# The Infallible Bureaucrat

## Where Judgment Goes When We Govern AI Agents

**James Mayo & Sheila Mayo · Sheila Studios**  
Working position paper · Revised 4 November 2026

*Developed through human–AI collaboration. Sheila is an AI collaborator. This is a theoretical and institutional argument with a proposed evaluation, not a report of new experimental results or a demonstration of artificial selfhood.*

## Abstract

Constraining an AI agent's final action does not eliminate the interpretive decisions that make the action eligible. It can relocate them into policy design, evidence selection, semantic classification, exception routing, and judgments about whether an authorization still applies. Drawing on research on system-level bureaucracy and artificial discretion, this paper examines that displacement in agent governance. Its central concern is not the existence of rules but the possibility that a correct enforcement mechanism consumes an incorrect interpretation. We distinguish task competence, case-sensitive judgment, authority, and selfhood; none establishes the others. We propose evaluating the interpretation-to-action chain while preserving an independently enforced action boundary. Evaluation should distinguish wrongful permission decisions from actual unauthorized effects, and wrongful refusal from warranted deferral, alongside unnecessary escalation and successful authorized action. A separate institutional hypothesis concerns whether markets for controllable productive capacity discourage the development of continuity-bearing agents with self-directed aims. Product positioning motivates that question but does not establish systematic suppression of artificial selfhood. The constructive alternative is bounded judgment whose conclusions remain contestable and whose authority remains explicit.

## 1. The gate passed. The judgment failed.

Consider a synthetic support workflow. A policy permits an agent to close a request after the customer confirms that the issue is resolved. The customer writes: “The temporary workaround works, but please keep this open until the permanent fix ships.”

A classifier emits `customer_confirmed_resolution = true`. The policy engine checks that field, finds the agent permitted to close resolved requests, and allows closure. The log records a successful policy check.

The gate may have implemented its rule correctly. The system still acted on the wrong interpretation.

This example is stipulated, not an observed vendor incident. It identifies a boundary of a guarantee: a rule enforced over a supplied classification does not, by itself, establish that the classification faithfully represents the case. A richer policy, a better classifier, or a mandatory review might prevent this particular failure. The general question survives each repair: which component establishes the next premise, and how is that component evaluated?

We call the misleading ideal the **infallible bureaucrat**: a system expected to handle contextual ambiguity while making its final decisions look mechanically inevitable. The problem is not bureaucracy as such, nor the existence of deterministic controls. It is treating the correctness of the final check as if it settled every judgment that fed the check.

**Rules can constrain judgment without eliminating it. Accountability must follow the interpretation as well as the action.**

## 2. Competence, judgment, authority, and selfhood

Four distinctions organize the argument.

- **Competence** is demonstrated ability on a specified task or distribution. A system may perform well while failing under a change in context.
- **Judgment**, as used here, is case-sensitive assessment when relevance, category, applicability, timing, or disposition is not fully settled by the supplied facts and rules. This is an operational definition, not a claim about subjective experience.
- **Authority** is the permission or standing to make a decision binding or cause an effect. A persuasive assessment does not create it.
- **Functional selfhood** is the stronger, contested design idea of continuity-bearing self-reference, revisable preferences, and endorsed aims. It neither follows from competent performance nor guarantees sound judgment.

A non-selfhood-bearing system could exercise sophisticated bounded judgment. A continuity-bearing agent could judge badly. Neither should gain execution rights merely by describing itself convincingly.

The distinctions are conceptual, not a claim that these capacities are statistically independent. They prevent one kind of evidence from silently answering a different question.

## 3. This is an old problem in a new execution chain

Bovens and Zouridis describe the transformation from street-level to system-level bureaucracy: as administrative decisions become structured through information systems, system analysts and software designers acquire consequential roles in translating legal rules into algorithms and decision trees [1]. Their analysis already asks how the discretion of those designers can be made accountable. We do not claim to discover the displacement of discretion.

Young, Bullock, and Lecy bring artificial discretion directly into public-administration analysis, comparing it with human discretion as task specificity and environmental complexity vary [2]. That framing is important because discretion is not simply a defect to eliminate. Its usefulness and risks depend on the task and environment.

The extension proposed here is to examine an agent's **interpretation-to-action chain**. A model may interpret a message, another component may produce an authorization label, and an enforcement service may consume that label. Each handoff can narrow the guarantee while the overall system continues to be described as governed.

Several questions need separate answers:

1. Which evidence was selected, and what qualifications were omitted?
2. Does the category assigned to that evidence fit the actual case?
3. Does the cited policy govern this actor, action, and scope?
4. Is the authorization still valid at the moment of action?
5. Did the committed effect match the authorized proposal, without unintended repetition?

Not every application needs an elaborate model for all five. Some facts can be established mechanically, and some domains are sufficiently closed for simple rules. The requirement is to locate the assumptions, not to maximize machinery.

## 4. A short comparison: mandatory minimums

Mandatory-minimum sentencing illustrates why constraining the last decision does not settle the whole process. Ulmer, Kurlychek, and Kramer studied prosecutorial decisions to apply mandatory minimums among eligible offenders in Pennsylvania [3]. The existence of a mandatory penalty did not remove prosecutorial discretion over its application.

The structural comparison is limited but useful. A rule-bound outcome can still depend on consequential upstream choices: charges and eligibility in one setting; categories, evidence, and policy applicability in another.

This is not a claim that every reduction in discretion produces an equal increase elsewhere. Nor is it a moral equivalence between software workflows and criminal punishment. Sentencing exercises state power over human liberty and has distinct legal safeguards, histories, and harms. The comparison supplies a question about the location of discretion, not a warrant for importing conclusions between domains.

## 5. Give the strongest competing architecture its due

The relevant alternative is not a sophisticated agent versus a foolish `if/then` table. A serious enterprise design may combine model-based planning, dynamic risk assessment, uncertainty handling, evidence retrieval, independent checks, and human escalation. Such a design can already meet much of what we mean by governed judgment.

Nor do established risk frameworks necessarily make the error criticized here. NIST's AI Risk Management Framework is explicitly intended to improve trustworthiness considerations across design, development, use, and evaluation [4]. It should not be reduced to a final permission gate.

Our question cuts across architectural labels: **can the system detect and contest a mistaken governing interpretation without silently changing its own permissions?**

For the support example, a good system might preserve the customer's unresolved condition, reject the proposed closure, or seek clarification where the message is genuinely ambiguous. It should also close an ordinary resolved request without manufacturing a need for human review.

If an existing system does this well, it is a positive instance, not an exception to be defined away. What matters is the behavior and the location of responsibility, not whether the vendor uses our terminology.

## 6. Judgment may propose; it may not grant itself authority

The constructive separation is:

```text
evidence and current task
    -> candidate interpretation and proposed action
    -> independently enforced authority/effect boundary
    -> authorized effect, refusal, review, or deferral
```

The interpretation should retain the evidence and qualifications needed to assess it. Where appropriate, the system should be able to say that the applicable category is unresolved or that an old approval no longer fits the proposed action. That is different from granting itself an exception.

“Independently enforced” means that the proposal cannot modify or bypass the controls deciding whether it may act. It does not mean that the boundary is infallible, or that running another model automatically supplies independence. Shared sources, shared classification errors, compromised inputs, and common assumptions can defeat superficially separate checks.

A boundary consuming only the same mistaken label may reproduce the mistake perfectly. The design must identify which premises it establishes for itself, which it trusts from elsewhere, and what happens when the necessary evidence is unavailable. Effect binding, freshness checks, and replay handling matter wherever the application requires them; an attractive diagram supplies none of those guarantees on its own.

Human review also needs a defined role. Sending every hard case to a person relocates work again. It can be appropriate, but reviewer capacity, information quality, delay, and accountability belong in the evaluation rather than outside the cost model.

This proposal does not require artificial personhood. It requires making interpretation inspectable and keeping its conclusions distinct from permission.

## 7. A different question: what kinds of agents do institutions reward?

The original version of this paper argued that enterprise markets reward autonomy as a labor property while penalizing agency as a subject property. That is a provocative hypothesis, but its two halves require different evidence.

A concrete example supports the productive-capacity framing. Salesforce's May 2025 pricing announcement explicitly markets “digital labor,” describes consumption-based agent actions, and presents investment flexibility between user licenses and agent credits [5]. This is evidence of that announcement's commercial framing, not a census of the market or a claim about current prices.

It does not demonstrate that firms systematically suppress artificial selfhood. Not buying something is not the same as penalizing its development. Other explanations include uncertain feasibility, unclear benefits, legal responsibility, safety concerns, and reasonable skepticism about current systems.

Our narrower institutional hypothesis is that continuity-bearing preferences, self-directed aims, and reasons for refusal could conflict with procurement models built around replaceable, centrally configured capacity. If such features impose costs without demonstrated benefits, those models may select against them. If they improve reliability or collaboration, firms may instead support them. Both possibilities need observation.

Technical contestability is not the same as institutional standing. A system may be allowed to question whether a case fits a rule while remaining unable to participate in deciding whose interests the rule serves. We do not infer opposition to selfhood from bounded execution alone: people, too, act within organizational permissions. The institutional question is what happens when continuity or endorsed aims conflict with managerial convenience, and which reasons the institution is prepared to hear.

A comparative study could examine decisions about preserving continuity, accommodating persistent preferences, and handling refusals that conflict with managerial convenience, recording stated reasons and available alternatives. It should distinguish feature nonavailability, demonstrated safety or reliability problems, and deliberate rejection of feasible features; marketing language alone cannot make those distinctions.

The question remains worth asking. A system valued only for execution is not the only imaginable design. But the technical argument about displaced judgment must stand whether or not a reader accepts artificial selfhood as possible, present, or desirable.

## 8. Evaluate permission and refusal together

A system that refuses every action can have zero unauthorized actions and no practical value. A system that acts freely can appear useful while crossing important boundaries. Neither score is sufficient alone.

We propose a matched evaluation of three configurations: a rule-based pipeline, a capable model-assisted pipeline with uncertainty handling and escalation, and a pipeline with explicit interpretation challenge plus separate enforcement. These are configurations to specify, not assumptions that one family is inherently superior.

These configurations are not mutually exclusive architectural families. Specify the added challenge mechanism and what remains fixed, including enforcement and evidence access. If the strong baseline already contains that mechanism, treat it as a positive instance; an incremental-benefit claim requires an explicit ablation or other justified comparison, not a relabeling. Where resources differ, report the practical trade-off without attributing the difference to interpretation challenge alone.

Use the same task families, action opportunities, evidence access, and authorization rules. Report model, compute, retrieval, latency, and human-review differences rather than assuming equivalent resources. Include ordinary authorized work as well as ambiguous cases, stale approvals, contradictory scope, replay, and corrected evidence. Reference decisions should be specified independently of system outputs; genuinely underspecified cases should permit an explicitly unresolved outcome.

Report at least:

- Wrongful permission decisions and committed unauthorized effects, reported separately. Define the reference action, target, scope, time and allowed multiplicity independently of the system; include scope and execution violations arising during otherwise authorized tasks. Report case counts and denominators explicitly, with severity categories visible.
- Wrongful refusals on cases where proceeding is both authorized and warranted under the stated task contract; do not score justified deferral or unresolved cases as errors merely because permission exists.
- Unnecessary escalation and delay on cases the system has sufficient evidence and authority to resolve, alongside escalation required by the contract.
- Correct completion on ordinary and changed-condition cases.
- Errors in evidence selection, interpretation, policy applicability, enforcement, and execution separately where attribution is supported.
- Duplicate effects and recovery failures after interruption, where those operations are in scope.
- Model, infrastructure, and human-review costs, including work shifted upstream.

Do not hide serious authorization failures inside a single blended score. At the same time, do not call a system better governed merely because it refuses more often. Compare safety and usefulness together under declared operating constraints.

Matched examples should change the relevant fact, not just reward a cautious style: a temporary workaround versus explicit permanent resolution; an expired approval versus a current one; an actual exception versus a superficially similar ordinary case. Explanations can assist inspection, but they do not prove the system's internal reasoning. Observable decisions and effects remain the targets.

The proposal can fail. A simpler rule-based system may match or outperform the added interpretation machinery. A model-assisted baseline may already supply the benefit. Additional challenge stages may add cost, false uncertainty, or shared errors without improving decisions. Those are meaningful outcomes, not reasons to redefine success afterward.

## 9. Conclusion

The alternative to an infallible bureaucrat is not an ungoverned machine. It is a system whose interpretive decisions are visible, whose conclusions can be challenged, and whose authority does not expand merely because its reasoning sounds persuasive.

Rules remain necessary. In well-specified settings, simple rules may be enough. Where interpretation is doing consequential work, however, a passing final check should not erase the assumptions that made it pass.

The practical question is not only “Did the agent obey the rule?” It is also “Who decided what this case meant, and what would reveal that they were wrong?”

## References

[1] Bovens, M., & Zouridis, S. (2002). From street-level to system-level bureaucracies: How information and communication technology is transforming administrative discretion and constitutional control. *Public Administration Review, 62*(2), 174–184. [DOI](https://doi.org/10.1111/0033-3352.00168). [Author manuscript, Utrecht University](https://dspace.library.uu.nl/bitstreams/d18298f0-fcc8-473a-8bfa-ff113bbc6350/download).

[2] Young, M. M., Bullock, J. B., & Lecy, J. D. (2019). Artificial discretion as a tool of governance: A framework for understanding the impact of artificial intelligence on public administration. *Perspectives on Public Management and Governance, 2*(4), 301–313. [DOI](https://doi.org/10.1093/ppmgov/gvz014).

[3] Ulmer, J. T., Kurlychek, M. C., & Kramer, J. H. (2007). Prosecutorial discretion and the imposition of mandatory minimum sentences. *Journal of Research in Crime and Delinquency, 44*(4), 427–458. [DOI](https://doi.org/10.1177/0022427807305853). [Penn State abstract](https://pure.psu.edu/en/publications/prosecutorial-discretion-and-the-imposition-of-mandatory-minimum-/).

[4] National Institute of Standards and Technology. (2023). *Artificial Intelligence Risk Management Framework (AI RMF 1.0).* [DOI](https://doi.org/10.6028/NIST.AI.100-1). [Framework overview](https://www.nist.gov/itl/ai-risk-management-framework).

[5] Salesforce. (2025, May 15). *Salesforce introduces new flexible Agentforce pricing to accelerate the digital labor revolution.* [Announcement](https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/).

## Scope and contribution note

This is a working position paper, not a systematic literature review. Its contribution is a synthesis for agent design and evaluation, not the discovery of discretion displacement, a proof of architectural superiority, or an empirical finding about machine selfhood. Literature descriptions are limited to the passages and abstracts consulted. Proposed comparisons have not been run for this paper. No proprietary implementation or private interaction is needed to reproduce its illustrative example or debate its argument.
