A recent Reuters report on how large tech companies are tracking employee behaviour to train AI systems has brought a growing industry tension into the spotlight: how far companies are willing to go in sourcing data and what it means for both people and the systems being built.
At first glance, using internal workflows as training data may seem like a natural extension of AI development. In practice, it raises more complex questions. When employee behaviour becomes a dataset, the line between innovation and extraction starts to blur.There’s also a fundamental flaw in the logic.
The truth is that most people will change how they work when they know they’re being observed. You’re not getting clean data. You’re getting performance data and that’s a pretty expensive mistake if accuracy is the goal.
Andy Berg, CEO at Rhapsody
This dynamic creates a dual problem. On one side, it compromises the integrity of the data itself. On the other, it introduces a human cost that is harder to measure but equally significant. When people feel monitored rather than trusted, it impacts how they make decisions, how they collaborate, and ultimately the quality of the work they produce.
Two diverging paths
The industry now faces two distinct approaches to integrating AI into work.The first focuses on capturing existing workflows in detail and training systems to replicate them at scale. It assumes that current processes are the right foundation and that efficiency gains come from accelerating what already exists.
The second approach questions that premise entirely. Rather than optimising legacy workflows, it asks whether those workflows still make sense in an environment where AI can remove many of the constraints that shaped them in the first place.
At Rhapsody, the focus has shifted toward rebuilding systems from the ground up.
Instead of making individual steps faster, the company is rethinking how those steps connect reducing handoffs, bringing decision-making closer to execution, and applying AI where it removes friction rather than simply increasing speed.
Rethinking the workflow, not accelerating it.
Retouching offers a clear example of this shift.
In a traditional setup, AI is added onto an existing process. A designer shares a PSD file with comments, a retoucher opens it in Photoshop, and AI assists in executing changes more quickly. The workflow remains unchanged, only faster.
Rhapsody approached the problem differently.
Rather than asking how to accelerate the process, the team questioned why the process existed in that form at all. At its core, the “brief” is simply feedback on an image, yet it is often packaged as a file, passed between tools, and managed across multiple versions.
The alternative was to remove the handoff entirely.
Rhapsody developed a shared, browser-based canvas where designers annotate directly on the image. Retouchers (whether human or AI) operate in the same environment, with changes visible in real time. There are no version files, no back-and-forth exchanges, and no need to reconcile parallel workflows.
In this case, AI didn’t improve the existing system. It made much of it unnecessary.
One system, not parallel tracks
This approach reflects a broader principle seen in other industries. Tesla, for example, did not design separate systems for human drivers and Autopilot. It created a unified interface that supports both.
The same logic is increasingly relevant in creative and production workflows: one system where humans and AI operate together, rather than separate tracks that require constant alignment.This shift highlights where AI’s real impact lies.
Historically, output has not been the primary constraint in creative work. The bottlenecks have been time, cost, and the structural complexity required to move from idea to execution. AI has the potential to rebalance those factors but only if it is applied within rethought systems, not legacy ones.
The role of human judgment
What AI does not replace is judgment.Deciding what is worth creating, what should be refined, and what ultimately delivers value remains a human responsibility. As the volume of possible output increases, that role becomes more critical, not less.
The decisions companies make now will shape how that balance evolves.What we’re seeing across the industry is unlikely to be an isolated case. It represents an early signal of one direction the industry could take one that prioritises data extraction and replication of existing behaviours.
The alternative is to use this moment to design systems that are not only more efficient, but also more aligned with how people actually work best.
A question of trust and accuracy
There is also a practical consideration that extends beyond ethics.If employees modify their behaviour under observation, the resulting data is not an accurate reflection of real workflows. It becomes a performance shaped by awareness rather than authenticity.
That introduces a data quality issue alongside the cultural one.Both are significant. And both are avoidable.As Berg puts it, the companies that will prove most resilient over the next decade are unlikely to be those that extracted the most from their people.They will be the ones that trusted them enough to keep them.