The end of the myth of easy
When ChatGPT was released in November 2022, it made powerful AI feel instantly accessible. For many people, that was the first shock: not that AI existed, but that almost anyone with a browser and a prompt could use it.
The first phase of generative AI carried a seductive message. AI was easy. Ask the right question, write the right prompt, automate the right task and value would follow.
That story is now running into reality. Major AI providers are investing heavily in deployment, integration and forward-deployed engineering. The lesson is becoming clearer: access to powerful models is not the same as the capability to apply them well.
This applies below enterprise scale too. Smaller businesses may not be hiring teams of forward-deployed engineers, but they face the same underlying challenge. AI has to improve real work, not merely make work faster or more impressive-looking.
That is where a quieter risk becomes important. The real AI risk is not adoption. It is abdication.
What is AI abdication?
AI abdication occurs when consequential parts of reasoning shift to the model without sufficient human scrutiny, judgement and accountability.
That is different from delegation. Asking AI to summarise material, generate alternatives, produce a first draft or organise information can be an effective division of labour. The human can delegate parts of the work while still owning the question, evaluating the evidence and deciding what should be done.
Abdication begins when that ownership weakens. AI starts framing the problem, selecting the emphasis, interpreting the evidence or producing the recommendation while the human role becomes increasingly passive.
This is not necessarily a failure of individual character. It is often a consequence of how the work is designed. People are encouraged to use AI, rewarded for speed and given tools that can make outputs look finished before the thinking is complete.
That creates an obvious temptation. Fluent output can feel like completed reasoning. But fluency is not judgement, speed is not understanding and a confident answer is not the same as a considered one.
The problem is knowing where capability ends
One of the strongest warnings from the research is that AI capability is uneven.
Dell’Acqua et al. describe this as a “jagged technological frontier”. AI can perform extremely well on some tasks while performing much less reliably on others that appear quite similar from the outside.
In their study of work conducted at a large consultancy, AI improved performance on a range of consulting tasks. But performance deteriorated on a strategic recommendation task that required participants to integrate spreadsheet data with interview evidence.
The significance goes beyond saying that AI is good at some things and bad at others. The harder management problem is that people cannot safely infer capability on one task from impressive performance on another.
Questions such as “Can AI do marketing?”, “Can AI do analysis?” or “Can AI do strategy?” are therefore too broad to be particularly useful. Within each category there are different kinds of reasoning, different evidence requirements and different consequences if the answer is wrong.
AI may be highly effective at generating alternatives, structuring rough material or summarising information. It may be less reliable where evidence has to be weighed, context interpreted and a commercially sensitive judgement made.
This is precisely why human judgement remains necessary. If AI capability were consistently high or consistently low, allocating work would be relatively straightforward. Jagged capability makes that allocation itself a judgement task.
It also creates a further problem: AI can make the wrong answer look more professional. That is not a reason to avoid it. It is a reason to stop confusing polish with reliability.
Why GenAI makes abdication easier
Generative AI makes the boundary between support and abdication unusually easy to cross because it can participate across an entire workflow.
Older forms of AI often appeared at a relatively contained point: scoring a lead, ranking options, flagging a risk or making a recommendation. There was usually a clearer moment at which a human could accept, reject or investigate the output.
GenAI is different. It can help define the issue, research the context, summarise evidence, develop options, produce recommendations, refine the wording and package the result for an audience. Randazzo et al. describe forms of human-AI collaboration that extend across the work rather than sitting at one isolated decision point.
That breadth is useful, but it creates a subtle risk. The person may feel they are simply working more efficiently. At every stage the material becomes cleaner, more coherent and easier to share. Yet if AI is shaping each stage, the human can gradually lose contact with how the conclusion was reached.
Abdication therefore does not need to look reckless. In an organisation under pressure to adopt AI, it can look modern, efficient and productive.
Marketing shows why task-level judgement matters
Marketing is a useful example because it is clearly knowledge work but not one kind of work. It includes research, interpretation, creative development, customer understanding, message testing, brand judgement, channel decisions and external-facing execution.
A large-scale experiment by Ju and Aral illustrates why broad claims about AI capability are inadequate. Human-AI teams produced more advertisements per worker and higher-quality advertising copy, suggesting meaningful productivity and text-quality gains. But the benefits did not apply evenly across the task. Human-human teams produced higher-quality images and human-AI teams produced more homogeneous outputs.
In practical terms, AI helped people produce more and improved some dimensions of quality while weakening others.
That matters commercially. Marketing effectiveness does not come simply from producing polished material. It depends on distinctiveness, customer relevance, timing, context and strategic fit. A campaign can sound fluent and plausible while still being ineffective because it repeats generic category language or fails to identify the real buying tension.
Research into advertising agencies by Serra-Simón et al. points in a similar direction from practice. GenAI is being incorporated into creative and production processes, but not as a simple replacement for human strategic and creative work. Human verification remains part of the process.
The lesson is not that AI is suitable for creative work but unsuitable for strategy, or vice versa. It is that capability has to be understood at the level of the actual task and workflow.
Where human control sits
Many organisations are still too vague about the human part of human-AI work. They provide access to tools, encourage or require their use and assume appropriate judgement will naturally form around them.
Randazzo et al. identify three broad modes of human-GenAI knowledge work: Cyborg or Fused Co-Creation, Centaur or Directed Co-Creation and Self-Automator or Abdicated Co-Creation. The labels are less important than the underlying distinction. Different ways of organising AI use produce different patterns of human control.
Consider two superficially similar uses. A sales director asks AI to draft several versions of a client follow-up, evaluates them against what they know about the relationship and substantially adapts one before sending it. AI has contributed heavily to production, but the human remains in control of the judgement.
A manager asks AI to analyse a complex client problem, recommend an answer and draft the proposal, then gives the finished document a quick skim. The model may have performed brilliantly, but the human is in a much weaker position to recognise when it has not.
The important organisational question is therefore not simply whether a human remains “in the loop”. It is where human control sits, which parts of the workflow require active judgement and where delegation can occur without losing the ability to interrogate the result.
Self-automation can look highly productive
The most dangerous form of abdication is not messy or obviously careless. It often looks productive.
Work appears quickly. It is coherent, well structured and easy to share. The recommendation sounds reasonable and the language creates an impression of control. That apparent quality can make it harder to see whether the issue was framed properly, the evidence was weighed appropriately or important weaknesses were missed.
There is also a developmental consequence. People do not simply produce work through organisational systems. They develop expertise through doing the work.
Randazzo et al. associate Self-Automation with more limited development of both domain expertise and AI-related expertise. If AI repeatedly performs the framing, analysis and drafting, people lose some of the opportunities through which judgement itself develops.
This matters well beyond individual productivity. Organisations need future experts as well as faster current employees. A system that produces excellent immediate outputs while progressively weakening the expertise required to evaluate them creates a different kind of risk.
The danger is not that AI writes badly. It is that it writes well enough to make weak judgement harder to see.
Judgement is the control mechanism
The productivity gain from AI comes partly from delegating generation. AI can produce options quickly, organise messy material, summarise information and turn rough notes into usable text. Those are valuable capabilities. They are not the same as knowing what matters.
In business, judgement involves understanding the problem, noticing what is missing, weighing conflicting evidence and risk, interpreting context and taking responsibility for the consequences of a decision.
AI can generate a recommendation that sounds balanced, cautious or commercially sophisticated. But it does not own the client relationship, the budget, the brand damage, the legal exposure or the strategic cost of being wrong.
This is why AI capability alone does not determine the quality of the resulting work. Performance emerges from the interaction between the capability of the model, the task it is given, the context and data available to it, the expertise of the human user and the controls built around the workflow.
Human judgement is therefore not simply a bottleneck to be automated away. It is part of the control mechanism that makes AI commercially usable.
What businesses should design for
Giving staff access to AI tools and hoping productivity improves is not an AI strategy. Neither is training everyone to write better prompts.
Mature adoption requires organisations to decide deliberately where AI should contribute and where human ownership must remain explicit. In practice, that means understanding capability at task level, identifying where important judgements enter a workflow and defining points at which outputs need to be verified, challenged or escalated.
The appropriate balance will differ by task. Some work can be delegated extensively because errors are cheap, visible and easily corrected. Other work requires humans to remain much closer to the reasoning because the evidence is ambiguous, the context is difficult to encode or the consequences of being wrong are substantial.
The objective should not be maximum human involvement. Nor should it be maximum automation. It should be an effective allocation of capability, with enough human control to understand what is being produced and why.
What the evidence does and does not show
The evidence does not support a simple rule that AI is good for creative work and bad for strategic work. Dell’Acqua et al. demonstrate precisely why such categories are too broad: performance can vary sharply across tasks that appear closely related.
Nor do the marketing studies establish that AI improves every dimension of marketing performance. Ju and Aral find productivity and copy-quality gains alongside weaker image performance and greater homogeneity. Serra-Simón et al. show that agencies are incorporating GenAI into their work while continuing to rely on human verification.
Similarly, the modes described by Randazzo et al. should not be interpreted as proving that one pattern of human-AI interaction is universally optimal. Different tasks create different requirements for control, expertise and delegation.
The practical conclusion is therefore not “use AI less”. It is to take the design of AI-enabled work more seriously.
The mature AI position
The mature position is neither AI avoidance nor blind adoption. It is disciplined human-AI work.
AI can reduce friction, increase productive capacity and improve the quality of some forms of work. But those gains become commercially valuable only when organisations understand where the technology performs well, where its capability becomes less reliable and where human judgement must remain active.
The next phase of AI is likely to reward something more demanding than rapid adoption: the ability to use increasingly powerful systems without surrendering judgement to them.
That is the distinction between delegation and abdication. AI should make human expertise more productive. It should not quietly replace the reasoning required to know whether the work is any good.
References
Dell’Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Organization Science.
Ju, H., & Aral, S. (2025). Collaborating with AI agents: Field experiments on teamwork, productivity and performance (arXiv:2503.18238). arXiv.
Randazzo, S., Lifshitz-Assaf, H., Kellogg, K. C., Dell’Acqua, F., Mollick, E. R., Candelon, F., & Lakhani, K. R. (2025). Cyborgs, centaurs and self-automators: The three modes of human-GenAI knowledge work and their implications for skilling and the future of expertise (Harvard Business School Working Paper No. 26-036). Harvard Business School.
Serra-Simón, J., Puntí-Brun, M., Espinosa-Mirabet, S., de Faria Nogueira, M. A., Martín-Guart, R., & Tôrres de Azevedo, S. (2025). Generative artificial intelligence in advertising: Field applications in Rio de Janeiro and Catalonia. Telecommunications Policy, 49(8), Article 103033.
