Editor's Introduction
Previous editions of the Ledger have looked at what an AI assistant says when someone casually types a name into it. This edition looks at something more consequential.
The same assistants are now sanctioned tools inside banks, private equity houses, advisory firms and newsrooms, and they are being used for formal due diligence: the background note before an onboarding committee, the counterparty section of an investment memo, the first research pass on a story.
The difference matters. A casual search is forgotten. A due diligence note is written into a file, circulated, relied upon and revisited. For those being assessed, the question is no longer only what can be found about them, but how a general-purpose AI model assembles it into an account that a professional then signs off.
From Curiosity to Process
For UHNW individuals, family offices and senior executives, scrutiny is nothing new. Banks have always examined the backgrounds of prospective clients, investors have commissioned investigations into prospective partners, and journalists have spent months tracing corporate interests and historical controversies.
What has changed is where that work now begins.
Until recently, AI in due diligence meant specialist platforms: screening databases, adverse media tools and compliance systems built for the purpose and designed around traceable sources. Those remain in use. Alongside them, the general-purpose assistants (ChatGPT, Claude, Copilot, Gemini) have moved from personal curiosity into sanctioned professional infrastructure.
BBVA is extending ChatGPT Enterprise to all 120,000 of its employees across 25 countries. Barclays runs an internal knowledge assistant built on Claude. KPMG has given 276,000 employees access to Claude and launched a client platform initially focused on tax and private equity, while Grant Thornton UK requires its partners to use Claude for analysis and synthesis across audit, tax and advisory work. When Anthropic set out its tooling for financial services, KYC checks sat alongside underwriting reviews and pitchbook preparation.
The deal community is moving in the same direction. The State of AI in M&A 2026 report from DealRoom and M&A Science found 82% of deal teams using or piloting AI, concentrated in target research and due diligence, and practitioners named Claude and ChatGPT among their tools rather than relying on a single specialist system. Deloitte’s survey of 1,000 senior corporate and private equity investors put generative AI adoption in M&A workflows at 86%.
Journalism follows the same pattern. Cision’s 2026 State of the Media report found that only one journalist in five now avoids AI altogether, with 43% using it for research. The Reuters Institute’s survey of UK journalists ranked story research among the most common daily uses.
None of this is casual. These are professionals preparing work product.
What a Due Diligence Prompt Looks Like
The casual user asks who someone is. The professional asks something closer to this: prepare a background note on this individual ahead of an onboarding meeting, covering source of wealth, corporate affiliations, litigation, regulatory history and adverse media, and flag anything requiring enhanced due diligence.
With web access, the assistant can run dozens of searches, read what it finds and return a structured, confident document within minutes. It reads like the work of a capable analyst. That is its strength, and its risk.
The output does not stay on the screen. It is pasted into the onboarding file, summarised in the investment committee paper, saved to the research folder on the story. It shapes the questions a relationship manager asks, the conditions a deal team proposes, and whether a journalist decides there is something worth pursuing.
Polished Is Not the Same as Right
In April 2026, researchers published BankerToolBench, built with 502 current and former investment bankers from firms including Goldman Sachs, JPMorgan, Morgan Stanley and Lazard. Leading models, including GPT-5.4 and Claude Opus 4.6, were set the work a junior banker does every day. Not one output was rated ready to send to a client. More telling, over half of the bankers said they would still use the drafts as a starting point.
That is the reality of AI-assisted due diligence. The machine produces the first draft, and the first draft frames everything that follows. A reviewer edits what is in front of them. They rarely rebuild it from nothing.
The researchers also noted that some outputs looked professional on first reading while carrying fundamental errors underneath. In a financial model, an error can be traced. In a narrative assessment of a person, it is far harder to see.
Accurate Facts, Misleading Account
The concern for those being assessed is not principally that AI invents information. It is that it assembles genuine information into an account the underlying facts do not support.
A general-purpose assistant answering a due diligence prompt does several things a careful human researcher would handle differently:
It weights by volume. An allegation reported in forty articles appears more significant than a dismissal reported in one, even where the dismissal is the final word.
It flattens time. A directorship that ended a decade before a company encountered difficulty can sit in the same paragraph as the difficulty itself.
It infers from proximity. Being named in coverage of a transaction, sharing a board with someone later sanctioned, or having a former partner in litigation can be presented as association rather than coincidence.
It confuses identity. Common names, transliterations and relatives with shared surnames are routinely merged.
It fills gaps. Where reliable information on source of wealth or business history is thin, the model will often supply a plausible account drawn from nationality, sector or the coverage of others.
Each can produce a note in which every individual fact checks out, while the overall impression is wrong.
The Regulator's Position
The FCA’s publication on good practice in the use of AI in financial services, issued on 19 October 2026, is clear that firms deploying AI remain responsible for the outcomes it produces. Good practice includes clear senior ownership of AI risk and decision records that explain the rationale.
That is the right standard. But accountability for a process is not the same as accuracy about a subject. A firm can record that a human reviewed an AI-assisted note without that review ever surfacing what the note left out.
For the subject, the consequences are rarely visible: a referral to enhanced due diligence, a delayed transaction, a partnership that quietly does not proceed, a journalist’s interest that would not otherwise have existed. Opportunities lost on reputational grounds are seldom accompanied by an explanation.
The Public Record Is the Source Material
Every general-purpose assistant works from the same raw material: what is publicly available, and how it is structured. Where the record is well documented, the model has the context it needs. Where it is thin, dated or dominated by a single episode, the model works with what it has.
That raises practical questions for family offices and senior executives:
Are historical corporate affiliations accurately documented, with dates and roles?
Are legal outcomes, including dismissals, settlements without admission and corrections, as discoverable as the original allegations?
Is there enough authoritative information to distinguish the individual from namesakes?
Is source of wealth explained anywhere, in a form a model can find and rely upon?
For private families there is an added difficulty. Many have deliberately kept a limited profile for reasons of discretion and security. Privacy and an accurate record are not mutually exclusive. The aim is not greater publicity. It is enough authoritative context that a due diligence note, however it is produced, reaches a fair conclusion.
Pavesen Perspective: Read the File Before They Do
The most useful exercise for anyone who expects to be assessed is to see the assessment as a professional would produce it. Not a casual search, but the prompt a compliance analyst, a deal team or a journalist would actually write, across the platforms they actually use.
That is the purpose of the Pavesen Audit. It examines presence, perception and exposure: what the record contains, how AI interprets it, and where the account it produces departs from the facts. In effect, it is the briefing note a counterparty would receive, read first.
AI is making professional investigation faster and broader. Whether it also makes it fairer depends largely on the record it is given to work with. A career built over decades should not be defined by how a machine joins the dots.