## Introduction The leaders in AI governance, including Credo AI, Holistic AI, and the rest of the Fast Company "Most Innovative" list, sell the same product. A platform where the company being audited configures the audit. Risk thresholds tunable by the customer. Remediation automated by the same vendor's AI agents. Compliance reports generated by an LLM the vendor controls. The product literally markets "customizable controls" and "AI-powered remediation" as features. This is not a niche pattern. It is the category. ## We have done this before In the 1990s, Arthur Andersen audited Enron's books while also selling Enron extensive consulting services. The audit firm's revenue from a single client made it structurally unwilling to challenge that client's accounting. The result is well known: Enron collapsed, Arthur Andersen dissolved, and the retirement savings of tens of thousands of employees evaporated overnight. Less well remembered: there was a moment before the collapse when this looked like a normal, well-regarded business arrangement. Andersen was the most prestigious audit firm in the world. Enron was Fortune's "Most Innovative Company" six years running. The lesson Congress took from Enron, codified in the Sarbanes-Oxley Act of 2002, was that audit and consulting cannot be sold by the same firm to the same client. The Public Company Accounting Oversight Board (PCAOB) was created to oversee auditors directly, separately from the companies being audited. Auditor independence became enforceable in two specific senses: independence in fact (no financial interest in the audited company) and independence in appearance (no relationship a reasonable outside observer would find compromising). These rules exist because we learned, expensively, that when the people checking the work are paid by the people doing the work, the work does not get checked. ## AI governance today The leading AI governance vendors do all of the following at once: - Audit AI systems on behalf of the customer deploying them - Allow that same customer to configure the audit thresholds - Automate remediation of any issues their own audit surfaces - Sell consulting and "AI Audit" advisory services to the same customer - In some cases, employ the same people who write or advise on the regulations they help customers comply with There is no independent body equivalent to the PCAOB. There is no separation between auditor and remediator. There is no requirement that AI governance vendors disclose their conflicts of interest. There is no fact-versus-appearance distinction. Fortune 500 buyers of these platforms understand exactly what they are buying: the appearance of compliance, configured in their favor. When the EU AI Act's enforcement obligations under Articles 12, 19, and 26 land in 2026, every one of these arrangements will produce technically valid documentation. Much of it will be evidence only of the buyer's compliance posture, not of the AI system's actual behavior. ## What independence would look like Auditor independence in AI governance, borrowing the same principles accountants codified after Enron, would mean four things: 1. Audit thresholds set by published external standards (the EEOC four-fifths rule under 29 CFR 1607 for disparate impact, FDIC SR 11-7 and OCC 2011-12 for model drift), not by the customer. 2. A structural separation between the entity that detects problems and the entity that remediates them. The auditor cannot pre-resolve findings before a regulator or affected party sees them. 3. Records the customer cannot tamper with after the fact, even through the customer's own administrative access. The integrity has to be cryptographic, not policy-based. 4. A public, prior commitment naming which findings the auditor will surface regardless of customer preference, and what kinds of customers or features the auditor will refuse. These are not radical principles. They are the same principles that govern every field where the public has decided that self-certification is insufficient. Financial audit. Drug safety. Aircraft certification. Building inspection. We have a hundred years of institutional memory about what happens when these principles are absent. ## The unequal stakes Enron's collapse hurt shareholders, employees, and pension funds. They organized, they lobbied, they sued. Congress responded. The people on the receiving end of biased AI decisions today do not have pension funds. They are job applicants screened out before a human ever sees their resume. Patients triaged by models nobody has independently checked. Defendants risk-scored by software no court has audited. Students whose financial aid algorithms flagged them as a risk. They have no organized constituency, no lobbying arm, no standing to sue an AI vendor's customer's vendor. This asymmetry has a corollary. Enron eventually collapsed, and Arthur Andersen with it, because there were organized stakeholders who could not be ignored once the truth surfaced. The collapse, devastating as it was, is what made Sarbanes-Oxley politically possible. AI governance has no equivalent forcing function. Credo AI and Holistic AI are not concealing fraud from shareholders. Their customers are getting exactly what they pay for: defensible documentation, configured in their favor. The transactions are honest. Everyone in the room is satisfied. The people not in the room have no way to make themselves heard. ## There will be no visible failure The natural arc of the Enron story is collapse, scandal, reform. That arc requires a forcing function. A moment when the harm becomes undeniable, the harmed parties organize, and the political system responds. The AI governance category has no such arc available to it. Credo AI, Holistic AI, and their peers will not collapse. Their customers are not their victims. Their celebrated status will continue. The Fortune 500 buyers will continue to deploy AI systems that harm people, with documented compliance to point to when questioned. The harm will continue to be paid by the people those AI systems decide against. One denied loan at a time. One screened-out resume at a time. One missed diagnosis at a time. None of those harms will aggregate into a single visible failure. Each will look like an individual unfortunate outcome, indistinguishable from the noise of any other decision. The textbook the accounting profession published after Enron is on the shelf. The AI governance industry has chosen not to read it. Without an Enron-scale forcing function to compel them to read it later, they will continue to operate exactly as they do now, indefinitely, against a population that has no standing to demand otherwise. ## What this asks of the rest of us If the category will not self-correct, correction has to come from somewhere else. **Regulators** at the FTC, EEOC, CFPB, FDIC, OCC, the European Commission AI Office, and national AI authorities across the EU: AI governance vendors that allow the audited party to configure the audit thresholds, automate their own remediation, or sell consulting services to the same client they audit should not satisfy any obligation under EU AI Act Articles 12, 19, or 26, or any future US enforcement framework. Independence in fact and independence in appearance are well-trodden legal concepts. They translate cleanly to AI auditing. Require them. **Buyers** on Fortune 500 compliance teams: when your AI governance vendor offers configurable risk thresholds as a feature, that is the problem the EU AI Act is meant to surface, not a solution to it. Ask your vendor, in writing, what findings they have committed to surfacing regardless of customer preference. If the answer is "whatever you configure," you are buying audit theater, and your name will be on that decision when it surfaces. **Engineers** building governance products: you can refuse to ship the suppression features. You can refuse to wire customer-configurable thresholds below standards-aligned defaults. You can refuse to build the self-remediation loop. The reference is on the shelf. So is the cost of pretending it isn't. **Founders** in this category: incorporate as a Public Benefit Corporation. Publish a Charter that names which customers and which features you will refuse. Require unanimous board approval to amend it. Take the refusal off the table before money can convince you to put it back on. ## A standing offer I founded AILedger to help. That offer is open. To any regulator drafting AI auditing rules: I will brief you, your staff, or your standards body on what enforceable AI auditor independence looks like at the schema and protocol level. No fee, no NDA required, no pitch attached. Just the substance. To any buyer's compliance team trying to write procurement criteria that distinguish audit substrate from audit theater: I will help you draft the language, free of charge, whether or not you ever buy from AILedger. To any founder building in this category who wants to incorporate as a PBC with structural anti-theater commitments: the AILedger Charter is public at [ailedger.dev/charter](https://ailedger.dev/charter). Adapt it. Improve on it. Make us look bad by going further. To any engineer who has been asked to build the suppression features and wants out: the open-source Detection layer is at [github.com/ailedger-dev/ailedger-detection](https://github.com/ailedger-dev/ailedger-detection). The detection primitives are Apache 2.0, anchored to published standards, and do not have configurable thresholds for that reason. You can use them. You can copy them. You can decide whether this is the kind of work you want to be doing instead. For any of the above, reach me at [help@ailedger.dev](mailto:help@ailedger.dev). --- This is not ok. We can do better than this. The danger is not that the AI governance category will collapse like Arthur Andersen did. The danger is that it will succeed, and the people it claims to serve will be harmed anyway. If you read this and felt something at the end, that feeling is correct. Act on it. If you want to help build it, I am hiring. Early equity. Serious inquiries only: [jobs@ailedger.dev](mailto:jobs@ailedger.dev). --- *Jake Joyner is the founder of AILedger, PBC. Contact: [help@ailedger.dev](mailto:help@ailedger.dev). The Charter referenced above is published at [ailedger.dev/charter](https://ailedger.dev/charter). The open-source Detection layer is at [github.com/ailedger-dev/ailedger-detection](https://github.com/ailedger-dev/ailedger-detection).* *This essay is opinion and commentary on a matter of public concern. Factual claims about named companies are sourced to those companies' own public marketing and press releases, documented separately in [a sourced vendor assessment](vendor-assessment/). Good-faith corrections of any factual error are welcomed at [help@ailedger.dev](mailto:help@ailedger.dev).*