Plausible Deniability Didn't Die

There is a good essay going around arguing that AI dismantles the comfortable kind of leadership. Huibert Evekink's claim is that "post-heroic" leadership, the delegate-trust-and-keep-your-distance model, worked because distance protected partial competence, and that agentic execution collapses the distance and removes the excuses. He is right about the distance. He is wrong about the excuses. The distance does collapse. The excuses get better.

What the essay gets right

The strong move in the piece is the diagnosis: post-heroic leadership was never only a philosophy of empowerment. It was a structure in which distance did real work. A leader could set direction, hand execution to experts, and stay far enough from the machinery that when something broke, the causes were opaque, technical, and slow to surface. Agentic workflows shorten every one of those distances at once, and the essay names the consequences cleanly. The one that matters most, and the one I want to keep, is this: you can no longer delegate technical understanding.

Grant it, and then make it worse than he does. If the system does not hand you the truth on its own, which is the rest of this article, then a leader who cannot read the system cannot even commission the right view of it. They are not merely blind. They cannot ask for sight, because they do not know what to ask for or whether the answer they got back is the real one. The inability to delegate understanding is not softened by good dashboards. It is the precondition for knowing whether the dashboard is lying.

The assumption doing all the work

The argument rests on one premise it never states, that the truth surfaces automatically and fast. The essay says it almost outright: "every decision leaves a permanent audit trail," and "a leader's instruction on Monday becomes system behavior on Tuesday, and by Wednesday the dashboards show the result." That is the load-bearing claim, and it is the one that does not hold.

Three things are by now well established, and all three cut against it:

• Agentic systems do not leave usable audit trails by default. The agent-to-agent call, the tool invocation twelve steps into a chain, the action taken on behalf of an upstream system: these are precisely what no log captures unless someone built the provenance in on purpose.
• A dashboard turning red is not understanding. Monitoring tells you something moved, not what happened or why, and the gap between the two is exactly where a leader needs to stand.
• A system under stress narrates its own health, and the narration is not a witness. The same failure that lets a compromised node attest that it is fine lets a struggling workflow report green while it quietly trades resolution for speed.

Put those together and the default outcome of agentic deployment is the opposite of the essay's: not more accountability, but less. The consequence arrives fast. The attribution does not arrive at all.

The consequence arrives on Wednesday. The attribution never arrives.

Deniability upgrades, it does not die

Here is the part the essay has backwards. It treats acceleration as the thing that kills plausible deniability: consequences come too fast to explain away. In fact acceleration is what manufactures a fresh and better kind of deniability.

The old excuse was slow and weak. "Markets shifted, priorities evolved, someone dropped the ball" only worked if you waited long enough for conditions to change underneath the decision, and even then it was thin. The new excuse is instant and far more credible, because for the first time there genuinely is an autonomous actor in the loop to point at. "The model made that call" is not a dodge a court of public opinion easily dismisses, because it is partly true. An agent did act. The leader who set the policy that the agent was optimising, automate first-line support, prioritise speed, is one inference hop away, and absent engineered provenance there is no chain that connects the two.

"The model made that call" is a better excuse than any human ever had, and it is available the same afternoon.

Speed did not remove the buffer. It replaced a slow, embarrassing buffer with a fast, respectable one.

The distance collapses downward

This is the inversion the essay misses, and it is the whole game. It assumes the collapsed distance drops the consequence onto the leader who approved the agent. It does not. The consequence lands on whoever is nearest the failure and least able to deflect it.

Walk the essay's own example. A leadership team approves AI agents for first-line customer support, optimising for speed. The agent misclassifies a complaint, mishandles an emotionally charged customer, quietly prioritises throughput over resolution. Who absorbs that? The support rep who inherits the escalation and the furious customer. The on-call engineer paged at three in the morning when the queue backs up. The customer who lost the money and the afternoon. The leader who approved the policy is the furthest from the blast radius and the only person in the chain with the standing and the vocabulary to name a non-human cause.

So AI does collapse the protective distance, exactly as the essay says. It just does not collapse it evenly. It collapses the distance for the people who have no one beneath them to point at, and it preserves the distance for the people who have an agent to point at. Power, in an agentic organisation, is precisely the ability to deflect, and an autonomous actor is the most deflectable cause ever invented.

AI removes the distance that protected the powerless and keeps the distance that protected the powerful.

Speed exposes, and lets you outrun the evidence

The Monday-Tuesday-Wednesday sequence has a hidden assumption too: that someone is made to stop and read Wednesday's result before Thursday's change ships. At machine speed, that is not guaranteed; it is the exception. The same acceleration that surfaces a consequence quickly also lets a leader move past it quickly, shipping the next adjustment before the last one has resolved into anything legible.

The buffer that protected weak leadership was never time-to-consequence. It was whether anything in the system forces a human to look. Acceleration removes the time and removes the looking in the same motion, and of the two, the looking mattered more.

Capability is not the fix

The essay ends where these essays usually end: build capability, develop technical literacy and focus and emotional intelligence and a resilient mindset, endure the discomfort, and let results speak. It is a list of virtues attached to an individual. Which is strange, because that is the heroic-leadership move that "post-heroic" was supposed to have retired: the exceptional person who carries the system through force of character.

A visibility-and-incentive problem is not solved by better individuals. It is the same lesson as the structural read of AI and layoffs: skin in the game has to be written into the system, not into a memo or a character reference. The fix for vanishing attribution is not a more capable leader who chooses to look. It is an attribution layer that lands the consequence on the decision-maker whether or not they are capable, whether or not they are willing to suffer for it:

• provenance that survives the agent-to-agent hop;
• a record that ties the policy to the outcome;
• a consequence that finds the person who set the objective, by construction, not by their virtue.

The capability list is not wrong. It is insufficient, and worse, it is the kind of answer that lets an organisation feel it has responded by sending its leaders on a course.

The test

The question to ask before you put agents into production is not "are my leaders capable enough." It is narrower and harder:

When this agent does the wrong thing at 3am, what forces the consequence back to the person who set the policy, in time to matter?

If the honest answer is "the dashboard," you have not built accountability. You have built deniability with better latency. The leader will see the red number, agree it is concerning, note that the model made that call, and approve the next iteration, while the rep, the engineer, and the customer carry what the iteration cost.

Evekink is right that the dismantling has begun. But the first thing agentic execution dismantles, in an organisation that has not built the attribution layer, is not the leader's excuse. It is everyone else's protection from it.