can generate ten fluent variations of a sentence in seconds. Choosing which one is actually right, or noticing that none of them are and the sentence needs a different idea entirely, is still a human read, made by someone who has sat with the brief long enough to know the difference.
Where judgement still has to hold
Two kinds of decision stay entirely human, and I do not expect that to change soon.
The first is anything with a real consequence attached: a regulated claim, a piece of financial advice, a system that touches somebody's money or their medical history. My background is in the Lloyd's and London insurance market, where the cost of getting a detail wrong is never hypothetical, and that discipline carries into every project regardless of sector. Where a brief touches personal data, compliance or clinical information, the work has to be explicit about what needs specialist legal, security or clinical input, and AI has no part in making that call. It can help draft the question worth putting to a specialist. It cannot be the specialist.
The second is accountability itself. Every project I take on is led and delivered personally, from concept through to build, rather than handed between departments or tools along the way. That is a structural choice as much as a stylistic one: when responsibility for a decision is spread thin, quality erodes quietly and nobody notices until it is expensive to fix. AI can produce the options. It cannot own the outcome, and somebody has to.
A rule of thumb
The distinction that actually holds up in daily practice is this: AI is trusted with anything reversible, and judged carefully on anything that is not. A layout can be generated, reviewed and discarded in seconds, so it gets generated freely. A commitment made to a client, a line of copy that will represent someone publicly, or a piece of architecture that other systems will come to depend on: that gets slower, more deliberate, more human scrutiny, precisely because it is expensive to undo later.
In practice, that means using AI heavily at the start of a project, while everything is still cheap to change, and pulling it back at the point a decision starts to carry weight: the moment a design is signed off, the moment code ships, the moment a claim is made in public. The tool does not become less useful at that point. It simply stops being the one making the call.
The point of the human in the loop
None of this is a caveat added to reassure anyone that AI is being kept in its place. It is closer to a working method: AI does the researching, exploring, drafting and testing, and a person who understands the brief, the risk and the audience decides what is actually good enough to ship. That division of labour is not new; it is how any studio has always worked, with experienced judgement sitting over faster, cheaper first passes. What has changed is how quickly the first pass now arrives, and how much more of it there is to look at. If anything, that makes the judgement matter more, not less.