The Performance of Judgement

Essay 5 of 5, Decision-Native 2040 series. Draft 1, 2026-04-23.

One of the first things an auditor learns is how to confirm a bank balance. The procedure is simple and it has a purpose. When a company tells you it has money in an account somewhere, you do not ask the company. You ask the bank. You write directly to the bank, addressed to someone with authority to answer, and you ask them to confirm the balance and the account holder’s authority to operate it. The company gives you an address. You check that the address belongs to the bank. You do not use the company’s postbox.

Between 2016 and 2018, EY Germany signed off on Wirecard’s audited accounts three times without making that confirmation directly with OCBC, the Singapore bank where Wirecard said it held approximately one billion euros.¹ EY instead accepted confirmations supplied through a trustee, Citadelle, and documents supplied by Wirecard itself. OCBC later stated that neither Wirecard nor Citadelle had ever had an account at the bank.

Day-one audit-school behaviour, three years in a row, omitted in exchange for a growing procedural file that documented every other step.

This is an essay about what happens to judgement when it gets absorbed into procedure. The cases below are early instances of a pattern that, by 2040, will define how most professional work is produced and assessed. It is not an essay about fraud. Fraud is actors gaming a system they do not believe in. What happened at EY is different. It is what happens to practitioners operating a system that has silently replaced their judgement with its representation, while the practitioners are still operating it in good faith.

The accumulation

Audit procedures were not always a regime. The procedures were designed, over decades, to force auditors to form independent views of the companies they audited. Confirm cash balances with banks. Observe inventory. Inspect contracts. Interview counterparties. Each procedure’s original function was to produce an encounter with reality that would force the auditor to notice when something did not match.

Post-Enron, post-Sarbanes-Oxley, the PCAOB set out to make audit quality measurable. The Public Company Accounting Oversight Board began inspecting Big Four engagements. Internal quality reviews multiplied. Each new failure mode produced a new required procedure. Documentation requirements grew. Over time, the time spent documenting that the procedures had been performed came to exceed the time spent performing them.

The metric that measured audit quality became did the firm complete and document every required procedure. Anyone optimising for that metric is no longer optimising for did I understand this company. The two were correlated for a while. Then they decorrelated, and nobody inside the system noticed.

The inner witness

EY had a forensic unit. In 2016 and 2017, that unit opened an investigation into Wirecard, internally nicknamed Project Ring, after a whistleblower flagged accounting manipulation and attempted bribery of an auditor in India.

The forensic unit produced a sixty-three-page deck documenting “red-flag indicators” of possible balance-sheet manipulation. The deck noted that the work had so far used only open-source background research and had not yet done interviews or email review. The investigation was not complete. The red flags were not resolved.

Three weeks after the forensic deck was produced, the audit team concluded that Project Ring was “completed,” with no evidence of flawed accounting found.²

Two functions of the same firm looked at the same evidence. The forensic function saw red flags requiring follow-up. The audit function saw a procedural question that had been resolved. Both functions were staffed by people acting in good faith, inside the same institution. The procedural conclusion won because the procedure asked a different question. The forensic deck asked is something wrong here? The audit procedure asked has the required follow-up been documented? The two questions produce different answers on the same facts.

The shopping test

In 2018, EY ran a merchant-verification test on Wirecard’s customer merchants, the businesses whose card-processing revenue flowed through the supposed third-party acquiring partners. The procedure involved buying something from those merchants using prepaid cards, to confirm they were real.

The prepaid cards were supplied by Wirecard.³

The test produced a documented result. The procedure was complete. The independence, which was the reason the procedure existed in the first place, was gone, but the documentation did not reflect that. Verification theatre at the finest grain: a test whose independent variable was supplied by the subject being tested.

What the journalist did

While EY was documenting complete procedures, Dan McCrum and Stefania Palma at the Financial Times were doing something simpler. McCrum and Palma were forming independent views of Wirecard using the methods EY was no longer using.

Palma, in particular, visited Wirecard’s third-party acquiring partners in the Philippines in person. She walked to the offices. She looked. She found the addresses belonged to bus companies, retired sea-captains’ homes, and empty rooms.⁴

None of this was hidden. The addresses had been in the audit file for years. The bus companies had been running at those addresses. EY had local staff in the same country, on the same client.

She was not doing forensic accounting. She was doing the thing the audit procedure was originally designed to enforce, producing an encounter with reality. Everyone at EY had access to the same addresses. No one walked there.

The German Federal Financial Supervisory Authority responded to the FT’s reporting by banning the short-selling of Wirecard stock in February 2019 and filing a criminal complaint against the Financial Times itself. The system, in multiple dimensions, had substituted the defence of its own process for the activity the process was originally supposed to support.

The Abschlussprüferaufsichtsstelle, Germany’s audit regulator, eventually ruled EY’s Wirecard work “grossly negligent” and imposed a two-year ban on EY Germany taking new public-interest audit clients.⁵ The Bundestag inquiry commission’s final report, 4,500 pages, concluded that EY failed to identify fraud-risk indicators, did not fully implement professional guidelines, relied on verbal assurances, and did not request independent bank confirmations.⁶

EY’s public defence has been that the fraud was elaborate and that many parties were deceived. The regulators and the commission have rejected that framing. The specific failure, they found, was procedural rather than unlucky. The procedure EY had retired was the one requiring direct bank confirmations.

The same failure mode at scale

The EY case looks exceptional because it collapsed a DAX 30 company. The failure mode itself is ordinary. It is running, right now, across at least one entire profession, not through fraud but through design.

In 2017, Wang and colleagues at UCSF analysed 23,630 inpatient progress notes from 460 clinicians at a single academic medical centre over eight months. In a typical note, clinicians manually wrote 18% of the text. 46% came from the prior note, copied forward. 36% came from templates. For residents, the manually entered share dropped to 11.8%.⁷

The progress note was designed, originally, to force the clinician to re-evaluate the patient daily. You examine the patient. You form an impression. You write it down. Each day’s note was supposed to be an encounter with a changed clinical reality.

It is now a copy of yesterday’s note, with template elements imported where required, and 18% new content. The note is not an encounter. It is a re-certification of the procedure of having written a note.

Sinsky and colleagues measured what happened to clinician time as EMR documentation expanded. For every hour of direct face-to-face patient contact, a physician spends nearly two hours on the EMR and desk work. Inside the exam room itself, more than a third of exam-room time goes to the screen. Then, nightly, one to two additional hours of after-hours documentation.⁸ Arndt and colleagues, measuring family physicians’ event logs, found 5.9 hours per day inside the EMR, 44% of it clerical rather than clinical.⁹

The 2021 CPT Evaluation and Management reform aimed to reduce this burden by removing history and physical exam from code-selection requirements. Note length grew by 8%.¹⁰ The procedure had become load-bearing for reasons that had nothing to do with why it was built. Shrinking one face of it did not shrink the face that mattered.

The recursion

In the last two years, ambient AI scribes have started writing the notes for clinicians. The scribes listen to the visit, generate a compliant note, and populate the EMR automatically. They are marketed as a solution to clinician burnout.

The early outcome data tells a different story. Riverside Health reports an eleven percent rise in work RVUs and fourteen percent more HCC diagnoses per encounter after deploying AI scribes. Texas Oncology’s documented diagnoses per visit rose from 3.0 to 4.1. Cigna began auto-downcoding level 4 and 5 E&M claims in October 2025 in response to this upward drift across the insured population.¹¹

The AI is doing exactly what it was trained to do. It is writing the most procedurally complete, highest-documented, highest-coded note that the billing system will accept. The substance of clinical reasoning is not what it was trained on. The clinicians are not making different decisions, and the patients are not getting different care. The documentation, which is the thing the billing system measures, has inflated to the ceiling of what the system will pay for.

The loop closes. The AI produces the notes that the coding rules demand, rules originally designed to measure something the notes no longer contain.

What performance is

When an organisation measures its own judgement, the metric starts to reshape the practice. The practice becomes performance of the measurement. The substance, the thing the measurement once tracked, evaporates quietly, in good faith, at industrial scale.

The auditors at EY were not cheating. They were doing what the profession now asks of them, which is to complete the procedures and document completion. The profession drifted, over thirty years, from a world where the procedures served independent judgement to a world where the documentation of the procedures was the judgement. The drift did not happen because anyone chose it. It happened because the metric that measured audit quality rewarded the documentation, and every incremental decision about how to spend audit time slowly flowed toward the documented version.

The clinicians at UCSF are not cheating. They are doing what the EMR now asks of them, which is to produce a documented note that clears coding and billing checks. The profession has drifted, over fifteen years, from a world where the note captured the clinician’s reasoning to a world where the note is produced to satisfy the system. That drift did not happen because anyone chose it either.

The AI scribe adds a darker variation. The system does not merely absorb the judgement that used to happen; it performs the representation of that judgement better than the humans ever did. The note reads more thoroughly. The codes are more complete. Whether any reasoning stood behind the note is a question the billing system was never going to ask.

The test that cannot be run

There is one more thing that makes this pattern dangerous. An outside observer, whether the regulator, the insurer, the board, the patient, or the shareholder, has no reliable way to distinguish a firm that is practising from a firm that is performing. From the outside, the outputs are identical. The audit reports have the same shape. The notes have the same fields. The quarterly pack has the same headings.

The difference, at the individual level, is a cognitive act that leaves no trace. The auditor who actually reconstructed the trust account from scratch produces the same signed report as the auditor who ticked the procedure box. The clinician who actually re-examined the patient produces the same Epic note as the clinician who copy-forwarded yesterday’s.

This is why the atrophy of an organisation whose procedure has absorbed its judgement cannot be caught by auditing its output. The output is built specifically to look the same regardless of whether the judgement is still inside.

The only reliable test is a tail event. For EY, it was Wirecard. For the medical system, the tail event is the one it has been running for a decade without noticing. The judgement of clinicians who trained before the current regime is still partially intact. The judgement of clinicians trained inside it has no baseline to compare to. The experiment is live. The readout lands around 2040. Nobody is going to stop it in the middle because nobody inside the system can see it.

What would not collapse

The pattern is not inevitable. Procedure absorbs judgement when the metric scores procedural completeness. It does not when the metric scores calibration against outcome. The audit profession measured itself on documentation. The EMR measures itself on coding. Both reward producing the artefact and ignore whether the artefact tracked anything real.

A different architecture would record the reasoning behind decisions, hold it against eventual outcomes, and score the alignment between the two. The substance of the decision becomes part of the record. The record updates against reality. That is the design the rest of this series has been pointing at, and the difference between a 2040 in which judgement walks out of every profession and one in which it gets compounded across them.

The series began with a question about what it takes to still be deciding in a world where most of the deciding has been automated for you. Five essays later, the question is narrower and harder. Not whether the systems are good; they are. Not whether you should resist them; you should not, mostly. What remains is whether there is someone inside your organisation who is still walking to the offices. Who is still asking whether the confirmation came from the bank. Who is still noticing what they do not understand, and choosing to investigate before closing the file. There is no external test for whether that person is there. There may be no internal test either.

Sources

1. Joe Brennan, “EY failed to ask for Wirecard bank statements for three years,” Irish Times, 26 June 2020. EY’s acceptance of trustee and company-supplied confirmations instead of direct OCBC confirmation, 2016–2018.

2. Joe Brennan, “EY fraud unit found ‘red-flag indicators’ at Wirecard in 2018,” Irish Times, December 2020. Project Ring forensic deck and the audit team’s procedural closure.

3. Joe Brennan, “Wirecard and EY Germany: the anatomy of a flawed audit,” Irish Times, 2021. Merchant-verification shopping tests using prepaid cards supplied by Wirecard.

4. Committee to Protect Journalists, Q&A with Dan McCrum and Stefania Palma, October 2020. Palma’s physical visits to Wirecard’s third-party acquirer addresses in the Philippines.

5. Joe Brennan, “German watchdog finds EY’s Wirecard audits ‘grossly negligent’,” Irish Times, April 2024. APAS €500,000 fine and two-year ban on EY Germany for new public-interest audits.

6. Joe Brennan, “EY audit failings on Wirecard laid bare in ‘dynamite’ report,” Irish Times, 2021. Bundestag inquiry commission addendum on the 2018 EY audit.

7. Michael D. Wang et al., “Characterizing the Source of Text in Electronic Health Record Progress Notes,” JAMA Internal Medicine 177(8): 1212–1213 (2017). 23,630 UCSF inpatient progress notes, 460 clinicians, eight months. 18% manual, 46% copied forward, 36% templated.

8. Christine Sinsky et al., “Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties,” Annals of Internal Medicine 165(11): 753–760 (2016).

9. Brian G. Arndt et al., “Tethered to the EHR: Primary Care Physician Workload Assessment Using EHR Event Log Data and Time-Motion Observations,” Annals of Family Medicine 15(5): 419–426 (2017).

10. Epic Research, “Two years after coding changes sought to decrease documentation, notes remain bloated,” 2023.

11. Jingjing Dai, Joseph Kvedar, and Daniel Polsky, “Ambient AI scribes and the coding arms race,” npj Digital Medicine, December 2025.