Managing digital labor - a case for an AR function
- Claas
- May 25
- 10 min read
We probably do not need to spend much time anymore explaining what AI can do or where it can be useful. Most people working in knowledge-intensive environments already use it in some form. Texts get drafted, presentations structured, research accelerated, reviews supported and intermediate results optimized before they move further through the organization.
Artificial work is already part of normal business activity, but some fundamental questions remain unsolved. We see very responsible usage in some places and surprisingly careless behavior in others. AI-generated content gets forwarded without proper review, unclear information gets mixed into existing material and responsibility for validation quietly moves from one person to the next.
The classic "human in the loop" often sounds better in theory than it works in practice. In reality, responsibility frequently gets delegated along with the document itself. Someone generates a first version, another person adjusts a few parts, a third continues working on it and eventually nobody is fully sure anymore what was original input, what was generated, what was actually reviewed and where assumptions or mistakes entered the process.
At the same time, uncertainty around policies, security and data handling remains visible in many organizations. What is allowed? Which tools can be used? Which data may be uploaded? What requires explicit review? And who decides where sensible limits actually are?
Despite all of this, adoption continues to grow quickly because the productivity gains are simply too attractive to ignore. And exactly there the organizational question becomes interesting. Companies are already using AI like additional capacity, but most organizations have not caught up with what that actually means.
From HR to AR
The idea of Artificial Resources (AR) or dedicated AI leadership roles such as a CAIO came up in some of my earlier posts. At the time, the discussion was still mostly linked to strategy, governance or AI adoption in general. With the increasing amount of actual artificial work that now flows through organizations every day, there are fresh perspectives to consider.
Once systems actively contribute to operational output, the similarities to traditional workforce questions become difficult to ignore.
Human Resources | Artificial Resources |
Recruiting | Selection of AI tools and models |
Onboarding | Integration into workflows and teams |
Training & Development | AI literacy and usage patterns |
Performance Reviews | Output validation and quality checks |
Compliance | AI governance and policy management |
Workforce Planning | Human vs AI task allocation |
Cost Management | Token and usage control |
Knowledge Management | Context, prompt and output management |
The interesting part is that most companies already perform parts of these activities implicitly. Teams experiment with tools, IT defines technical boundaries, Cybersecurity creates restrictions, Legal reviews risks and employees develop their own usage patterns somewhere in between. But there is usually no coherent organizational model behind it.
Artificial work increasingly behaves like an operational capability, while most organizations still manage it through fragmented tooling decisions, local experimentation and isolated governance mechanisms. And that is why the discussion around Artificial Resources no longer feels entirely theoretical.
The organizational gap
One of the more striking observations is that almost nobody seems fully responsible for this topic yet. IT looks at AI primarily through a tooling, platform and architecture lens. HR naturally focuses on people, roles and organizational development. Cybersecurity and Legal concentrate on risks, compliance and restriction mechanisms. Business teams mainly optimize for local productivity and delivery speed. All of these perspectives are reasonable, but none of them fully covers what actually happens once generated work starts moving across teams and workflows.
That becomes especially visible when organizations start introducing AI policies. The underlying motivation usually makes complete sense. Sensitive data, unclear security implications, regulatory concerns, intellectual property risks or uncontrolled external usage are all legitimate issues. At the same time, many organizations are still struggling to define what responsible and practical AI usage actually looks like in day-to-day work.
Without clear ownership, the result often drifts toward two extremes. Either usage becomes heavily restricted and people stop using AI in meaningful ways because the process becomes too complicated or too risky. Or the opposite happens: AI usage spreads informally across teams with very little transparency, inconsistent validation and unclear responsibility for the generated output. Neither approach scales particularly well.
The underlying difficulty is that artificial work cuts across several existing organizational domains at the same time. It is partly a technology topic, partly a workforce topic, partly a governance topic and partly an operational delivery topic. That makes it difficult to assign clear ownership using existing structures.
IT can define platforms and technical boundaries, but usually not what constitutes responsible usage in day-to-day knowledge work. HR understands organizational development and workforce structures, but typically not the operational characteristics of AI-generated work. Cybersecurity and Legal can define restrictions and risk controls, but they are not designed to manage artificial work as an operational capability.
As a result, many organizations currently govern AI indirectly through scattered policies, local optimization and individual judgement calls. That may be sufficient while usage is still relatively contained. It becomes significantly harder once artificial work starts scaling across teams, processes and operational delivery.
Artificial work scales differently
One reason why the organizational side becomes important so quickly is that artificial work scales very differently from human work. Companies normally do not hire unlimited numbers of people without thinking about supervision, onboarding, mentoring, leadership capacity or quality control. Organizational growth creates natural friction. Teams grow gradually, responsibilities become visible and costs are relatively predictable.
Artificial work behaves differently. New capacity can appear almost instantly. Additional models, more agents, more generated output, more automation, parallel processing of tasks that previously required significant manual effort. From a technical perspective, scaling is comparatively easy. Organizationally, it becomes difficult much faster. The amount of generated work can increase dramatically without equivalent growth in review capacity, supervision or operational understanding. And unlike human work, artificial work often arrives without much transparency around how exactly a result was produced.
Half-finished AI outputs are a good example. They are often only partially usable because the surrounding context is missing. Assumptions are unclear, intermediate reasoning is gone and the next person working on the material may not fully understand why certain decisions were made, which sources were used, what was generated or what still requires validation.
At the same time, the economic model behaves differently as well. With employees, organizations usually understand the cost structure relatively clearly. Salaries, contracts, leadership capacity and budgets create visible constraints. Artificial work initially looks much cheaper and almost infinitely scalable. Until usage starts compounding. Then suddenly new categories of overhead appear: token consumption, repeated iterations, validation effort, redundant processing, multiple models reviewing each other, context reconstruction, operational inefficiencies caused by low-quality intermediate output.
Unlike traditional automation, these costs scale dynamically with usage rather than through fixed implementation costs. Familiar budgeting logic stops working.
When agents scale, structure becomes governance
The shift from individual AI usage to agent-based workflows changes the organizational question fundamentally. A single user generating a draft, reviewing it and deciding what to use next is still a recognizable work pattern. There is one person, one context, one point of accountability.
Multi-agent workflows behave differently. One agent structures the problem, another conducts research, a third challenges assumptions, a fourth consolidates output. Each step produces intermediate results that feed the next. This already happens in practice. In content pipelines that move from research to analysis to drafting to review without a human touching every transition. In consulting delivery where agents handle data aggregation, benchmarking and first-draft narrative in sequence before a senior reviewer sees the output. In financial reporting where multiple agents pull data from different sources, reconcile figures and generate commentary before a human signs off.
The operational challenge is not that agents produce bad work. The challenge is that errors, assumptions and gaps compound across steps without natural checkpoints. By the time output reaches a human reviewer, the connection between original input and final result may be genuinely difficult to reconstruct. Not because anyone was careless, but because no organizational structure required it to remain visible.
This is where individual discipline stops being a sufficient answer. Telling people to document assumptions or limit context treats an organizational design problem as a personal habit. What organizations actually need are structural decisions: where do handovers between agents require explicit output documentation? Which intermediate results need human validation before the next step starts? Who owns quality when no single person touched every step?
These are not technology questions. They are workflow design and governance questions. They require the same kind of deliberate organizational thinking that companies apply to any other critical delivery process. The difference is speed. Agent workflows can scale in days. Organizational responses typically take months. That gap is where risk accumulates, quietly and quickly.
This is precisely why the question of organizational ownership becomes urgent rather than theoretical. If agent workflows already operate across teams and functions, and in many organizations they already do, informally, then the absence of clear governance is not a neutral state. It is an active choice to let artificial work self-organize. That choice gets harder to reverse the longer it runs.
Where it breaks
Even where structure and validation are in place, there are still areas where artificial work becomes difficult to use efficiently. Some outputs can be reviewed relatively quickly because mistakes are easy to spot or consequences are limited. Brainstorming, early drafts, structuring, idea generation or exploratory work often fall into this category. Even imperfect results can still create value because they accelerate thinking or reduce repetitive effort.
Contracts, pricing documents, regulatory content, critical customer communication or anything with significant legal, financial or operational impact usually require near-complete verification. In these situations, a single unnoticed error can create disproportionate consequences. And that changes the economics completely. If generated output still requires a full manual review, the productivity gain quickly starts shrinking. In some cases, the validation effort approaches the original creation effort closely enough that the AI contribution becomes questionable from an efficiency perspective.
The problem becomes even harder when outputs appear highly plausible while containing subtle inconsistencies or hidden mistakes. Completely wrong results are often easier to identify than "almost correct" ones. That is one reason why some of the public AI failures in consulting, legal work or customer communication are probably less about carelessness than people assume. Once generated work becomes deeply embedded into larger delivery processes, maintaining complete visibility over every intermediate step becomes genuinely difficult.
Another limitation appears in long-running or highly iterative workflows. At some point, context windows become overloaded, assumptions drift, intermediate decisions disappear and the relationship between original input and final output becomes increasingly unclear. More iterations do not necessarily improve quality. Sometimes they simply increase complexity.
Artificial work is not equally suitable for every type of task. Some activities tolerate approximation and iterative refinement quite well. Others depend heavily on traceability, accountability and complete understanding of how a result was produced. The challenge for organizations is to decide where artificial work creates real leverage and where the surrounding validation and control requirements outweigh the benefit.
Organizational setup
The question of where Artificial Resources would actually sit inside an organization is less about org chart preferences and more about which function is structurally capable of owning the problem.
The obvious answer would probably be IT, simply because most AI initiatives currently start there. Platforms, tooling, architecture and security are all heavily technology-driven topics and naturally create a gravitational pull toward IT ownership. At the same time, the operational questions increasingly extend beyond classic IT responsibilities. The topic touches workforce structure, quality management, delivery processes, governance, training, supervision, organizational behavior, operational risk and cost management all at the same time.
HR covers some of these dimensions, but usually from a people perspective rather than an artificial work perspective. Cybersecurity and Legal focus heavily on restriction and risk reduction, which is necessary but insufficient on its own. Business units optimize locally for productivity and delivery pressure, often without a broader organizational model.
And that is probably why many current AI setups are fragmented or temporary. The structures around artificial work often emerge indirectly through AI task forces, Digital Labs, local champions, governance boards, CAIO roles, transformation initiatives or informal power users inside teams. All of these approaches can work to some degree. But they also reflect that organizations are still searching for a stable operating model around artificial work.
Organizational Function | Typical Focus Today | Gap Around Artificial Work |
IT | Platforms, tooling, architecture, integrations, operations | Usually not responsible for defining practical usage models, validation logic or operational quality of AI-generated work |
HR | People, roles, workforce development, organizational culture | Focused on human workforce structures rather than artificial work and AI-supported delivery processes |
Cybersecurity & Legal | Risk reduction, compliance, data protection, restrictions | Strong on control and policies, weaker on enabling practical and scalable day-to-day usage |
Business Units | Productivity, delivery pressure, local optimization | AI usage often grows pragmatically and inconsistently without shared organizational standards |
Digital Labs / Innovation Teams | Experimentation, prototyping, early adoption | Often temporary structures without long-term operational ownership |
CAIO / AI Leadership Roles | Coordination, strategy, governance alignment | Frequently broad strategic roles without operational control over daily artificial work |
Artificial Resources (AR) | Cross-functional management of artificial work | Potential future capability combining governance, delivery, quality, cost control and operational AI usage |
Comments