Machine Representation Without Machine Domination: AI Appeal Agents, Duties of Loyalty, and Human-Controlled Authority

Full report

On this page

1. Executive Summary

The transition from human-driven bureaucracy to algorithmic administration has precipitated an acute crisis in due process, administrative law, and legal standing. As institutions increasingly deploy artificial intelligence to adjudicate benefits, moderate digital platforms, and authorize healthcare at machine speed, individuals are systematically forced to interact with opaque, high-velocity decision systems. In response to this asymmetry, a secondary market of consumer-facing AI agents has emerged. These systems promise to act as tireless digital advocates capable of contesting automated denials, filing complex appeals, and navigating labyrinthine administrative architectures on behalf of unrepresented individuals. However, this phenomenon—termed the "agentic shift"—represents a movement from AI-as-tool to AI-as-agent, fundamentally disrupting the established legal frameworks that govern professional services, advocacy, and representation1. When an artificial intelligence acts on behalf of a human principal, the foundational legal architectures of agency law, fiduciary duty, and professional responsibility are violently tested. Unlike a human advocate, an AI possesses no innate conscience, no professional license to protect, and no organic capacity for loyalty. If an AI agent is provided by the very institution it is meant to contest, or if it operates on the same foundational infrastructure as the adjudicatory system, the human user faces a severe risk of structural disloyalty and systemic exploitation. This research comprehensively examines the legal, ethical, and human-computer interaction (HCI) paradigms necessary to govern AI appeal agents. By synthesizing the Restatement (Third) of Agency, unauthorized practice of law (UPL) statutes, and foundational procedural due process jurisprudence—specifically the balancing test established in Mathews v. Eldridge2—this report constructs a robust governance architecture. It proposes rigorous frameworks, including a Machine Representative Duty of Loyalty, a Human Authorization Ladder, and a No-Silent-Waiver Rule. These frameworks are designed to ensure that as legal and administrative representation accelerates to machine speed, the human being retains the temporal accommodation necessary for genuine comprehension, consent, and control over their legal standing.

2. Types of Machine Representation and Domain Applications

The deployment of artificial intelligence to assist individuals in adversarial or bureaucratic settings spans a vast array of legal and administrative domains. Understanding the variance in these deployments is critical, as the consequences of algorithmic failure, hallucination, or disloyalty scale directly with the severity of the human interest at stake.

Current and Emerging Uses of AI Agents

The legal and administrative ecosystem is currently witnessing the deployment of AI agents across ten critical domains. Each domain presents unique regulatory challenges regarding the unauthorized practice of law and the unauthorized waiver of human rights.

1. Administrative Appeals: Citizens increasingly utilize large language models to draft appeals against algorithmic welfare state determinations. This includes challenging the reduction of Supplemental Nutrition Assistance Program (SNAP) benefits or automated Medicaid disenrollments. The AI parses the state agency's notice, identifies procedural defects, and generates a formal request for a fair hearing.

2. Legal Information: Consumer-facing chatbots generate tailored explanations of jurisdiction-specific statutes, attempting to bridge the access-to-justice gap. These systems must navigate the precarious boundary between providing general legal information (protected speech) and dispensing individualized legal advice (the unauthorized practice of law)4.

3. Benefits Applications: AI agents autonomously compile financial records, extract relevant medical history from uploaded documents, and complete complex state and federal entitlement applications. These agents act as persistent navigators through multi-step bureaucratic processes that typically cause human abandonment.

4. Insurance Claims: In response to mass automated claims denials by private health insurers, patients deploy AI tools to instantly generate appeal letters. These agents cite specific medical necessity criteria, cross-reference the patient's electronic health record, and quote the insurer's own policy language to demand coverage overturns.

5. Financial Disputes: Algorithmic agents draft dispute letters to credit bureaus and initiate chargeback protocols with financial institutions to combat fraud or incorrect billing. These agents leverage statutes like the Fair Credit Reporting Act to enforce statutory timelines on institutions.

6. Healthcare Navigation: Specialized AI patient-advocates coordinate prior authorizations, schedule specialist visits, and dispute medical billing codes on behalf of incapacitated or overwhelmed patients. They act as relentless administrative proxies in a fragmented healthcare system.

7. Workplace Grievances: Employees utilize AI platforms to anonymously draft human resources complaints, analyze complex employment contracts, and document patterns of discrimination or wage theft, often referencing collective bargaining agreements or federal labor laws.

8. Education Appeals: Parents of children with disabilities employ AI advocates to parse the Individuals with Disabilities Education Act (IDEA) requirements. The agents draft formal demands for Individualized Education Program (IEP) accommodations and request due process hearings against school districts.

9. Platform Moderation Appeals: Digital creators rely on AI agents to automatically submit counter-notices against algorithmic copyright strikes and appeal automated account suspensions on social media platforms, navigating the strict timelines of the Digital Millennium Copyright Act (DMCA).

10. Accessibility Support: Cognitive prosthetic AI agents serve neurodivergent users by translating dense bureaucratic jargon into plain language, maintaining persistence in multi-step administrative processes, and executing tasks that require high executive function.

Distinguishing the Taxonomy of Representation

The regulatory exposure of an artificial intelligence system depends entirely on its functional classification. The law distinguishes between passive tools and active agents, assigning liability, professional responsibility, and fiduciary duties accordingly.

3. Agency, Loyalty, Confidentiality, and Consent

The application of twentieth-century professional responsibility and agency law to twenty-first-century autonomous agents exposes significant doctrinal gaps. As users delegate complex legal and administrative tasks to artificial intelligence, the legal frameworks governing human behavior must be radically adapted to govern algorithmic outputs.

Agency Law and the Fiduciary in the Machine

The law of agency governs the relationship where one person (the agent) acts on behalf of another (the principal) in a fiduciary capacity, altering the principal's legal relations with third parties11. Under the Restatement (Third) of Agency § 8.01, an agent owes a fundamental, non-negotiable duty of loyalty to the principal, requiring that all actions taken within the scope of the relationship be exclusively for the principal's benefit11. Furthermore, § 8.02 prohibits the agent from acquiring material benefits from third parties in connection with their representation11. However, the computer science definition of an "agent" diverges sharply from the legal definition. A software agent lacks intentions, a conscience, and independent assets to satisfy liability judgments11. The law cannot ascribe moral culpability to a matrix of weights and biases. Therefore, the legal responsibility for an AI agent must be ascribed to the human beings or corporations that design, deploy, and host the system12. When a principal uses an AI to enter into a contract or file an appeal, the principal cannot later object that they did not know what their algorithmic agent was doing; they are bound by the AI's objective manifestations of authority10. The "agentic shift" removes the human learned intermediary, creating a profound "supervision paradox"1. In traditional professional settings, a human expert actively supervises subordinates. If an AI initiates, decides, and executes an appeal autonomously, the professional is no longer the decision-maker, and the fiduciary anchor disappears1. To bridge this gap, Information Fiduciary theory argues that entities holding vast asymmetries of power, data, and technical capability over users must be bound by duties of care, loyalty, and confidentiality16. In the context of AI agents, this dictates that the agent's mathematical optimization functions must be rigidly aligned with the user's best interests, categorically avoiding self-dealing and third-party incentives1.

Professional Responsibility and the Unauthorized Practice of Law

When AI agents provide services historically reserved for lawyers, they invariably collide with UPL statutes. Statutes across all 50 states penalize UPL to protect the public from unqualified, incompetent legal advice4. The functional scope question is not whether the AI "understands" the law, but whether the system's output—as delivered to a specific person in their specific legal situation—constitutes the practice of law4. In Illinois, the Attorney Act (705 ILCS 205/1) strictly prohibits non-licensed entities from receiving compensation for legal services or holding themselves out as providing legal services13. The prohibition extends to corporations under the Corporation Practice of Law Prohibition Act (705 ILCS 220/1)19. The risk of AI UPL was most prominently highlighted in the litigation surrounding DoNotPay, an AI service marketed as the "world's first robot lawyer"20. While a federal class-action lawsuit against the company was dismissed on standing grounds (the plaintiff law firm failed to prove direct financial injury)15, the Federal Trade Commission (FTC) aggressively pursued the matter. The FTC ultimately reached a $193,000 settlement with DoNotPay, prohibiting the company from making unsubstantiated claims about its ability to substitute for professional legal services without rigorous evidence20. For licensed attorneys utilizing AI, the American Bar Association (ABA) Formal Opinion 512 is definitive: there is no "AI exception" to the rules of professional responsibility6. Lawyers must maintain competence (Model Rule 1.1) by understanding the limitations and hallucinatory risks of generative AI5. They must supervise their non-lawyer AI assistants (Model Rule 5.3), bearing absolute strict liability for any errors the AI produces5. Most critically, the duty of candor to the tribunal (Model Rule 3.3) requires lawyers to meticulously verify all AI-generated outputs to prevent the submission of fabricated citations to courts5.

Confidentiality, Privilege, and Informed Consent

The ingestion of sensitive health, financial, or legal data into a third-party large language model introduces severe, structural risks to confidentiality and data protection. Under ABA Model Rule 1.6 and equivalent state rules, a lawyer must not reveal information relating to the representation of a client5. Using consumer-grade generative AI models that retain user inputs to train future iterations effectively destroys attorney-client privilege and violates informed consent norms6. When unrepresented consumers use AI appeal agents directly, they face identical risks. The agent requires access to intimate medical records or financial histories to draft competent appeals. If the AI operates on a cloud architecture that ingests this data for model weight updating, the user's privacy is permanently compromised. Informed consent in the age of algorithmic representation requires far more than a boilerplate click-wrap agreement. Users must be explicitly educated on where their data flows, whether it is processed locally or in the cloud, who owns the underlying foundational model, and whether the deploying entity is monetizing the user's administrative trauma.

4. Conflicts and Vertical Integration

The most profound, systemic threat to the duty of loyalty in machine representation arises from provider conflicts and the vertical integration of algorithmic infrastructure. A fatal conflict of interest occurs when the corporate entity supplying the user's AI appeal agent possesses competing financial or operational obligations to the institution being appealed against. The modern digital economy favors massive consolidation. Consider a hypothetical scenario where a massive healthcare technology conglomerate operates the algorithmic prior-authorization system utilized by a major health maintenance organization (HMO). This system is designed to maximize insurer profitability by efficiently denying claims. Simultaneously, this same conglomerate develops and sells an "AI Patient Advocate" subscription directly to consumers, promising to appeal those exact prior-authorization denials. To complete the ecosystem, the conglomerate also provides the compliance auditing software used by state insurance regulators, and hosts the entire infrastructure on its proprietary cloud servers. In this vertically integrated ecosystem, structural disloyalty is not a risk; it is mathematically guaranteed. The AI appeal agent cannot aggressively litigate a denial without indicting the logic of its sister-system. The optimization weights of the consumer-facing agent will inevitably be tuned to prioritize the conglomerate's overall liability exposure, systemic efficiency, and server-compute costs over the individual patient's right to life-saving care. The agentic loyalty problem recognizes that the AI deployer is constantly in a position to influence the agent to the deployer's benefit over the user's benefit11. To resolve this catastrophic failure of due process, the law must mandate strict structural separation. An AI agent cannot owe a duty of loyalty to a principal if its developer derives material benefit from the opposing party. Fiduciary law handles human conflicts of interest through mandatory disclosure and recusal; a human lawyer cannot represent both the plaintiff and the defendant in the same litigation. In the algorithmic context, vertical integration of adjudicatory AI and advocacy AI must be governed by aggressive antitrust principles and strict, absolute conflict-of-interest prohibitions. If a vendor supplies the decision-making algorithm, it must be legally barred from supplying the consumer-facing appeal algorithm in the same administrative domain.

5. Machine-Speed Advocacy and Human-Speed Understanding

The primary commercial promise of AI agents is velocity: the ability to parse a 300-page dense insurance policy, cross-reference it against an algorithmic denial code, and generate a meticulously formatted, jurisdictionally accurate appeal within milliseconds. However, the American legal and administrative systems are fundamentally designed around human temporal rhythms. When machine-speed advocacy collides with human-speed understanding, the foundational requirements of due process are severely threatened.

**The Mathews v. Eldridge Balance**

The Supreme Court articulated the standard for procedural due process in the landmark 1976 case Mathews v. Eldridge2. The Court established a three-factor balancing test to determine what process is required before the government may impair a protected interest: first, the private interest that will be affected by the official action; second, the risk of an erroneous deprivation of such interest through the procedures used, and the probable value of additional safeguards; and third, the government's interest, including the administrative burdens that additional procedural requirements would entail2. Algorithmic administration radically tilts this balance. Institutions utilize AI to maximize administrative efficiency (the third factor), terminating benefits or restricting rights at a speed and scale that vastly increases the risk of erroneous deprivation (the second factor)2. If individuals deploy AI agents to fight back, the resulting "bot-on-bot" litigation occurs at a frequency entirely beyond human comprehension. While the AI agent can act quickly, the human principal must retain the temporal accommodation necessary to understand the agent's actions, review the evidence, approve settlements, or revoke authority. A legal system operating entirely at machine speed strips the human being of their fundamental agency.

Machine-Speed Automatic Stays

In traditional administrative law, the timely filing of an appeal often pauses the execution of an adverse action, preserving the status quo. For example, under federal Medicaid regulations (42 CFR § 431.230), if a state agency sends an advance notice of termination and the beneficiary requests a fair hearing before the action takes effect, the agency must maintain the beneficiary's services pending the hearing decision—a vital lifeline known as "aid paid pending"27. As institutions issue algorithmic denials instantaneously, human beings cannot reasonably be expected to meet traditional 10-day or 30-day appeal deadlines without the assistance of AI. However, a new procedural standard—the Machine-Speed Automatic Stay—must be enacted to bridge the temporal gap. If a denial is generated by an automated system, the user's AI agent must be empowered to instantly trigger an automatic stay of the adverse action. This instantaneous machine-speed filing halts the deprivation, but it must be coupled with a statutory "human temporal accommodation" window (e.g., 60 days). During this window, the human principal is afforded the time to review the machine's preliminary filing, consult human legal counsel, understand the basis of the denial, and authorize a substantive, formalized appeal.

6. Error, Hallucination, and Verification

The delegation of legal and administrative authority to generative AI systems carries severe, documented risks of factual fabrication and procedural default. Large language models function via probabilistic next-token prediction rather than deterministic fact-retrieval; they are designed to generate plausible text, not necessarily true text. This architecture results in the phenomenon of AI "hallucination"—where an LLM generates highly convincing but entirely fabricated information. The consequences of hallucination in adversarial settings are catastrophic. In the widely publicized case of Mata v. Avianca, attorneys were fined after submitting a ChatGPT-generated brief containing completely fabricated judicial opinions and citations6. Similarly, in Johnson v. Dunn, a federal court disqualified an entire law firm after confirming that the citations in their filing did not exist6. Because LLMs possess a dangerous capacity for false confidence, they present citations in perfect Bluebook format, deceiving both pro se litigants and unwary professionals. Courts have established a consistent standard: ignorance of the AI's internal logic or failure to verify its output is not a defense to professional misconduct or UPL6.

Limits on Machine Authority

Given these fundamental failure modes, AI agents must be strictly prohibited from executing irreversible legal actions without cryptographic, synchronous human verification. The delegation of absolute authority to a machine violates the principles of informed consent and due process. Specifically, an AI agent cannot be permitted to:

• Concede Facts: An AI might hallucinate a damaging admission based on misinterpreting a user's prompt or a poorly formatted uploaded document.
• Waive Rights: A machine cannot knowingly, intelligently, and voluntarily waive a constitutional or statutory right (such as the right to a hearing or the right to cross-examine witnesses).
• Accept Settlements: Financial finality requires human judgment regarding adequacy, fairness, and long-term consequences. An AI optimizing for speedy resolution may accept a fraction of a claim's true value.
• Abandon Claims: Missing a machine-speed deadline, miscategorizing evidence, or hallucinating a statute of limitations must not result in the permanent forfeiture of a human's legal claim.

To enforce these limits, all evidence parsed by an AI, and all citations generated in an appeal, must be subjected to a strict Evidence and Citation Verification requirement. The human user—or a licensed human supervisor—bears absolute liability for the outputs. Furthermore, auditability of the AI's decision-making process must be maintained without exposing privileged or intimate user data to third-party auditors. This requires the implementation of localized processing, zero-knowledge proofs, and the cryptographic logging of the agent's reasoning chains, allowing for post-hoc review of why the AI took a specific action without revealing the underlying sensitive data.

7. Ten Scenarios of Machine Representation

The following ten fictional scenarios illustrate the varied application of AI agents across diverse domains, demonstrating the requisite data access, authority parameters, human approval requirements, and conflict risks associated with each deployment.

Scenario 1: Benefits Denial (Medicaid Unwinding)

Scenario 2: Insurance Claim (Prior Authorization)

Scenario 3: Housing-Screening Dispute

Scenario 4: Employment Discipline

Scenario 5: School Appeal (IEP Accommodations)

Scenario 6: Platform Account Restriction

Scenario 7: Healthcare Authorization

Scenario 8: Immigration Filing Support

Scenario 9: Financial Fraud Dispute

Scenario 10: The Unresolved Failure (Vertical Integration)

8. Required Frameworks for Machine Representation

To mitigate the catastrophic risks outlined in the scenarios above, the legal and administrative systems must adopt new architectural constraints for the deployment of AI agents. Current tort law and product liability frameworks are insufficient to govern systems that act in a representational capacity1. The following nine frameworks must be integrated into administrative law and professional ethics codes.

1. Machine Representative Duty of Loyalty

A statutory or common-law requirement stating that any entity deploying a consumer-facing AI agent for legal, financial, or administrative advocacy assumes a strict, fiduciary-like duty to the user1. The agent's algorithms must be mathematically and operationally optimized solely for the user's benefit9. The deploying company is strictly liable for any self-dealing, third-party preference, or optimization for institutional efficiency embedded in the system's weights or prompts11. If the AI acts against the user's interests to save the deploying company compute costs, it is a breach of loyalty.

2. Conflict-of-Interest Matrix

A mandatory disclosure and restriction framework evaluating the developer's broader corporate ecosystem. Before an AI agent can be deployed in an administrative setting, the vendor must pass the following matrix:

3. Human Authorization Ladder

Legal representation requires the explicit escalation of consent based on the severity of the intended action. A blanket Terms of Service agreement is legally insufficient to authorize life-altering administrative actions.

4. No-Silent-Waiver Rule

An evidentiary and procedural rule stipulating that an AI agent cannot waive a human principal's statutory, constitutional, or procedural rights (e.g., the right to a physical hearing, the right to cross-examine) through omission, hallucination, or default. Any waiver of rights generated by a machine is per se void unless accompanied by a separate, specific, human-executed affirmative waiver. The law cannot recognize a machine's probabilistic output as a "knowing and intelligent" waiver of human due process.

5. Revocable Delegation Protocol

A technical standard requiring that all AI agents possess a highly visible, universally accessible "Kill Switch." Revocation of the agent's authority must be instantaneous, executable via the lowest-technology common denominator (e.g., an SMS text replying "STOP" or a single prominent button in a mobile app). Activating this protocol must instantly sever the agent's API access to the user's data and automatically notify the opposing institution that the machine's representation has been terminated.

6. Machine-Speed Automatic Stay

An administrative rule expanding upon the precedent of 42 CFR § 431.23027. If an institution uses an algorithmic system to deny a benefit or restrict a right, an automated challenge generated by a user's AI agent must immediately halt the execution of that denial. The status quo must be preserved. Institutions cannot be permitted to execute adverse actions at machine speed while demanding that citizens fight back at human speed.

7. Human Temporal Accommodation Standard

An administrative procedure standard requiring institutions to hold machine-speed proceedings open for a mandatory "human processing window." While the AI agent can file an appeal in milliseconds, the institution must grant the human principal a minimum statutory period (e.g., 30 to 60 days) to review the machine's filing, seek human legal counsel, amend the filing, or cure any AI-generated defects without penalty.

8. Evidence and Citation Verification Requirement

A professional responsibility and UPL mandate enforcing the strict standards of ABA Formal Opinion 5127. Any document submitted to a tribunal, agency, or institution by an AI agent must pass through a strict verification layer. For pro se users, the agent's user interface must force the human to physically click through hyperlinks to the actual, verified primary source text of any cited case or policy before the document can be transmitted, preventing the catastrophic submission of hallucinated law6.

9. Agent Handoff and Succession Protocol

A data-portability standard ensuring that if a user revokes an AI agent's authority—or if the AI identifies a case as too complex for safe automation—the agent must generate a standardized, human-readable continuity file. This file must contain the complete procedural history, the logic tree of the AI's actions, and the raw evidence. This ensures a licensed human attorney can seamlessly take over the representation without starting from scratch or losing statutory deadlines.

9. Counterarguments and Access-to-Justice Benefits

Critics of stringent AI governance argue that overly burdensome regulations—particularly those enforcing strict UPL statutes and demanding heavy human-in-the-loop verification—will stifle legal-tech innovation and price the poor completely out of the market. The access-to-justice benefits of low-cost, scalable machine representation are profound and undeniable. The vast majority of low-income individuals facing evictions, debt collection, or benefits denials have absolutely no access to human legal representation. The justice gap is a systemic failure. A well-designed, open-source AI agent can parse complex administrative codes, identify procedural defects in algorithmic denials, and generate highly competent legal prose that a pro se litigant could never produce on their own. Restricting these tools entirely to licensed attorneys, or crushing their developers under the weight of strict fiduciary liability, effectively traps marginalized populations in an unrepresented void, leaving them entirely defenseless against the government's and corporations' own algorithmic efficiency. However, the counterargument is that bridging the justice gap with unaccountable, hallucination-prone, vertically integrated AI creates a dystopian two-tiered system of justice: bespoke, fiduciary-bound human lawyers for the wealthy, and disloyal, error-prone bots for the poor. True access to justice is not achieved by providing vulnerable consumers with flawed, conflicted tools, but by ensuring those tools are bound by the exact same duties of loyalty and care that protect paying clients. Frameworks like the Human Authorization Ladder and the Machine Representative Duty of Loyalty do not ban AI agents; rather, they legitimize them by creating a safe, accountable, and legally sound environment for their deployment.

10. Open Questions

As the architecture of machine representation rapidly evolves, several critical legal and technical questions remain unresolved by current doctrine:

1. Liability Assignment in Open-Source Ecosystems: If an open-source, non-profit AI agent hallucinates a legal citation that results in a user's deportation or bankruptcy, who bears the ultimate liability? The unrepresented user, the non-profit deployer, or the corporate developer of the underlying foundational model?

2. First Amendment Boundaries vs. UPL: At what exact point does an AI's highly specific, personalized legal output cross from constitutionally protected free speech (legal information) into the heavily regulated, criminalized domain of unauthorized practice of law (legal advice)? How do courts apply the "reasonable reliance" standard to machines4?

3. Institutional API Blockades: Can a government agency or private corporation legally block consumer AI agents from accessing public portals via API to prevent server overload, or does that constitute an unconstitutional denial of the citizen's right to petition the government?

4. Cross-Jurisdictional UPL Enforcement: If an AI agent operates on cloud servers in California, advising a user located in Illinois on a federal administrative matter, which state's UPL statutes, legal ethics codes, and regulatory bodies govern the machine's behavior31?

The resolution of these questions will determine whether AI appeal agents become a democratizing force for human standing, or a mechanism for complete, opaque algorithmic subjugation.

Web-Ready Package

Public Summary (150 Words)

As government agencies, hospitals, and employers increasingly use algorithms to deny benefits, restrict accounts, and make life-altering decisions at machine speed, citizens are fighting back using AI "appeal agents." While these digital advocates promise low-cost, instant representation, they pose severe legal and ethical risks. Unlike human lawyers, AI agents lack a conscience, cannot be inherently loyal, and frequently hallucinate fake laws. Furthermore, if the company that built your AI advocate also provides the software to the institution denying you, the AI may secretly sabotage your case to protect its creator's profits. This report explores how to safely integrate AI into administrative law, proposing new legal frameworks like the "Machine Representative Duty of Loyalty" and "Machine-Speed Automatic Stays." By establishing strict rules for human authorization, conflict of interest, and error verification, we can use AI to bridge the justice gap without sacrificing human due process or institutional accountability.

Ten Key Findings

Ten Frequently Asked Questions

1\. What is an AI appeal agent? An AI appeal agent is a software tool, usually powered by a large language model, that autonomously reads denials, gathers evidence, and drafts or files appeals on behalf of a human user. 2\. Can an AI agent legally represent me in court? No. Under current statutes in all 50 states, only licensed human attorneys can practice law in court. AI agents can act as drafting assistants, but they cannot formally appear as your legal representative. 3\. What is the "Unauthorized Practice of Law" (UPL)? UPL occurs when an entity (including software) that is not a licensed attorney provides specific, individualized legal advice tailored to a person's specific facts, rather than providing general legal information. 4\. What happens if my AI agent hallucinates a fake law? If you submit a document containing fabricated cases, you (and your attorney, if you have one) can be severely sanctioned, fined, or have your case immediately dismissed by the court. 5\. What is the Machine Representative Duty of Loyalty? It is a proposed legal standard requiring the companies that deploy AI agents to mathematically and operationally align the AI to act solely in the best interest of the human user, forbidding self-dealing. 6\. How does vertical integration affect my AI agent? If the company that built your AI also provides software to your adversary (like your health insurer), your AI might be secretly programmed not to aggressively fight for you in order to protect the company's broader business interests. 7\. What is a Machine-Speed Automatic Stay? A proposed procedural rule where an automated appeal filed by your AI agent instantly pauses any adverse action against you, giving you the necessary time to review the case at human speed. 8\. Can my AI agent accept a settlement for me? Under the proposed Human Authorization Ladder, an AI agent can never accept a settlement or waive your rights without explicit, synchronous human approval (such as multi-factor authentication). 9\. How do I fire an AI agent? The Revocable Delegation protocol mandates that all AI agents have an instant "kill switch," allowing you to text "STOP" or click a single button to immediately revoke its access to your data. 10\. Will AI replace human lawyers? AI will increasingly handle routine administrative disputes and draft complex forms, but human lawyers will remain necessary for complex litigation, empathy, ethical accountability, and verifying the AI's work.

Glossary of Twenty-Five Terms

"Who does this AI agent actually serve?" Checklist

Before authorizing any AI agent to handle an administrative appeal or legal dispute, the human user must verify its true alignment:

• \[ \] Model Ownership: Is the developer financially independent from the institution you are appealing against?
• \[ \] Monetization: Do you pay for the service directly, or is the software company selling your appeal data to third parties?
• \[ \] Optimization Goals: Is the AI mathematically trained to maximize your financial recovery, or to minimize friction for the legal system?
• \[ \] Conflict of Interest: Does the developer provide auditing, hosting, or adjudicatory software to your adversary?
• \[ \] Data Architecture: Does the AI retain your medical/financial data to train future models, or is it processed locally?
• \[ \] Revocability: Is there a clearly accessible, low-tech button to instantly revoke the agent's authority and API access?

Six Warning Callouts

⚠️ WARNING: HALLUCINATION RISK: AI agents frequently invent fake case law, court rules, and medical codes. Submitting unverified AI documents to a tribunal can result in severe legal sanctions, fines, or permanent case dismissal.

⚠️ WARNING: UNAUTHORIZED PRACTICE OF LAW: Using an AI agent to represent you in formal court proceedings does not make it a licensed lawyer; doing so may expose you to UPL violations and leave you entirely unprotected by professional malpractice insurance.

⚠️ WARNING: PRIVACY WAIVER: Pasting confidential health, financial, or legal information into public or consumer-grade AI chatbots effectively destroys legal privilege and exposes your intimate data to third-party developers.

⚠️ WARNING: VERTICAL INTEGRATION: Never use an appeal agent built by the same corporate conglomerate that provided the software used to deny your claim. The structural conflict of interest guarantees disloyalty.

⚠️ WARNING: SILENT WAIVERS: Do not grant an AI agent authority to automatically accept settlements. Machines cannot comprehend the finality of waiving your rights to future legal recourse.

⚠️ WARNING: MACHINE SPEED DEADLINES: Just because your AI agent can file an appeal in milliseconds does not mean the institution will review it quickly. Always ensure you have a human-readable confirmation of the filing to protect your statutory deadlines.

Delegation-and-Approval Diagram

(Text Description of Diagram)

• Top Layer (The Human Principal): The user initiates the process. Arrows flow downward indicating "Grants Access" and "Sets Boundaries."
• Middle Layer (The AI Agent): Divided into three functional processing nodes: Data Ingestion (reads institutional portals and uploads), Analysis (identifies errors in the algorithmic denial), and Drafting (generates the legal appeal).
• Bottom Layer (The Institution): The adjudicating body receiving the appeal.
• The Firewall (Human Authorization Ladder): Between the Middle Layer and Bottom Layer sits a rigid gate labeled "Approval Matrix."
• Drafting & Research flows freely back to the Top Layer.
• Filing requires a "Click-to-Approve" human intervention.
• Waiving Rights/Settling requires a "Cryptographic Signature" (Biometrics/MFA) from the Top Layer.
• Side Loop (Revocation): A prominent red line arcs from the Top Layer directly to the Middle Layer labeled "Instant Kill Switch," severing all data connections and API authorizations instantly.

Page-Title, Meta-Description, and Search-Intent Suggestions

Works cited

1. The Fiduciary in the Machine \- ZwillGen, https://www.zwillgen.com/artificial-intelligence/the-fiduciary-in-the-machine/

2. Due Process Test in Mathews v. Eldridge | U.S. Constitution Annotated, https://www.law.cornell.edu/constitution-conan/amendment-14/due-process-test-in-mathews-v-eldridge

3. Mathews v. Eldridge | 424 U.S. 319 (1976) \- Justia Supreme Court, https://supreme.justia.com/cases/federal/us/424/319/

4. AI and the Unauthorized Practice of Law: Regulatory Boundaries in, https://ailegalauthority.com/ai-unauthorized-practice-of-law/

5. AI and Attorney Ethics Rules: 50-State Survey \- Justia, https://www.justia.com/trials-litigation/ai-and-attorney-ethics-rules-50-state-survey/

6. AI Legal Risks Posing Malpractice Threats to Illinois Lawyers, https://www.isbamutual.com/liability-minute/ai-legal-risks-posing-malpractice-threats-illinois-lawyers

7. Illinois AI Legal Ethics: A Practical Guide for Solo and Small Firm, https://clearbrief.com/blog/illinois-ai-legal-ethics-a-practical-guide-for-solo-and-small-firm-attorneys

8. How the ARDC tackles the Unauthorized Practice of Law, https://www.illinoiscourts.gov/News/847/How-the-ARDC-tackles-the-Unauthorized-Practice-of-Law/news-detail/

9. Fiduciary duties in AI | AI Governance Lexicon | VerifyWise, https://verifywise.ai/lexicon/fiduciary-duties-in-ai

10. AI Agent Legal Liability: Who Pays When AI Signs, https://blog.promise.legal/ai-agent-legal-liability-contracting-authority/

11. AI Agents and the Law \- arXiv, https://arxiv.org/html/2508.08544v1

12. The Law of AI is the Law of Risky Agents Without Intentions, https://lawreview.uchicago.edu/online-archive/law-ai-law-risky-agents-without-intentions

13. 705 ILCS 205/1, https://ilga.gov/documents/legislation/ilcs/documents/070502050K1.htm

14. COURTS (705 ILCS 205/) Attorney Act. \- Illinois General Assembly \- \-, https://ilga.gov/Legislation/ILCS/Articles?ActID=1853\&ChapterID=50

15. MILLERKING LLC v. DONOTPAY INC (2023) \- FindLaw Caselaw, https://caselaw.findlaw.com/court/us-dis-crt-s-d-ill/115536248.html

16. Information Fiduciaries in the Digital Age \- Balkinization, https://balkin.blogspot.com/2014/03/information-fiduciaries-in-digital-age.html

17. AI Fiduciaries and the Law \- ResearchGate, https://www.researchgate.net/publication/404170573\_AI\_Fiduciaries\_and\_the\_Law

18. Illinois Statutes Chapter 705\. Courts § 205/1 \- Codes \- FindLaw, https://codes.findlaw.com/il/chapter-705-courts/il-st-sect-705-205-1/

19. 220/1. Unlawful for corporations to practice law | Statutes \- Westlaw, https://content.next.westlaw.com/practical-law/document/NC18C0870DAFA11DA9F00E4F82CEBF25B/220-1-Unlawful-for-corporations-to-practice-law?viewType=FullText\&transitionType=Default\&contextData=(sc.Default)

20. DoNotPay has to pay $193k settlement over 'world's first robot lawyer', https://www.reddit.com/r/technology/comments/1fpfnu4/donotpay\_has\_to\_pay\_193k\_settlement\_over\_worlds/

21. 'Robot lawyer' DoNotPay reaches settlement in suit alleging it's, https://www.abajournal.com/news/article/robot-lawyer-donotpay-reaches-settlement-in-suit-alleging-it-is-neither-a-robot-nor-a-lawyer

22. AI 'Robot Lawyer' Not Licensed to Practice Law in Illinois, Class, https://www.classaction.org/news/ai-robot-lawyer-not-licensed-to-practice-law-in-illinois-class-action-claims

23. DoNotPay | Federal Trade Commission, https://www.ftc.gov/legal-library/browse/cases-proceedings/donotpay

24. Legal Ethics and Practical Considerations for Business Lawyers, https://www.americanbar.org/groups/business\_law/resources/business-law-today/2026-july/legal-ethics-practical-considerations-lawyers-using-ai-modern-legal-practice/

25. Exploring Generative AI Ethics in Illinois and the Broader Midwest, https://www.bakersterchi.com/exploring-generative-ai-ethics-in-illinois-and-the-broader-midwest

26. View of AI Agents and the Law, https://ojs.aaai.org/index.php/AIES/article/view/36705/38843

27. Must ABA Services Continue During a State Fair Hearing?, https://www.finnihealth.com/resources/clinicians/aba-services-maintained-pending-state-fair-hearing

28. 42 CFR 431.230 \-- Maintaining services. \- eCFR, https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-C/part-431/subpart-E/subject-group-ECFR8f26a68cc25862f/section-431.230

29. 42 CFR Part 431 Subpart E \- Procedures \- eCFR, https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-C/part-431/subpart-E/subject-group-ECFR8f26a68cc25862f

30. How to Appeal a Medicaid Denial: Fair Hearing Rights (2026) \- Brevy, https://brevy.com/medicaid/appeals-fair-hearings

31. Illinois Licensed Lawyer Practicing Illinois Law From a Geographic, https://www.isba.org/ethics/opinions/2203