Named Owner of the Residual

What remains of a 'Chief Technology Officer' once the technology is operated by agents, and there are no officers left to be chief of? The question usually arrives as a joke about headcount. It deserves a colder answer. What remains is a name.

The 'residual' - in the strict statistical sense, observed minus modeled - is whatever survives of a system's behavior after everything that can be written down, rehearsed, or automated has been. It has two axes. On the axis of decision it is the judgment no policy covers; on the axis of operations it is the failure no plan covers. Both are measured against the same thing, the capable default, and the capable default is now a machine that improves every quarter. The residual is what the machine leaves.

Infrastructure supplies the standard instance. A routine, scheduled change to a critical system - approved, precedented, performed by competent hands - destroys the component it was meant to maintain, through an interaction nobody had modeled. The fallback, which had been prepared, agreed with everyone concerned, and never exercised, fails in its turn on a defect that only execution could have revealed. Service returns in fifteen minutes because one person knows where in the system an undocumented intervention is possible, is permitted to make it by hand on a live system, and will answer for the result if it makes matters worse. Every layer of preparation failed. The name held.

How does 'ownership' work? The machinery is not hard to describe, once its three parts are pulled apart:

(1) Knowledge: where the system can be reached into, what can be altered under pressure.

(2) Authority: permission to do an unreviewed and risky thing at once.

(3) Accountability: a standing relation between a person and everyone in a position to punish them.

Agents hold the first (frequently better than the incumbent), can be granted the second by policy, and can bear the third no more than a runbook can.

In sum, automation strips ownership for parts, carrying off every transferable component and leaving behind the one that was never a skill. It follows, therefore, that the function can be shrunk without bound and never to zero. "One person plus AI", the slogan of the moment, turns out to be a description of the residual owner rather than a staffing plan.

Anyone for whom the 3am phone call is an abstraction can stop here; nothing below will cost them anything. Whoever takes that call should check the arithmetic.

Accountability obeys no conservation law. It is a tree whose root always stands at one hundred percent; delegation partitions the root's share among the branches and never reduces it, so an engineer replaced by an agent adds nothing to a total that was already total. What rises is exposure - the probability that the next incident reaches the root undamped, multiplied by the cost of handling it unrehearsed. Each departure raises both factors. The product does not add. It multiplies.

The cost factor has a literature. Lisanne Bainbridge described the mechanism in 1983: automate the routine, and the human keeps only the hardest cases, monitoring a system no longer operated by hand and taking over exactly when taking over is hardest. Air France 447 is the standard exhibit. The owner of the residual under agents is out of practice at precisely what is left. (Fifteen minutes becomes two hours, if it becomes anything.)

Nor can the agent be the last plan, since the last plan must survive the failure of the thing it protects. An agent needs a network, name resolution, multiple remote endpoints; when the outage is the network, the agent is inside the outage. The human needed a terminal and a text editor. A double failure lies, by definition, outside the tested envelope - where machine confidence is cheapest and worth least.

The engineers who fear the agents are reading the situation correctly. Knowledge and authority migrate to the machine while accountability stays put, which leaves a person carrying the consequences of actions neither performed nor approved in detail. (Every employee has met this arrangement before, in the form of a manager.) The residual is unbounded only until someone writes down, for each class of action, who reviews it, what it can break, and how it is undone.

To propose that a system be run unsupervised by agents is to propose owning that system's residual. A board cannot own it in the operational sense: it holds authority and the root accountability, but no knowledge, and it cannot take the 3am call. It can own a residual only in its own currency, by naming an owner and paying the premium for the position. A mandate that does neither manufactures an unowned residual, which drifts to whoever is left - a strategy slide in the boardroom, a 3am call for someone who was never asked. A free single point of failure is invisible on the ledger, and the invisibility is the risk. The board was accountable all along; it chooses only whether to know.

The owner's standing duty is to shrink the residual, shipping everything documentable into the default until what cost fifteen minutes of attention costs fifteen seconds. This is a job description with an asymptote built in - self-liquidation that cannot complete. The better it is done, the less there is of the doer, down to the signature and no further. This essay is itself such a shipment: a piece of judgment written out, and thereby made available for any agent to recite. (It was drafted by one. It can hold the argument. It cannot be held to it.)

The whole matter compresses into one sentence:

The residual is the part of the job that cannot be delegated to a plan or an agent, and therefore must be delegated to a name.

What a name is, the sentence does not say. A name is the unit in which a society stores blame, and what it stores there is ruin - dismissal and liability. Work goes to the machines, then knowledge, then most of the deciding. Liability stays. Operations is a repeated game with unbounded downside, in which ruin is a matter of time rather than probability, and the position now concentrates on the one party least able to diversify it. 'We' had assumed the last human job would be something only a human could do. It is something only a human can suffer.

Nobody has yet priced the premium …

Written by

l4rz