Traces, Power, and Mediated Work

Abstract

To understand AI-mediated work, leaders need to stop looking only at named roles on an org chart and start looking at traces. The real action is often visible in prompts, comments, approvals, exceptions, version history, dashboards, rules, and handoffs. Those traces show how coordination actually happens, where power sits, and how humans and tools mediate one another.

This essay uses three concepts carefully. One explains how power works through observation, normalization, and examination rather than only through direct command. Another explains why agency in modern work is distributed across humans, interfaces, policies, artifacts, and systems. A third explains the risk that summaries, dashboards, and generated outputs become a substitute for the thing they are meant to represent. None of these lenses should be used as academic performance. They are only useful if they make leadership more precise.

The practical point is simple: if you cannot trace how work is made, you cannot redesign it. In an AI-heavy environment, the trace is not a side effect. It is the evidence of the operating model.

Start With Traces

Most organizations still describe themselves in terms of roles.

The sales manager owns sales. The product lead owns product. The engineer builds. The operator coordinates. The executive decides. That language is useful as far as it goes, but it hides the real mechanism of work. In an AI-mediated environment, the visible role often matters less than the sequence of traces that actually shape outcomes.

A trace can be any artifact left behind by action:

a prompt
a comment
a redline
a ticket
a Slack thread
a review note
an approval
a metric
a version diff
a log entry
a policy exception
a dashboard glance

Those traces tell you more than a job title does. They show what signal triggered the action, what constraints governed it, what got reviewed, what got ignored, and what was actually authorized.

That is especially important in AI-mediated work because the most consequential activity often happens between roles rather than inside them. A human writes a prompt. A system generates a draft. Another human edits it. A policy blocks one path. A manager approves another. A dashboard reports success. The org chart may still say the same thing, but the work has already been reorganized.

So the first operational question is not, "Who owns this role?" The first question is, "What traces show how this work is really being done?"

Power Is In The Arrangement

One useful lens moves the conversation away from command-and-control caricatures of power.

In this account of disciplinary power, control is not only exercised through a boss issuing instructions. It also works through observation, normalization, and examination. Power is embedded in systems that make people visible, comparable, and correctable. The important point for business is not the historical detail. It is the mechanism: people adapt to the regime they are placed inside.

That matters for AI work because AI systems do not just produce output. They change the way work is observed and judged.

Examples:

a dashboard makes one form of performance legible and another invisible
a review system changes what people pay attention to
a prompt library standardizes how a task is framed
a policy model makes some exceptions cheap and others expensive
a system of approvals teaches people what they are allowed to attempt

In other words, the structure of visibility changes behavior.

That is where this kind of reading becomes useful. Not as ideology. As diagnosis.

If a team says it wants to be innovative but every new idea must pass through ten gates, the traces will show that the organization is not merely cautious. It is producing a particular kind of subject: one that learns to wait, conform, and reduce ambition to what the system rewards. If the team says it wants ownership but all decisions are hidden behind an AI summary with no review path, the traces will show something else: a false autonomy, where responsibility is blurred even if output is fast.

This is why "human-centred" cannot be a vague moral slogan. It has to be visible in the traces.

You should be able to see:

where the human framed the problem
where the human authorized the move
where the system made the option space wider
where review happened
where accountability sits after the fact

If those traces are missing, the organization may still say humans are in charge, but the actual power arrangement may tell a different story.

Actors, Scripts, And Mediation

Another useful correction is to treat agency as distributed rather than neatly contained inside a single human actor. Work is carried by collectives of humans and nonhumans related in specific ways. Tools have scripts. Interfaces invite some actions and discourage others. Artifacts shape behavior. The result is not pure human control and not machine control either. It is mediation.

This is a good fit for AI-mediated work because modern systems are never just "people doing things with tools." They are composite arrangements:

people
software
documents
prompts
templates
permissions
models
workflows
incentives
history

Once you see that, the question changes. The question is no longer whether the agent is smart enough. The question is what the system is inviting the agent, and the human, to do.

That is the practical meaning of a script. A script is not a screenplay. It is a structured invitation to act in a particular way. A button implies one sequence. A form implies another. A review queue implies a third. A prompt template can steer the entire shape of the output before the human even starts.

In that sense, AI work is never just about the model. It is about the whole composite arrangement around the model. The strongest organizations are the ones that design those arrangements deliberately.

That gives Reworking Order a useful rule:

If the system changes behavior, treat the interface, policy, and artifact as part of the operating model, not as plumbing.

This is also why traces matter so much. They show which actants are doing the work. They show whether the model is merely generating options or quietly becoming the default authority. They show whether the human is actually deciding or merely rubber-stamping.

When Representation Replaces Reality

The danger here is that representation can start to replace reality. In AI work, this shows up whenever the summary becomes more trusted than the underlying work, or when the dashboard becomes the work itself. A clean-generated output can create a comforting illusion of completeness. A polished status report can hide the lack of actual understanding. A well-worded AI answer can become a simulacrum of clarity.

That is not a philosophical edge case. It is a management problem.

If the organization starts making decisions off summaries that no one has actually audited, the trace chain begins to break. People no longer know whether the output reflects the world, a model, a policy, or a prior summary. At that point, the system becomes hyperreal in the practical sense: the representation is treated as if it were the thing itself.

Leaders should be especially careful here because AI makes this failure mode seductive. The output is fluent. The report is neat. The slide is polished. The summary feels exhaustive. But the work underneath may be thinner than it looks.

That is why traces are the antidote.

If the system is healthy, the trace chain can answer:

what was observed
what was inferred
what was generated
what was checked
what was approved
what was changed
who is still accountable

Without that chain, the organization can drift into a world of perfect-looking representations and weak reality.

What Leaders Should Actually Look For

If you want to understand AI-mediated work in practice, stop asking only about roles. Start mapping traces.

Look for:

where the first prompt comes from
what policy shapes the prompt
what the model returns
what humans edit
what gets approved automatically
what gets escalated
what exceptions are frequent
what version of truth the team uses
which dashboards are trusted
which summaries get copied into decisions

Then ask:

where is power exercised?
where is it hidden?
where is it visible?
what behavior is being normalized?
what part of the workflow is actually being changed?

That is a more serious operating question than "how many agents do we have?"

It also gives the practice a better way to talk about transformation. Reworking Order is not about imposing a theory on organizations. It is about making the traces legible enough that leaders can see where coordination, power, and mediation are actually happening.

That is the real intervention.

Diagram

The trace chain is the operating model made visible. What matters is not just the output, but the chain of prompts, policies, reviews, approvals, and releases that produced it.

Why This Matters For Reworking Order

This essay does not replace the core thesis. It sharpens it.

If AI changes the task chain before it changes the org chart, then the organization must learn to read the chain. Traces are how you read it. They reveal the real order of work. They show whether humans are framing and authorizing change, whether the tools are merely assisting, and whether the system is becoming legible or merely more automated.

That is why these concepts matter in a practical practice:

one helps you notice how power is arranged
one helps you notice how agency is distributed
one helps you notice when representation is replacing reality

None of those lenses should become a performance. They should remain instruments.

The practical outcome is a better business question:

Can we trace how this work is actually made, and does that trace support the kind of organization we want to become?

If the answer is yes, then the work is probably being redesigned well. If the answer is no, then the organization is probably optimizing appearances faster than it is improving reality.

That is the distinction Reworking Order exists to make visible.