The Work Moves Downstream

The First Draft Became Cheap

I am an enthusiastic user of AI. I use it to research, write, test ideas, build software and challenge my thinking. It has allowed me to produce things I would not have attempted a year ago. It has made me faster, but its more important effect is that it has made me more capable.

I have also received AI-generated work that creates more effort than it saves.

This is not always obvious at first. A poor human draft tends to announce itself. It is incomplete, uneven or visibly uncertain. A poor AI-assisted draft may be coherent, comprehensive and professionally structured. The confidence of the presentation raises the cost of discovering that the thinking underneath it is thin.

The first draft became cheaper. Attention did not.

When four revisions can be generated in two days, the scarce resource is no longer the ability to produce another version. It is the time and judgement required to determine whether the version is any better.

I have begun to think of this as iteration inflation. The ability to revise quickly creates the appearance of movement, even when the underlying problem remains where it was. Feedback that should trigger investigation or a decision instead triggers more text.

The output is faster. The outcome is not.

The Work Moves Downstream

The person producing the document may genuinely believe they have become more productive. They completed the task sooner, responded rapidly and delivered something polished. The additional effort is no longer visible from where they sit.

It appears later, in somebody else's calendar.

A manager spends an hour identifying unsupported claims. A specialist has to check whether a control was actually tested. A colleague rewrites the recommendation so that it reflects the organisation rather than a generic one.

Several people attend another meeting because the document did not resolve the question it was commissioned to answer.

The work did not disappear. It changed owners.

That matters because organisations often measure activity close to its point of production. Time to first draft is visible. The time senior people spend repairing it is dispersed across meetings, comments, conversations and quiet rewriting. The person generating the material appears efficient; the people absorbing the verification debt simply appear busy.

An organisation cannot reward speed, ignore the cost of review, and then be surprised when it receives polished drafts faster than it receives better decisions.

This is not only an individual failure. It can be the predictable result of what the organisation has asked for.

Use AI. Move faster. Produce more. Demonstrate innovation.

If those instructions are not accompanied by a clear standard for what is ready to be reviewed, people will optimise for visible output. They may be doing exactly what they think leadership wants.

A Checkbox May Be an Excellent Use of AI

There is an easy response to all of this: distrust anything that appears to have been generated by AI.

I do not think that response survives contact with reality.

Some work is repetitive. Controls have to be checked. Evidence has to be gathered and mapped against criteria. Reports have to be assembled. If AI can perform those tasks accurately, preserve traceability and release a skilled person to investigate exceptions and provide better insight, that is not cheating. It is a good use of technology.

The problem is not that AI can complete the checkbox. It is that organisations may mistake a completed checkbox for completed thinking.

At the time of writing, Wellington mayor Andrew Little has alleged that large parts of a NZ$435,000 Deloitte staffing report were written using AI. Deloitte has said it uses AI-enabled tools where appropriate and that humans signed off the analysis and conclusions. The controversy is not really about whether a machine helped form the sentences. It is about the quality of the work: disputed assumptions, incorrect data use, double-counting and an alleged NZ$21.5 million overstatement in staffing costs. The original report presented a headline conclusion that Wellington City Council had 330 more full-time-equivalent staff than its benchmark suggested. (Original report; RNZ coverage)

If those criticisms are well founded, the presence of a human sign-off is not especially reassuring. It may only demonstrate that human-in-the-loop can become another checkbox.

A meaningful human in the loop is not a person whose name appears at the end. It is someone with the context, time, authority and accountability to challenge what the system produced.

I support AI-augmented work when augmentation improves the outcome. I become cautious when it merely accelerates production, obscures responsibility or moves verification downstream.

The useful question is not: Did AI write this?

It is: Did AI reduce the total effort required to reach a trustworthy outcome, or did it reduce the author's effort by increasing everybody else's?

The Capable Operator

I have seen the other outcome too.

A capable person uses AI to explore alternatives, find gaps, interrogate assumptions and improve work before anyone else sees it. They do not pass each generated version to their manager. They use the cheap iterations privately and submit the result only when they understand and endorse it.

A good AI operator does not simply produce faster. They absorb more of the checking themselves before burdening the next person.

Under the right conditions, that person really may produce in two days something that previously required four weeks. They have not asked AI to replace their judgement. They have used it to expand the amount of investigation, drafting and testing their judgement can direct.

But even this claim requires care.

A business case is not merely a document. It is the visible record of choices, evidence, consultation and organisational commitment. AI may compress the writing. It cannot necessarily compress the time required to consult affected people, understand operational reality, resolve disagreement or earn permission to proceed.

Sometimes four weeks of drafting can become two days. Four weeks of organisational sense-making probably cannot.

This distinction will be easy to lose as highly capable AI users become more visible. Leaders will see what they can produce and begin recalibrating their expectations of everybody else.

That creates a genuine capability divide.

The People Who Are Left Behind

In leadership discussions, I have wondered whether knowledge work is developing a new hierarchy: people who have learned to work effectively with AI, and people who have not.

I do not mean that every enthusiastic user belongs in a superior class. Some people use AI constantly and produce very little of value. Some thoughtful sceptics use it rarely and continue to produce excellent work. Usage is not literacy.

But the performance difference can be real.

A capable person using AI well may research faster, consider more alternatives, communicate more clearly and attempt work that previously required specialist help. Each successful task gives them more confidence and experience. They receive more opportunities because leaders learn that they can deliver. The advantage compounds.

Meanwhile, somebody who is uncertain may receive fewer difficult assignments because supporting them requires more time. They gain less experience. As the strongest operators accelerate, the distance between them grows.

The people left behind may not have been less capable. They may simply have been given less opportunity to become capable.

This seems unfair because it is unfair, at least sometimes.

Access to an AI tool is not the same as access to the conditions required to learn it. People need time, useful examples, feedback, domain knowledge, permission to fail safely and colleagues who can show what good practice looks like. A licence and a one-hour prompt-writing session do not create capability.

There are also legitimate reasons for hesitation. People may be concerned about privacy, accuracy, professional obligations, the effect on their role, or the environmental cost of rapidly expanding data-centre infrastructure, including its demand for energy and water. They may have encountered so much AI-generated rubbish that distrust has become rational. Others may simply produce excellent work without it.

An organisation should judge the quality and timeliness of the outcome, not demand loyalty to a particular tool.

But that cannot mean protecting every existing way of working forever. If AI becomes a normal and valuable instrument of knowledge work, refusing to engage with it may eventually become a capability issue.

Fairness cannot mean freezing performance expectations at the point before a useful tool existed. But it also cannot mean redefining competent work around the fastest early adopters before everyone has had a fair opportunity to learn.

Governance Arrives Late

Technology leadership has dealt with capability transitions before. Work moved from centralised terminals to personal computers, through successive cycles of distributed and centralised systems, and eventually into cloud services. Each change affected skills, operating models, control and expectations.

AI feels different because it is happening so quickly and so personally.

Previous technology transformations often arrived as programmes. Equipment was purchased. Systems were implemented. Training was scheduled. Migration dates made the transition visible.

Previous technology transformations arrived as programmes. AI is arriving as behaviour.

People begin using it through tools already on their desks or browser tabs the organisation cannot see. They develop practices before policy exists. By the time governance, training and assurance catch up with one capability, the tools have changed and the most engaged users have already moved further ahead.

I have experienced this directly. We establish guidance for the uses we understand, then almost immediately encounter a new one that does not fit. The responsible response cannot be to play governance whack-a-mole forever: discover a new behaviour, write another rule, discover another behaviour, write another rule.

Static governance will always be late to a capability that changes continuously.

But abandoning governance in favour of general principles is not enough either. Principles such as transparency, accountability and human oversight sound excellent until someone has to decide whether a particular output can be trusted, whether information should have been entered into a tool, or what human-in-the-loop means in practice.

The organisation therefore has to govern both the enduring questions and the changing uses.

Who owns the outcome? What evidence supports it? Where did the information come from? Who can challenge the decision? What happens when confidence is low? Which actions remain human? How is harm detected and recovered from?

Those questions can survive the next model release. The specific controls around them will still need to evolve.

What Does an Organisation Owe Everyone Else?

When AI makes some people dramatically more capable, what does an organisation owe everyone else?

It does not owe them indefinite protection from change. Technology has always changed which capabilities are valuable, and individuals retain some responsibility for developing their practice.

But the organisation created the incentives, selected the tools and benefits from the productivity. It therefore carries responsibility for making the transition more than a private contest between enthusiastic early adopters and everyone else.

People need protected opportunities to practise. They need real work rather than abstract prompting exercises, with feedback on judgement rather than tricks for producing fluent text. They need clear expectations about evidence, accuracy and ownership. They need permission to admit when an experiment failed without concluding that all AI use is unsafe.

Managers also need to stop quietly repairing every inadequate AI-assisted output. That may be necessary under pressure, but if it becomes normal, the organisation establishes an unhealthy division of labour: staff generate and senior people supply the missing judgement.

Returning the work with clear questions is slower today but may build capability for tomorrow.

The obligation is not to ensure nobody ever falls behind. It is to ensure nobody is abandoned there without access, explanation, support and a fair opportunity to learn.

That obligation becomes harder as the frontier keeps moving. The people providing the training are learning too. The technology leaders writing governance are also experimenting, revising and occasionally discovering that last month's confident position no longer survives.

I find that exciting. I also find it exhausting.

What Are We Actually Measuring?

Organisations will be tempted to answer all of this with adoption statistics.

How many people have access? How many use the tool each week? How many hours have been saved? How many documents were produced?

Those measures are easy to collect and may tell us almost nothing about whether work improved.

The more useful measures sit further downstream: time to a usable result, number of review cycles, reviewer effort, defects discovered later, confidence in the decision and whether saved capacity was converted into better service or deeper insight.

The real unit of AI productivity is not time to first output. It is total time to a trusted outcome.

That is more difficult to measure because it crosses people and teams. It may also reveal that some celebrated gains were merely costs moved somewhere less visible.

This does not make AI a bad investment. It makes organisational honesty part of realising the investment.

The Work, the Judgement and the Gain

I do not want organisations to slow AI adoption until every uncertainty has been resolved. That would leave people experimenting privately, governance becoming less relevant and genuine opportunities unrealised.

I also do not want speed to become the entire definition of progress.

Technology leadership in this moment is not simply choosing tools or preventing misuse. It is helping an organisation revise its understanding of competent work while the work itself is changing. It means supporting people without pretending nobody will need to adapt. It means recognising exceptional capability without immediately turning it into the minimum expected of everyone. It means allowing automation to remove effort while refusing to let it remove accountability.

The divide is real.

Some people are already becoming substantially more capable through AI. Others are producing more material without producing more value. Some remain cautious or disengaged. Many are somewhere in between, trying to learn while the target moves.

I do not yet know what a perfectly fair transition looks like. I doubt one exists.

But I know the questions I want leaders to ask.

Did the work improve, disappear or merely move?

Who received the productivity gain?

Who inherited the verification?

Who still owns the judgement?

And who was given a fair chance to learn?

Yes, an AI Helped Write This Too

This essay began during a Friday evening conversation with an AI system. I supplied the experiences, the unresolved leadership problem and my discomfort with the easy answers. The AI helped distinguish fast output from faster outcomes, and supplied language I immediately recognised: iteration inflation, verification debt, and the work moving downstream.

There is an obvious risk in using AI to create a polished essay about the danger of polished AI-assisted work.

The contradiction belongs here too.

If I publish this without checking the public example, reconsidering the claims and editing the argument into something I understand and endorse, I will have demonstrated the problem rather than examined it.

AI has helped me reach a credible draft quickly. Whether the thinking is finished remains my responsibility.