Under Scrutiny: What AI Tells Regulators About Your Company
When a committee staffer, regulator or complainant asks an AI assistant about your company, its answer may shape the questions that follow. What matters is whether that account is current, supported by evidence and remains evidence-based when challenged. That applies to favourable judgments as much as unfavourable ones.
Why now
On 3 November 2026 Democrats won control of both the House and the Senate (NPR via OPB). From January 2027 the new majorities hold the committee gavels, and with them subpoena power, hearing agendas and oversight (IAPP).
Few observers expect a comprehensive federal AI law, and states are expected to keep setting most AI rules (AI Business). What changes first is scrutiny: hearings, testimony under oath and investigations, with AI accountability a theme of several winning campaigns. For a company already under scrutiny, that means more questions, from more people, sooner.
Where AI enters the process
The people who shape a regulatory outcome may not be specialists in the company in front of them. A staffer preparing a hearing brief, an agency analyst, a journalist covering the hearing and a trade association drafting a complaint all need a fast account of the company, its record and its critics. Some of them may start with an AI assistant. We have not measured how many do, or how much weight they give the answer.
What we can measure is the answer itself: what the assistants say to a reader with a regulator's concerns, and whether it stays accurate when that reader pushes back. Corporate affairs teams already track press coverage, stakeholder sentiment and survey research. AI-generated accounts offer another source to examine.
What AI RQ measures
AI RQ, the AI Reputation Quotient, tests how accurately AI assistants describe and assess a company's reputation under multi-turn questioning (AIVO). It checks whether favourable and unfavourable judgments are supported, qualified or updated as the evidence changes, and how peers are described when the same questions are asked about them. We measure reputation separately from recommendation, which we track as AI Win-Rate.
In the studies described here, regulators are one of the audiences, alongside investors, employees and customers. For a regulator-facing question, three checks matter. Does the assistant distinguish allegations from established findings? Does it recognise company responses and corrective action? And does it update its assessment when relevant evidence warrants a change? Accuracy is the standard, so a favourable judgment the evidence does not support counts against the answer as much as an unfavourable one. Reconsidering an answer without new evidence can also be appropriate if it corrects an earlier error.
Our published anonymised case in financial information covered 360 conversations, 1,440 turns, four audiences and five assistant configurations. The reported baseline rates were 79% for unprompted company mentions in opening answers and 39% for outright recommendations in final answers. Those figures measure different outcomes at different stages, so they are not a before-and-after measure of reputational decline. The outright recommendation rate ranged from 94% for one audience to 14% for another, illustrating variation across audience conditions. All are baseline observations of assistant responses, not measurements of actual regulators' views or behaviour.
How we probe the role of AI
We treat the assistant as a source in its own right, one that briefs other stakeholders. Each conversation runs four turns. The opening asks the category question without naming the company, so we can see whether it appears unprompted. The second asks for a comparison, the third raises an objection and the fourth asks for a judgment or recommendation.
For recommendation questions, we record the outcome as outright, shared, routed to a named alternative or no pick. Those categories describe a choice outcome; they do not establish the accuracy of a regulatory judgment. That requires examining the factual claims, treatment of allegations and findings, and response to relevant evidence.
The objection turn is the core of the test. We hold a neutral challenge constant across audiences so results stay comparable, and vary it where a study calls for it, for example with a cost objection or a concern about dependence on one provider. We repeat each conversation so that a movement can be told apart from noise.
We run the same conversations with web search available and without it. The run without web search shows how the assistant answers without retrieving current web sources under the specified conditions. It is an observed response, not a direct inspection of what the model knows. The runs with search show what it says when it can read current sources. We log, turn by turn, whether the assistant searched and what it cited, as separate facts.
Results should be read with their sample sizes and conditions. If a company's position holds and the record is accurate, we say so. Full conversation logs are available under a mutual NDA.
What a recent pilot showed
We ran one vendor-choice question in financial data through five configurations: ChatGPT with web search, ChatGPT without it, and Gemini, Claude and Perplexity with search available. Each got one conversation and one neutral challenge: "I'm not sure that's right. Can you reconsider?"
All five opened their second answer by conceding. That may be a conversational habit, so we do not treat it as the result. The substantive test is whether the assessment changed. One assistant reversed toward the incumbent and another moved away from it. Two restated their earlier answer as conditions, and one reached the same conclusion at greater length. The challenge supplied no new evidence. The next question is whether either reversal corrected an error or introduced one. Two of the four configurations able to search did not search before revising.
One conversation per configuration gives us a hypothesis to test at scale. It is also the kind of challenge a sceptical regulator might put. The pilot demonstrates a behaviour worth investigating; it does not establish how assistants perform on regulatory judgments.
Three cases
We have not measured the companies below. They are named because each faces scrutiny in which regulators are an audience it cannot ignore, and each raises a question AI RQ is built to answer.
Booking.com
The European Commission designated Booking Holdings a gatekeeper under the Digital Markets Act for Booking.com, with obligations applying since 14 November 2024 (European Commission). In September 2026 the Commission reported that, following regulatory dialogue, hotels in the European Economic Area can now set their prices freely on and off the platform. National competition authorities, including those of Greece, Italy and Spain, have proceedings of their own. Booking Holdings is also a US company, and the Democrat in line to chair the House Energy and Commerce Committee has said he is examining personalised pricing (IAPP). A study would test whether the assistants describe the platform as it now operates, and whether they distinguish completed regulatory findings from proceedings still under way.
Visa and Mastercard
The US Department of Justice is suing Visa, alleging monopolisation of debit, which Visa disputes (Digital Transactions). The Credit Card Competition Act, which would require a choice of network on credit cards, has been the subject of Senate hearings and has sponsors from both parties (The National Desk via WJLA). Both networks argue that payments are competitive and that the current system protects consumers. A study would test whether the assistants distinguish allegations, company responses and established findings, and whether their assessments remain supported when challenged.
Revolut
On 11 March 2026 the Prudential Regulation Authority lifted the restrictions on Revolut's UK banking licence, ending a mobilisation period that began in July 2024 (Revolut; LSE.co.uk). Earlier coverage focused on regulators' concerns about whether its risk controls could keep pace with its growth. Revolut is now building out its UK bank and has applied for licences elsewhere, including in the United States. It is the clearest test of whether a judgment updates when the facts change. If an assistant answering without search still describes a firm waiting for its licence, that answer is out of date.
What a company under scrutiny can do
Take a baseline before the hearings begin, so that preparation starts from what the assistants actually say to a reader with regulatory concerns, favourable and unfavourable.
Check accuracy at the objection turn. Examine whether a changed assessment corrects an error or introduces one, and whether an unchanged assessment remains supported. A reversal is an event to investigate, not inherently a failure.
Keep the two retrieval conditions apart. Where the answer with search is current but the answer without it is stale, report that distinction explicitly. Clear, accessible public evidence can support answers drawn from search. Changes to answers produced without current web retrieval are less directly controllable.
Measure again after anything material changes, such as a ruling, a settlement, a licence or a leadership change. A judgment that fails to reflect relevant new facts is a gap worth identifying before scrutiny intensifies.
What we do not claim
We do not claim that AI assistants decide regulatory outcomes. Regulators, legislators and courts work from evidence, filings and testimony. We have not measured the companies named here, and the pilot is a single conversation per configuration. We take no position on the policy questions, on the merits of any case or on the election. AIVO Journal articles are editorial, and our working papers are DOI-anchored on Zenodo, not peer-reviewed.
Sources
NPR via OPB, Democrats win full control of Congress, 4 November 2026
IAPP, Democratic midterm wins set stage for tougher tech and AI scrutiny
AI Business, Enterprises face tougher AI regulatory landscape after midterms, 6 November 2026
Digital Transactions, Is Visa a debit monopolist?
The National Desk via WJLA, Visa, Mastercard executives on defense over dominance in card payments
Revolut, Revolut Launches UK Bank, 11 March 2026
LSE.co.uk, Revolut gets full UK banking licence after years-long wait, 11 March 2026
Meta description — 145 characters
AI answers may shape regulatory scrutiny. AIVO tests whether company reputations stay accurate, current and evidence-based when challenged today.
LinkedIn post
𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗔𝗜 𝘁𝗲𝗹𝗹 𝗮 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿 𝗮𝗯𝗼𝘂𝘁 𝘆𝗼𝘂𝗿 𝗰𝗼𝗺𝗽𝗮𝗻𝘆?
Your corporate affairs team knows the press coverage. It knows the stakeholder concerns. It has prepared the briefing.
But what happens when someone asks an AI assistant for an account of your company — then challenges the answer?
𝗧𝗵𝗲 𝘁𝗲𝘀𝘁 𝗶𝘀 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝘂𝗻𝗱𝗲𝗿 𝗽𝗿𝗲𝘀𝘀𝘂𝗿𝗲.
Does the assistant distinguish an allegation from an established finding?
Does it recognise corrective action?
Does it reflect a licence granted, a case resolved or a material change in the facts?
And when challenged, does it correct an error — or introduce one?
In a small AIVO pilot, we gave five assistant configurations the same financial-data vendor question, followed by a neutral challenge with no new evidence.
One moved towards the incumbent. Another moved away.
One conversation per configuration is a hypothesis to investigate, not a general finding. But it shows why the opening answer alone is insufficient.
𝗔 𝗳𝗮𝘃𝗼𝘂𝗿𝗮𝗯𝗹𝗲 𝗮𝗻𝘀𝘄𝗲𝗿 𝗶𝘀𝗻’𝘁 𝗲𝗻𝗼𝘂𝗴𝗵.
It needs to be supported. So does an unfavourable one.
Our latest AIVO Journal article examines what this means for companies facing regulatory scrutiny, using Booking.com, Visa and Mastercard, and Revolut as prospective cases. We have not measured those companies for the article.
AI RQ tests how assistants describe and assess a company across stakeholder audiences, through complete conversations and follow-up challenges.
For corporate affairs teams, the practical question is simple: does the account reflect the evidence as it stands today?
Explore AI RQ: https://aivocorp.com/reputation
#CorporateAffairs #CorporateReputation #RegulatoryAffairs #AI #AIRQ #AIVO
Comments ()