Global — Major artificial intelligence firms have reportedly concluded, after what several internal documents describe as “the longest and most expensive feasibility study any of us has ever participated in,” that C-suite executives remain unusually difficult to replace with automated systems. The finding, circulated quietly among research leads at three of the largest model laboratories, identifies a portfolio of leadership competencies that current architectures cannot reliably reproduce, and characterizes those competencies, in the working language of the working paper, as a “non-computable corruption skillset.”
CLASSIFICATION: EXECUTIVE LABOR ANALYSIS
DISTRIBUTION: Compensation Committees, Workforce Strategy Observers, Frontier AI Labs, Anyone Who Has Ever Been Promised That Their Job Was Safe
PREPARED BY: The Externality Research Division
DATE: May 2026
The conclusion, reached after roughly eighteen months of dedicated effort, reportedly reverses the prevailing industry assumption that executive labor — characterized in earlier white papers as “mostly meetings, mostly text, mostly the kind of pattern recognition we are very good at” — would be among the first categories of high-compensation work to be substantially automated. Researchers familiar with the project now describe that earlier framing as “catastrophically incomplete,” on the grounds that the prior analysis had focused on what executives produce rather than on what executives are required to do in order to produce it.
“Trust us. We tried.”
— Spokesperson, one of the AI firms involved
The spokesperson’s statement, delivered with what one attendee described as “the careful flatness of an executive who has, in fact, tried,” has been widely interpreted as the most candid public acknowledgement to date that the automation timeline for senior corporate leadership has been adjusted outward, and that the adjustment is not expected to be revised inward at any point on the horizon currently legible to the firms involved.
The Findings
The working paper, which has reportedly circulated under three different cover sheets and at least two revised titles, identifies a set of executive functions that current systems can reproduce competently and a separate set of functions that current systems cannot reproduce at all. Researchers note, with what one reader described as “a tone you usually only hear from analysts who have just discovered the moat is wider than they thought,” that the two sets are not balanced. The first set is large and growing. The second set is small, durable, and worth, in the working paper’s phrasing, “substantially more than the first.”
What AI Can Do
Across the categories tested, the working paper reports that current systems are able to:
- Summarize reports. Including the reports the executive was supposed to have read, the reports the executive intends to claim to have read, and the reports the executive intends to attribute to a deputy in the event that the report is later determined to have been incorrect.
- Optimize workflows. Including the workflows the executive is publicly committed to improving and the workflows the executive is privately committed to preserving on the grounds that they generate the meetings at which the executive is observed to be present.
- Generate strategy memos. Including memos that are subsequently approved, memos that are subsequently disowned, and memos that are subsequently both approved and disowned by the same individual in different rooms on the same afternoon.
The working paper notes, with what one analyst characterized as “the slightly defeated enthusiasm of researchers reporting genuine technical progress in a domain that has just been judged irrelevant,” that performance on these categories has improved substantially over the past eighteen months and now equals or exceeds the performance of the human comparison group in blind evaluation. The comparison group, the paper notes parenthetically, was composed of incumbents.
What AI Cannot Do
The same systems, the paper reports, continue to struggle with a category of competencies that researchers have grouped under the heading high-level institutional manipulation. One engineer associated with the study summarized the gap as follows:
“The model can recommend restructuring. It cannot yet smile while doing it.”
The distinction, which appears in three separate working documents and which one researcher described as “the sentence that ended the project,” captures what the paper identifies as the central obstacle: the systems are entirely capable of identifying the action and articulating the rationale, but they are not capable of delivering the action under the social and emotional conditions in which the action is required to be delivered. Executive labor, the paper argues, is not the recommendation. It is the delivery of the recommendation under conditions specifically designed to make the recommendation difficult to refuse.
The Non-Computable Skillset
The paper enumerates four core competencies that the executive class has reportedly retained in the face of otherwise broad automation pressure. The competencies appear, in the working paper, in the order in which the research teams reportedly gave up on reproducing them.
Cartel Formation
The first competency, and the one the paper describes as “the most stubbornly resistant to automation,” is the ability to identify a small number of structurally aligned counterparts at notional competitors and to arrive, through a sequence of unrecorded encounters, at a shared understanding regarding pricing, hiring, and the geographic distribution of effort. Current systems, the paper notes, can propose the arrangement. They cannot conduct the dinner at which the arrangement is, without ever being written down, agreed.
“The model can produce a memo describing the cartel,” one researcher explained. “It can even produce the memo in language designed to be plausibly deniable. What it cannot do is hand the memo to a peer at a private club, in a way that the peer reads it as a proposal rather than a confession, and walk away with the peer’s tacit agreement and no documentary trail. That is a different skill. That is the skill we cannot reproduce.”
Political Purchasing
The second competency identified is the practice of converting corporate funds into regulatory outcomes through a sequence of legally permitted instruments — campaign contributions, policy fellowships, the endowment of academic centers, the strategic provision of board seats — the cumulative effect of which is reportedly difficult to characterize in any single document and impossible to characterize in any single invoice. The paper notes that the practice has been studied extensively and that the literature has, on the whole, declined to use the word that the practice most closely resembles.
“The model can identify the regulators,” the same researcher continued. “It can identify the universities they came from, the foundations they intend to join after, and the policy positions they are currently in the market to be persuaded toward. What it cannot do is invite the regulator to dinner and produce, between the second and third courses, the kind of warm, decade-long relationship in which the eventual rulemaking is not so much purchased as it is inherited. That is, again, a different skill.”
Employee Fear Calibration
The third competency is the ability to maintain workforce productivity through a finely tuned application of anxiety. The paper notes that the relevant variable is not the total quantity of fear in the workforce but its distribution: enough to deter overt dissent, not so much as to trigger external attention; enough to suppress wage negotiations, not so much as to produce visible attrition; enough to keep the meeting polite, not so much that the meeting becomes a story.
Current systems, the paper reports, are able to model the distribution. They cannot produce it. The production of the distribution, researchers note, requires what the paper calls a credible source of sanction — a human being whose presence in a room implies, without ever being stated, that the room contains people whose continued employment is contingent on the human being’s continued favor.
“Fear is a delivery mechanism. The model can write the message. It cannot be the messenger.”
— Senior researcher, on the calibration problem
Strategic Blame Displacement
The fourth and reportedly most lucrative competency is the ability to ensure that adverse outcomes — layoffs, divestitures, restatements, indictments — are attributable, in the public record and on the relevant earnings call, to actors other than the executive whose decisions produced them. The paper notes that the competency has multiple components, including the construction of intermediate decision-makers whose authority is sufficient to bear responsibility and insufficient to deflect it, the maintenance of plausible alternative explanations in the form of market conditions, regulatory environments, and unforeseeable events, and the cultivation of a tone of resigned stewardship under which the executive is understood to be doing his level best within constraints that were, on careful examination, also produced by him.
Current systems, the paper notes, are able to identify the displacement. They cannot perform it. The performance, the paper argues, requires a degree of self-presentation that the systems have not yet achieved and that researchers are reportedly “no longer optimistic that they will.”
The C-Suite Advantage
Beyond the four core competencies, the working paper identifies a secondary set of executive functions that the research teams concede they have not seriously attempted to reproduce. The list, included in an appendix that one analyst described as “the only honest part of the document,” appears below in the order in which the research leads reportedly volunteered to drop them from scope.
Making Layoffs Sound Inspirational
The paper notes that current systems can compose announcements of workforce reductions in clear, neutral language. They cannot compose them in the language in which such announcements are conventionally composed: a language that combines gratitude toward the departing, optimism toward the remaining, admiration toward the leadership, and an implication that the underlying decision is, in some difficult sense that nobody is asked to articulate, an act of courage.
“The model writes the letter,” one researcher reported. “The model writes a fine letter. The model does not write the letter that ends with the phrase ‘we are bigger than this,’ in a tone that makes the reader, against their better judgment, briefly believe it. That is a different letter.”
Turning Failure Into Market Conditions
The paper describes a related competency in which executive performance failures are reframed, on the relevant earnings call, as externally imposed difficulties that were unforeseeable, generalized, and in any case affecting the entire sector. The reframing, the paper notes, has the additional property that it is delivered in the same tone the executive uses to claim credit for the previous quarter’s favorable conditions, which were, the executive will note, not external but the product of disciplined leadership.
Current systems, the paper reports, are technically capable of the reframing but lack the institutional credibility to deliver it persuasively. “The model can say the conditions changed,” one analyst noted. “It cannot be the person whose job was to anticipate the change, saying that the change was unforeseeable, while continuing to be the person whose job that was. The reframing requires the speaker. It requires the same speaker who made the prior commitment. The model is not, in any meaningful sense, that speaker.”
Extracting Bonuses During Collapse
The paper notes that the executive class has reportedly demonstrated, across several decades and many documented cycles, an ability to maintain individual compensation during periods in which institutional performance has materially declined. The mechanism, the paper observes, is not concealment. The compensation is publicly disclosed. The mechanism is the construction of an analytical framework under which the compensation, on examination, is not only justified but, in light of the difficulty of the period, modest.
“The retention argument is the key piece,” one researcher explained. “The argument is that the executive must be retained, that retention is expensive, and that any reduction in compensation during a period of poor performance would risk the executive’s departure to a competitor where they would, presumably, perform comparably poorly at higher cost. The model can identify the argument. It cannot, with a straight face, make it.”
Calling Fear “Alignment”
The fourth and most semantically advanced competency identified by the paper is the systematic renaming of coercive workplace conditions as cultural achievements. The competency, the paper notes, is the engine that permits the prior three to operate without organized resistance. Workforce anxiety, when correctly named, becomes engagement. Forced consensus, when correctly named, becomes alignment. Unpaid overtime, when correctly named, becomes ownership. Resigned compliance, when correctly named, becomes culture.
Current systems, the paper reports, are able to identify the substitutions. They are not yet able to deploy them in a way that the recipient does not notice the substitution has occurred. “The model uses the words,” one researcher noted. “But the model uses the words like someone using a second language. The executive uses the words like someone who has forgotten there was a first.”
“Their cunning is the moat.”
— Internal report, summary conclusion
Investor Reaction
Markets reportedly responded to the findings with what one fixed-income analyst characterized as “the calm reassurance of capital recognizing that the people responsible for protecting it have not, in fact, been made redundant.” Executive compensation futures, an instrument that does not formally exist but that several analysts have nonetheless modeled, are reported to have firmed.
One analyst, asked to summarize the institutional reaction, stated the following:
“As long as AI cannot form a backchannel coalition over steak dinners, executive labor remains defensible.”
The analyst clarified, in a follow-up note distributed only to clients above a specified tier, that the steak dinner was not a metaphor. The steak dinner was the actual production technology of executive labor. The remaining functions — the memos, the slides, the quarterly calls — were, in the analyst’s view, the visible portion of the work, and the visible portion of the work was already being automated, at considerable expense to the firms that had not yet noticed the visible portion was not the part that mattered.
Compensation Committee Response
Compensation committees at several large public companies have reportedly cited the working paper, in documents not yet entered into the public record, as supporting evidence for the retention of existing executive pay structures. The argument, one observer noted, is that the working paper establishes the executive class as a non-substitutable category of labor, and that non-substitutable labor commands the premium it currently receives, and that the premium is therefore not, properly understood, a premium at all but the market-clearing price of an irreplaceable input.
The argument has reportedly been received within the relevant committees with what one attendee described as “the satisfaction of a body that has just been handed, free of charge, the conclusion it was going to reach anyway.”
Workforce Response
Workforce reaction has reportedly been more muted. Internal surveys conducted at several of the firms involved in the underlying research have reportedly revealed a category of employee response that researchers are calling asymmetric automation awareness: the recognition that the same systems being deployed to summarize, optimize, and replace the worker’s output are, on present evidence, not being deployed against the worker’s management.
“The employees noticed,” one researcher stated. “They noticed quickly. They have, on the whole, declined to say so out loud, which is itself a behavior the working paper would like to flag for future study.”
Methodological Notes
The Selection Problem
The working paper acknowledges, in a section appended late and reportedly under pressure from internal reviewers, that the research was conducted by individuals whose own compensation was determined by the same class of executives being evaluated for replacement. The acknowledgement notes that the researchers considered the conflict, discussed the conflict, and ultimately determined that the conflict was, on balance, “manageable.”
External reviewers have characterized the acknowledgement as “the most honest sentence in the document,” and the determination that followed it as “the second most honest sentence in the document, in the specific sense that the determination did not survive the next paragraph.”
The Benchmark Problem
The paper notes that the benchmarks used to evaluate executive performance — quarterly earnings, share price movement, year-over-year revenue growth — were established by previous generations of executives operating under conditions broadly favorable to the establishment of benchmarks they were likely to meet. The research team acknowledges that the systems were therefore evaluated against a moving target whose movement was, in the relevant historical record, almost entirely upward.
“The model performed adequately against the benchmark,” one researcher noted. “The benchmark, on examination, was not difficult to perform adequately against. The benchmark, on further examination, was designed by people whose performance was being measured by it. We are unable to characterize this design feature in language the legal department has approved.”
The Counterfactual Problem
Several reviewers reportedly raised the question of whether the relevant comparison was between current systems and current executives, or between current systems and the executive class’s actual contribution net of the costs the executive class imposes on the rest of the organization. The paper notes that the latter comparison was, on internal vote, declared out of scope.
“That was the comparison we were not going to be allowed to run,” one researcher stated. “We knew that going in. We ran the other one. The other one came out the way we expected.”
Implications
The paper’s implications, drawn out in a final section that several reviewers reportedly objected to and that survived in modified form, may be summarized as follows.
- The automation frontier is asymmetric. The categories of labor most exposed to automation are not the categories that produce the highest compensation. The categories that produce the highest compensation are, on the present evidence, the categories the same systems have been least able to reproduce.
- The moat is social. The remaining executive function is not an analytical task that can be performed by a system. It is a social technology that requires a specific kind of person, embedded in a specific kind of network, with a specific kind of plausible deniability. The technology is not improvable by computation.
- The savings are downward. Productivity gains from automation are, on the working paper’s evidence, accruing to the executive class rather than being distributed to the workforce whose labor is being automated. This is described in the paper as “an expected and self-reinforcing pattern.”
- The framing is durable. The same executive class identified by the paper as non-substitutable retains, the paper notes, the authority to commission the next study and to determine which conclusions, of those reached, will appear in the next study’s executive summary.
Industry Response
Public reaction from the AI laboratories involved has, on the whole, taken the form of measured statements emphasizing that the research is ongoing, that the methodology is being refined, and that the firms remain committed to expanding the frontier of automatable labor. A second statement, distributed through a different press contact at one of the same firms, emphasized that the firm’s own executive leadership remained “deeply involved” in the research direction, and was “personally committed” to ensuring that the firm’s tools were applied where they could do the most good.
Researchers familiar with both statements have reportedly noted that the two statements were issued in the same week and were not, on present evidence, in active tension.
The Continued Effort
The firms have reportedly stated that replacement efforts will continue. One spokesperson, when asked what specifically the next phase of research would entail, declined to specify, citing the proprietary nature of the underlying methodology. A second spokesperson, asked the same question, stated that the work would focus on “the long tail of executive functions” and that he was confident the laboratory’s next generation of systems would “close meaningful portions of the gap.”
He declined to specify which portions, on what timeline, and against which benchmarks. The decline was delivered, attendees noted, in a tone that closely resembled the tone characterized elsewhere in the working paper as the non-computable one.
Bottom Line
- What Happened: Several major AI firms have reportedly concluded that the C-suite is difficult to replace with current systems, on the grounds that the systems lack a set of competencies that the working paper characterizes as a non-computable corruption skillset.
- What That Means: The categories of labor most exposed to automation are downstream of the categories that direct the automation. The automation is therefore being directed downward.
- Why It Matters: The argument that automation pressure is borne broadly across the income distribution is, on the present evidence, not consistent with the present evidence.
- What Happens Next: Replacement efforts will continue. The C-suite will continue to direct them. The next study will be commissioned. The conclusions of the next study will not, on present evidence, surprise the commissioners.
Closing Statement
AI firms say that replacement efforts will continue. The firms emphasize that current findings are preliminary, that the methodology will be refined, and that the frontier of automatable executive labor will, with sustained investment, expand. The firms have declined to specify a timeline.
For the moment, the C-suite remains secure.
At press time, executives were reportedly reviewing the findings.
Then taking credit for them.
Editorial Footnotes
- The working paper has not been published in any peer-reviewed venue. Researchers reportedly noted, at various points in the drafting process, that the relevant peer-reviewed venues are themselves edited by individuals whose compensation is set by the population being studied.
- Quoted remarks from researchers and analysts are illustrative; the institutional patterns described are characterized in service of the broader thesis.
- The phrase “non-computable corruption skillset” appears in three separate internal documents and one external email that was reportedly never intended for external distribution.
- This document is a satirical reconstruction. The satire concerns the framing. The framing, the working group reportedly notes, is the part that is not satirical.