When talking about the impact of AI on jobs, the dominant narrative is augment vs automate. The technology is described as either a substitute for or an enhancer of human labour. However, when we look closely at what humans really do when they use AI at work, we see, in addition to the familiar flows of automation and augmentation, two other flows often neglected by the dominant narrative. On one hand, AI creates new tasks that did not exist before. On the other hand, people take on new tasks previously done by colleagues. AI reconfigures work, and the main strategic implication is that organisation design is an important strategic lever and source of competitive advantage.

In 2016, Geoffrey Hinton famously suggested that because image recognition was improving so fast, we would no longer need radiologists. Technically, he wasn’t wrong: today’s AI systems regularly outperform humans at recognising anomalies in medical images. And yet, nearly a decade later, in the UK, the number of radiologists employed by the NHS has increased by more than 40 per cent since that prediction. AI did not arrive as a substitute that cleanly removed human labour from the system but as a complement that remodelled the work around it. Radiology is not an isolated case. Research published this year found that the companies investing most heavily in AI are adding employees faster than their peers, with white-collar hires rising by 10.2 per cent, a finding that runs directly against the prediction of broad, technology-driven job losses. Something other than substitution is going on, and it is visible in the aggregate, not only in one profession.

Four flows: Automated, Augmented, Created, Expanded.

We can summarise how AI reshapes and expands work in four flows. First, it automates parts of tasks that humans previously did. Writing lines of code for producing software, for example. Second, AI augments human work by enriching a human piece of work or by making a human decision more accurate or complete. Third, AI creates new tasks made necessary by the technology. Training on AI, for example. Last, AI expands work beyond the initial scope of the job description. For example, when graphic designers produce software based on their work.

Figure 1: The four flows of AI at work

Automated

Much research has shown AI’s ability to automate some tasks. With AI producing code, slides, and analysis, there is less need for entry-level positions, and several studies have documented a drop in employment among young graduates. For example, a study by Stanford researchers (among whom E. Brynjolfsson) concludes that “since the widespread adoption of generative AI, early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 13 per cent relative decline in employment”. Read on its own, this is the flow that drives the headcount arithmetic. Read against the three that follow, it is the most visible of the four and the easiest to overstate. The effect is also uneven, which is easy to miss when the flow is read as a single number. PwC argues that AI is splitting the job market in two rather than shrinking it uniformly. That split matters for what follows: if the technology destroys entry-level work in one part of the market while creating and expanding work in another, then a net headcount figure will conceal both movements at once.

Augmented

AI is not always automating human tasks but, in many instances, it augments them. Humans continue playing a role and see their productivity or quality increase because they use AI at some point in the process. The question of human-AI collaboration has been widely discussed over the last few years. A recent study analysed pricing managers’ decisions regarding AI-made pricing recommendations. The study intends to understand situations when managers override algorithmic pricing recommendations. The central finding is that the value of human intervention depends on when it occurs. Early interventions can create value by improving what the algorithm learns, while later interventions are most valuable when they correct algorithmic errors. The best outcomes come not from fully automated or fully human decisions, but from a carefully designed collaboration between the two.

In current research on the impact of AI on jobs in scriptwriting, we found similar evidence. Specifically, we identified the roles AI plays in the creative process. Some were illustrative of automation, and others of augmentation. Analysing data from interviews with professional scriptwriters, we identified 4 different roles that GenAI plays during the scriptwriting process.

• GenAI as a creative Guide. GenAI is used to structure thinking, test narrative possibilities, and explore thematic or dramatic options. Writers rely on it for brainstorming within established frameworks or constraints, often using it to challenge their assumptions or expand their field of vision.
• GenAI as a proofreader and rewriter. GenAI is also used to polish, simplify, or compress existing material, especially to rewrite scenes, adjust tone, or summarise lengthy documents.
• GenAI as a content producer. GenAI acts as a generative tool, especially for repetitive or format-driven writing tasks such as sketch comedy, synopses, or support materials for residencies or funding bodies. Here, GenAI isn’t just assisting the writer; it is generating material that may be used in final outputs.
• GenAI as a creative collaborator. The final role we identified was more dialogic in nature, an affective relationship with GenAI, using it as a stand-in for a co-writer or sparring partner in the early stages of idea development. Though its contributions are limited, GenAI offers a low-stakes way to bounce ideas and ‘talk through’ narrative choices.

The strategic clue is already here: the value does not sit in the task the machine performs, but in the design of the collaboration around it: who intervenes, and when.

Created

When analysing the data in our research, we observed another phenomenon. Scriptwriters described a significant number of new tasks. These tasks were directly related to using the technology. We found 20 of them and grouped them into 5 families. Where the four roles above describe what GenAI does, these new tasks describe what the writers themselves now have to do because of it:

• Framing: tasks through which scriptwriters construct and maintain the contextual conditions that make GenAI’s contributions relevant. Context becomes an explicit object of work. As one interviewee explained, “you have to remind it what happened before, otherwise it forgets and goes in another direction”.
• Generative orchestration: tasks through which scriptwriters steer and shape the generative behaviour of GenAI. One scriptwriter described using the system to explore narrative options quickly: “I asked it to list potential family secrets between two sisters. In ten seconds, we had fifteen ideas.” Others highlighted how GenAI enables them to think against themselves by surfacing counterproposals or unexpected direction.
• Evaluative curation: tasks through which scriptwriters assess and refine GenAI outputs to ensure logic and project alignment. This refers in particular to checking hallucinations or sorting through a large volume of propositions.
• Normative and affective regulation: tasks through which scriptwriters correct or constrain GenAI outputs to match emotional resonance, moral boundaries, and creative intent. As one scriptwriter observed, “ChatGPT can describe emotions, but it cannot make you feel emotions”.
• Training: effective GenAI use requires an ongoing investment of time and effort in learning the tool, experimenting with its limits, understanding how it processes requests, attending trainings, and building personal heuristics.

These are not overhead to be minimised. They are where the real work of using AI now lives and, as we will see, where a firm’s advantage in the technology is built or lost.

Expanded

Contrary to the vision of a tool automating tasks and reducing the demand for work, GenAI appears to intensify work.  Two researchers from the University of California, Berkeley studied how AI was changing work habits at a US tech company. They saw an increase in both the number of hours worked and the intensity of that work, as well as a broadening of the range of tasks people took on. Most notably, the company neither asked for these changes nor mandated the use of AI tools; the behavioural shifts happened organically.

In the same direction, according to a study by the OpenAI research lab, « AI changes the work that people do ». In an analysis of more than 800,000 messages from U.S. ChatGPT users, the research concludes that about half of occupation-specific messages are about tasks associated with another occupation. This suggests a changing division of work: some activities that once required a handoff can now be done by the person who first encounters the need. The researchers term the phenomenon task crossover: work historically associated with one occupation appearing in the AI use of people in another. One particular crossover is visible in the research: marketing. Marketing tasks now appear significant in all worker occupations studied (including engineering and finance). These professionals used to work with their marketing colleagues and are now turning to their AI system for marketing tasks. The same phenomenon is visible for engineering.

Once the lines between jobs become this porous, the question is no longer who does which task, but how the firm chooses to draw the lines at all, a question of organisation design. And that question has a cost attached. Early evidence suggests that heavier AI adoption may require more managers rather than fewer, because once tasks are redistributed across people and machines, coordination, not execution, becomes the binding constraint.

Strategic implications

If work is reconfigured instead of being divided, the strategic question changes. It is no longer “which tasks do we automate?” but “how do we redesign the whole system of tasks made by human and machine around the technology?” The point is not only analytical. One of the more convincing explanations for why AI is not yet showing up on corporate bottom lines is precisely this: organisations have added the technology without reorganising around it. Hence, the gains stay trapped at the level of individual tasks.

The value at stake is not the productivity gain. Reading AI’s value as a productivity boost captures only the first-order effect, and it consistently overstates the net benefit while missing where value actually settles. Two mechanisms are at work. First, a cost illusion: making a task cheaper raises the demand for it, and it spawns the very tasks we described above, which absorb part of the saving. Second, the productivity of a single task tells you little about the performance of the system it sits in. Value creation with AI is a property of the configuration, not of the task; it depends on how the reconfigured tasks fit together, not on how fast any one of them now runs.

Two recent reversals show what the first path looks like in practice. Ford deployed 900 AI-powered cameras to perform quality checks, then rehired more than 300 veteran quality inspectors when the technology underperformed. What the returning inspectors now do is instructive: they train the AI systems and mentor young workers, which means they perform the created tasks rather than the ones the cameras replaced. As Charles Poon, Ford’s vice president of vehicle hardware engineering, put it, artificial intelligence is a fantastic tool, but it is only as good as the information you use to train it. Starbucks, for its part, quietly retired an AI agent a few months after deployment when it hallucinated store inventories and slowed baristas down. Neither case is a story about a technology that does not work. Both are stories about a technology dropped into a process that was not redesigned around it, and in both the cost surfaced in the system rather than in the task.

The scarce capability moves from doing the task to orchestrating the system. If AI creates a new layer of work, then these become the capabilities that separate the firms that extract value from those that merely deploy the technology. They are largely tacit, learned by doing, and unevenly distributed, which in resource-and-capability terms is precisely what makes them a candidate source of advantage rather than a commodity bought off the shelf. The practical implication is a deliberate investment in capability-building, not training in “using ChatGPT,” but in the orchestration work itself: how to frame, how to curate, how to regulate.

Organisation design becomes the primary strategic lever. This is where the three threads meet. Reconfiguration means the unit of decision is no longer “automate or augment this task,” but “how do we allocate the whole portfolio of tasks, the old ones as well as the new ones, across humans and machines, and how do we redesign jobs around the result?” It is the make/buy/partner question brought inside the firm and applied at the level of the task: which tasks to hand to the machine, which to keep human, and which to recombine into new roles: the analyst who now writes software, the marketing work now done inside engineering. Two things hang on the answer. The cost of the structure, or how lean it can be. And, more importantly, what the firm can achieve in quality and speed that its competitors cannot. Because job and process design is causally ambiguous, socially complex, and hard to observe from the outside, a superior configuration is difficult to copy. That is what turns organisation design from an administrative afterthought into a genuine source of competitive advantage. The firms that win the AI transition may not be the ones with the best models, since everyone rents the same ones, but the ones with the best-designed organisations around them.

Conclusion

The augment-versus-automate debate asks a binary question about tasks. Watching what people actually do with the technology reveals something richer: four flows, not two. Tasks are automated, tasks are augmented, tasks are created, and the scope of human work expands. AI does not simply divide the existing work between people and machines; it changes the work itself, adding as fast as it subtracts.

That reframing changes the strategic conversation. It moves the question from cost (how many people can we remove?) to design (what can we now build?). If the technology adds tasks as fast as it removes them, then counting only the tasks it removes measures the wrong side of the ledger. The other side is the one worth counting: what the reconfigured organisation makes possible in quality, in speed, in scope, once the work is redesigned around AI rather than merely sped up by it. The firms that ask only “what can we automate?” will find a modest, one-off saving. The firms that ask “what can we now become?” are the ones for whom AI is a strategic lever rather than an efficiency tool.

So, the closing question is really two.

     If AI adds tasks as fast as it removes them, are you measuring the wrong side of the ledger?

     What previously impossible things, whether in quality, speed, or scope, become possible with the organisation design that AI now allows?

Photo de SwapnIl Dwivedisur Unsplash