Six ways judgment wears down

The failure modes from Judgment Capital, each with an owner, a reading for the board and a working method from Appendix A.

Score your organisation

Order the book

Sample organisation · 640 people · Tabled 21 Jan 2027

Mode	Owner	Reading this quarter	Threshold	Tier
Pipeline collapse	CHRO	Junior intake 4, down from 9	Floor 6	Concern
Competence drift	CRO	Override rate 2.1%; blind re-judgment agrees 91%	Re-judge 1 in 8	Watch
Inquiry groupthink	CRO	Divergence method run on 3 of 5 board decisions	Every consequential decision	Watch
Detection blindness	CRO · CHRO	Claims desk caught 31 of 40 planted cases (78%)	85%, set by the board	Concern
Authorship erosion	CRO	Tools-down rebuild: 6 of 10 completed unaided	Yearly, last Nov 2026	Watch
Agency erosion	CRO	0 overrides exercised; halt points named in 2 of 7 flows	A tested halt point in every flow	Watch

Verified by Head of Internal Audit (sample) · signed 18 Jan 2027Next reading 15 Apr 2027

“A register needs a fixed vocabulary of failure modes, or every owner will describe the same decay in different words and nothing will aggregate. There are six.”

Judgment Capital, chapter 10

Two lines, two owners

Pipeline collapse sits on the succession line, owned by the CHRO. Competence drift, inquiry groupthink, authorship erosion and agency erosion sit on the decision-quality line, owned by the CRO. Detection blindness sits on both: it wears down the bench being formed and the review running today.

Succession line

CHRO

Pipeline collapseA.1 · Pairing on live work

Decision-quality line

CRO

Competence driftA.2 · Monthly defect review Inquiry groupthinkA.3 · Divergence method Authorship erosionA.5 · Tools-down rebuild Agency erosionA.6 · Override channel

Detection blindnessOn both lines · A.4 · Seeded known-bad cases

Each mode below follows the same order: what it looks like, who owns it, what the app reads, the method that holds it, the workshop that installs it.

Pipeline collapse

The juniors who would become your seniors are no longer hired, or no longer formed.

The work that used to train them is automated, so the climb closes. A specialist leaves and no one has been formed behind her.

“A programme is a declaration. A calendar is behaviour.”

Judgment Capital, chapter 12

Owner: CHRO.
The app reads: Intake against a minimum junior share; coverage of the protected practice steps; time to reach competence without the tools; successor coverage for key judgment.
Method: Senior–junior pairing on live work, logged in both directions, with a floor of hours per junior per quarter (Appendix A, A.1).
In the book: Chapter 10 names it; chapter 11 sets the condition for bringing it down; chapter 14 on intake and pairing; Appendix A, A.1.
Workshop: W5 Pairing That Runs Both Ways

A young woman taking notes while a senior colleague walks her through a document — Pairing on live work: the junior does the work, beside someone who has done it before.

Pipeline collapse · SuccessionSample data

Junior intake: 4 this year, down from 9 · floor 6
Protected steps covered: 9 of 12
Pairing at the floor: 11 of 14 juniors this quarter
Successor coverage: 7 of 11 key-judgment roles
Tier: Concern

Owner CHRORead at every owning-committee meeting

Competence drift

The people in the role slowly settle into agreement with the machine.

Overrides fall, and nobody can say whether the machine got better or the reviewers stopped looking. The override rate alone cannot tell you.

Owner: CRO.
The app reads: The override rate beside a blind re-judgment floor: for example, one decision in eight re-judged without seeing the machine’s answer.
Method: A monthly defect review, ten failures read by eye against outcomes (Appendix A, A.2), and the blind re-judgment floor (chapter 11).
In the book: Chapter 10 names it; chapter 11 sets the blind re-judgment floor; Appendix A, A.2.
Workshop: W6 Overrides That Count

A man at a desk reading a printed page closely, pencil in hand — A reviewer reading the case before the machine’s answer is shown.

Competence drift · ClaimsSample data

Override rate: 2.1%, down from 3.4% a year ago
Blind re-judgment: 1 in 8 · 212 this quarter · agrees 91%
A.2 defect review: 10 failures read in January · next 3 Feb
Tier: Watch

The two numbers are read together, never the override rate alone

Inquiry groupthink

Every seat in the room draws on the same fluent source, and the room stops producing alternatives.

The gain is real: every director can be briefed to the same depth. The risk is that unanimity now certifies one account reflected in every chair. This is a risk for the board, the investment committee and the strategy room.

Owner: CRO.
The app reads: The share of consequential decisions where a divergence method was run before the recommendation was seen; how often dissent was recorded.
Method: A divergence method and a decision registry (Appendix A, A.3).
In the book: Chapter 10 names it; chapter 11 sets the condition for bringing it down; Appendix A, A.3.
Workshop: W7 The Divergence Round

A board meeting by tall windows: a grey-haired director speaks while the others at the table listen — The room where one fluent account can sit in every chair.

Decision registry · BoardSample data

Consequential decisions: 5 this quarter
Divergence run first: 3 of 5, before the recommendation was seen
Dissent recorded: 2 of 5
Next registry entry: Investment committee, 11 Feb
Tier: Watch

Threshold every consequential decision

Detection blindness

The people who review the machine’s work lose the ability to catch what it missed.

It usually shows up as the loss itself: the error nobody caught, arriving as a bill. So it is tested on purpose. A small cell outside the desk plants known-bad cases from your own history into the live queue. Nobody on the desk knows which.

“A monthly handful reads the desk, not the person.”

Judgment Capital, Appendix A

Owners: CRO presents it; CHRO owns the formation half.
The app reads: Caught out of planted, the miss rate and the time to catch, by desk, against a pass rate the board sets in advance.
Method: Seeded known-bad cases (Appendix A, A.4).
In the book: Chapter 10 names it; chapter 11 has the board set the pass rate in advance; Appendix A, A.4.
Workshop: W3 Fire Drills for Judgment

An old coloured postcard of an insurance company's main floor, rows of clerks at desks with files, in the claims and accounting departments — A claims department on a 1930s postcard, Keene, New Hampshire. Review by eye is older than the machine.

Detection blindness · Claims deskSample data

Caught of planted: 31 of 40 (78%)
Miss rate: 22%
Time to catch: Median 2.4 days
Pass rate: 85%, set by the board 12 Sep 2026
Tier: Concern since 14 Jan 2027

Reported by desk, never by person

Authorship erosion

The work reads well, and nobody in the building could rebuild it with the tools off.

The reasoning has moved out of the people and into the machine. A fluent output hides it until someone tries.

Owner: CRO.
The app reads: The unaided completion rate on a defined task set, tools off, supervised, at least once a year.
Method: The tools-down rebuild, measured (Appendix A, A.5). The book is plain about its limit: run it to detect, not to cure.
In the book: Chapter 10 names it; chapter 11 sets the condition for bringing it down; Appendix A, A.5.
Workshop: W3 Fire Drills for Judgment

A man in a green shirt writing on a contract at a table, a colleague's hand holding another page — Tools off: the same work, rebuilt by hand, supervised.

Tools-down rebuild · Credit memosSample data

Task set: 10 defined tasks, tools off
Completed unaided: 6 of 10
Last run: Nov 2026, supervised
Next run: Nov 2027
Tier: Watch

Run to detect. The cure sits on the succession line.

Agency erosion

Someone sees the machine is wrong, and the process gives them no way to make it count.

It fails two ways. The person overrides and is punished for it. Or the process admits no human input at all, and the machine’s output is the decision.

“Zero is not good news and not bad news; it is an unanswered question.”

Judgment Capital, Appendix A

“One punished override destroys the channel.”

Judgment Capital, Appendix A

Owner: CRO.
The app reads: Overrides exercised per quarter, and the share of automated flows with a named, tested halt point.
Method: A sanctioned override channel and drilled halt points (Appendix A, A.6).
In the book: Chapter 10 names it; chapter 11 sets the condition for bringing it down; Appendix A, A.6.
Workshop: W6 Overrides That Count

A grey-haired woman in glasses reading a document at a desk by a window, pen in hand — The person who sees it first needs a channel that counts.

Agency erosion · PaymentsQuestion to answer

Overrides exercised: 0 in Q4 2026
Named, tested halt point: 2 of 7 automated flows
Next halt-point drill: 9 Mar 2027
Tier: Watch

Sample data. Zero is asked about, not reported as good news.

Before you automate

One check serves all six: a one-page audit of each step before the automation case is approved.

The process owner and a delegate of the CHRO walk each step as it is actually worked. What does this step teach, and what will the changed role need that nobody has done before? The answer is one page per step. Each process owner then signs a map: the steps to protect, and the steps safe to hand over.

The app reads: The share of automation cases above your threshold that passed the audit before approval.
Method: The pre-automation value audit (Appendix A, A.7) and the process owner’s decision (Appendix B).
In the book: Chapter 10 on what to protect; Appendix A, A.7; Appendix B, the register on one page.
Workshop: W4 Mark the Work

Protect and hand-over map · Claims intakeSample data

Value audit, one page per step · walked by the process owner and a CHRO delegate

Read the first notice of lossTeaches what a normal claim looks likeProtect
Check the policy is in forceTeaches nothing the role still needsHand over
First-pass fraud checkTeaches the patterns the drills testProtect
Request missing documentsTeaches nothing the role still needsHand over
Set the first reserveTeaches the judgment a senior is promoted forProtect

Signed by the process owner (sample) · 9 Dec 2026Audited before approval 4 of 5 cases

A woman at a whiteboard explains a plan to colleagues seated in front of her — Walking each step as it is actually worked, before the case is approved.

The hardback of Judgment Capital: What Boards Must Preserve When AI Does the Thinking, by Oksana Sivokobilska and Alex Blumentals

Read them in the book

Chapter 10 names the six. Chapter 11 sets the conditions for bringing each back down. Appendix A gives the working method for each, with its sources and how well grounded it is. Appendix B puts the register on one page.

Chapter 10, The Register. Two lines, three tiers, six modes.The register
Chapter 11, Running the Register. Thresholds, the conditions for coming back down, the decisions no human saw.Thresholds and drills
Chapter 12, The Ladder. Five rungs, ten questions, the zero rule.The quarterly score
Chapter 13, The Agenda Item. Five questions for Monday, the motion, the verifier.The board reading
Appendix A, The Working Methods. Seven methods, one per mode and one before you automate.What the app schedules and logs
Appendix B, The Register on One Page. The entry, and each process owner’s protect or hand-over decision.Entries and the map

Order the book

Score your organisation

Ten questions, scored on evidence. You see your rung, your zeros, and the five questions to take to your next board meeting.

Rather talk first? Send an enquiry.