The AI transition framework

Failure at Strategy: adoption without diagnosis

The tool is adopted because competitors have one, not because a diagnosed problem calls for it. Because no baseline exists, no result can ever be validated — which is precisely the condition behind the headline statistics: MIT's finding that 95 per cent of generative AI pilots produced no measurable profit-and-loss impact (a preliminary, contested figure whose diagnosis — a learning gap, not a capability gap — survived the criticism), and S&P Global's finding that 42 per cent of companies abandoned most AI initiatives in 2025, up from 17 per cent a year earlier.

Early-warning driver: the absence of a written diagnosis and a captured baseline before money is spent. If the business case cannot state the do-nothing counterfactual, this failure mode is already active.

Failure at Outcomes: unvalidated results

Projected savings are announced before outcomes are measured; volume and cost are tracked instead of quality and resolution. The test every claimed outcome must survive: could an adversarial party — union, regulator, journalist, plaintiff — examine it without overturning it?

Early-warning driver: counter-driver divergence — cost falling while quality, satisfaction, or complaint metrics deteriorate.

Failure at Drivers: vanity metrics and gaming

Either no measure connects the tool to the outcome — the default condition of AI programmes — or the dashboard fills with auto-generated metrics nobody can name from memory. When usage itself becomes the target, Goodhart's law applies: when a measure becomes a target, it ceases to be a good measure, because people optimise the number rather than the outcome it was meant to indicate.

• Licence theatre: 30–50 per cent of provisioned AI licences typically show zero or near-zero usage; weekly active usage of a major workplace copilot runs at roughly 20–35 per cent of purchased seats.
• Mandate gaming: once AI usage enters performance reviews — Shopify's chief executive declared AI use "non-optional" and built it into reviews, and Meta announced "AI-driven impact" as a core performance expectation — prompts and sessions become the quantity employees inflate.
• Workslop: AI-generated volume that looks like work costs recipients nearly two hours of rework per incident and damages the sender's standing.
• Code churn: duplicated code blocks rose eightfold in 2024; churn rose from 3.1 to 5.7 per cent of changed lines.

Early-warning driver: a driver set that fails the Memory Test, lacks any counter-driver, or contains no vitality metric at all.

Failure at Actions: bolted-on AI and the wrong autonomy level

Two related mechanisms. First, AI bolted onto an unchanged process: exceptions still flow to humans who are now fewer or slower. Second, customer-facing or destructive authority granted beyond demonstrated reliability — the common final pathway through which upstream failures become public incidents.

Early-warning drivers: human-override and escalation rates; adversarial test results before every release; the share of target workflows actually redesigned.

Failure at Structure: nothing underneath

No grounding source of truth, no sanctioned tooling, no governance, no audit trail — and shadow use fills the vacuum. More than 80 per cent of employees use unapproved AI tools; breaches involving shadow AI cost about US$670,000 more than those without; among organisations suffering AI-related incidents, 97 per cent lacked proper AI access controls.

Early-warning drivers: share of AI traffic through sanctioned tools; registry coverage of the AI estate including vendor-embedded AI; update regression tests; vendor-assessment currency.

The eight failure modes: the reference catalogue

The eight modes below recur across the documented record. Each row states what happens, the Performance Bridge link where the chain actually breaks, the early-warning driver that would have surfaced it in time, and the anchor cases — every one of which appears in the case library with its full diagnosis.

Two observations cut across the catalogue. Almost every public incident combines FM3 with another mode, because the wrong autonomy level is the common final pathway through which upstream failures become visible explosions. And the modes cluster at the quiet ends of the chain — no diagnosis at Strategy, no validation at Outcomes — while the middle links produce the headlines.

Design on paper

At this point the top-down work is done: the diagnosis is written, outcomes have baselines and targets, the driver set has passed its tests, and the actions and structural requirements are specified. The Performance Bridge is coherent — on paper. Chapter 12's warning is that most transformations move straight from this point to execution without asking what is preventing the designed system from becoming operational, and then spend months wondering why nothing moves.

Blocker: the legal and compliance sign-off queue

Every AI use case waits on a review function that was never resourced for the volume. Everything downstream measures as slow, and no dashboard fixes it, because it is not a driver problem. The structural resolution is risk-tiered review — high-risk uses (employment, credit, clinical, safety decisions) get formal assessment; low-risk uses get lightweight registration with deliberately low friction so teams register rather than route around governance — plus resourcing the review function itself.

Blocker: the unfunded 70 per cent

Business cases funded for licences and integration but not for training, workflow redesign, and change management have allocated budget in inverse proportion to where the effort belongs. BCG's 10-20-70 rule puts 70 per cent of successful AI transformation effort in people and process change; the commonly observed failure is roughly 80 per cent of budget on platforms and tools, followed by surprise that adoption stalls. The transformation halts with the tooling deployed and nothing else funded — a wall, not friction, because no amount of good execution elsewhere compensates for work that has no budget line.

Blocker: data access

Teams that cannot reach the data their use case needs — because access sits with another function, or classification rules were never written with AI in mind — report exactly the "we know what to do but cannot do it" pattern that defines a blocker. Anthropic's survey of more than 500 technical leaders puts system integration (46 per cent) and data access and quality (42 per cent) as the top barriers to agent deployment. No metric clears this blocker; a decision does — someone with authority grants the access, writes the classification gate, or funds the pipeline.

Blocker: the sponsor conflict left unresolved

The department head whose function loses visibility, the business unit that will not share its data, the executive whose preferred vendor lost the selection — political blockers that only sponsor authority removes. AI transitions generate more cross-functional conflict than most programmes: 54 per cent of surveyed C-suite executives describe AI adoption as "tearing their company apart". The book's warning applies with force — acknowledged but unactioned blockers become permanent features. The empirical stakes: organisations where the chief executive is personally accountable for AI decisions report established returns at 14 per cent against 4 per cent elsewhere.

Blocker: AI procured like ordinary software

Enterprises that run AI implementation as a procurement exercise buy the tool but not the workflow redesign, training, and governance the tool needs — and the contract misses the clauses that matter: notification before material model updates, retention and exit-deletion terms, indemnity for AI outputs, audit rights over the vendor's own AI integrations. Two incidents show what those missing clauses cost: a vendor silently added generative AI to the US National Eating Disorders Association's approved rule-based helpline bot, which then gave harmful advice, and the Salesloft Drift breach spread through a vendor's own integrations. Clearing this blocker means rewriting the procurement pathway for AI purchases, because pushing more of them through the old pathway reproduces the same gaps.

Blocker: the unbuilt labour agreement

In co-determination jurisdictions, the workforce path runs through negotiated agreements. German works councils hold a mandatory co-determination right — not merely consultation — over technical systems capable of monitoring employee behaviour or performance, which covers most AI tools in human-resources contexts; the employer must reach agreement before introduction. A rollout plan that discovers this after deployment has found its wall. In Singapore the same institutional layer is an asset rather than an obstacle: Company Training Committees under the National Trades Union Congress give firms a ready-made joint forum where AI adoption, job redesign, and training are negotiated together.

Clearing the hinge: identify, own, act

Identify. List every blocker, then find the one that controls the others: not what makes execution harder, but what makes it impossible or close to it. If you removed this one, would the rest become manageable? That is the critical blocker. Until it is removed, improving everything else produces zero improvement in outcomes.

Own. The highest authority in the business owns blocker removal — the person who can reallocate budget, change reporting lines, override resistance, and make decisions that stick. The transformation lead identifies and escalates; the senior leader removes. If that leader treats blockers as someone else's problem, the bridge does not become operational.

Act. Set a deadline and treat it as non-negotiable. Blockers that are acknowledged but not acted on become permanent: people learn to work around them, workarounds become culture, and the blocker calcifies into "the way things are done here". Once the critical blocker is cleared, a new constraint will surface; that is the sequence operating as designed, because removing one constraint exposes the next one.

Execution in reality

With the hinge cleared, the bottom-up flow can carry the design: structure gets built and used, redesigned workflows run, drivers start moving, and outcomes begin to materialise against their baselines. The evidence signature of a cleared hinge is simple — the same drivers that sat flat during the blocked period begin to respond to actions. If they still do not, either another wall is standing or the driver set itself failed its tests.

The readiness check

The bridge rests on a foundation of empowerment; in command-and-control environments, drivers become surveillance and teams learn to hit metrics rather than improve outcomes. An AI programme on the same foundation produces mandates and performative usage. Five questions, translated for AI:

1. Can teams adjust how they use AI without seeking approval for every variation?
2. When an AI metric moves the wrong way, do leaders ask what happened, or who is to blame?
3. Do teams see their own usage and quality data in real time, or is it locked in leadership reports?
4. Is there room to experiment with different tools and workflows within guardrails?
5. When someone surfaces an AI failure — a hallucinated answer sent to a customer, a near-miss data leak — are they thanked or punished?

The fifth question carries extra weight in Asia-Pacific deployments, where face-saving and hierarchy dynamics demonstrably suppress error reporting: an incident process that depends on frontline staff volunteering bad news will underreport unless leaders de-personalise error and visibly reward the report. If readiness is missing, build safety and decision space first — the methodology can wait.

The data maturity check

Driver management depends on data: a driver can only be watched if it is measured, and only measured if the process captures the information in the first place. The four maturity levels below govern two distinct AI questions: whether the AI itself can be grounded in reliable data, and — more often missed — whether the outcome can ever be validated. A pilot in a process with no baseline data cannot be proved to have worked or failed, which is precisely the condition behind the many pilots claiming success nobody could demonstrate.

The trap: teams select ideal drivers they cannot track, then quietly substitute weaker ones because the data is available. Never accept a weak driver merely because it is easy to measure — and remember Gartner's prediction that organisations will abandon 60 per cent of AI projects unsupported by AI-ready data.

The sponsor question

Every sponsor says they are committed; commitment reveals itself only under pressure — when the transformation conflicts with something the sponsor values more, or when holding an influential person accountable would cost a relationship. Ask directly before starting, and record the answer: when this transformation creates conflict with other priorities, will you personally engage to resolve it? When obstacles require your authority to remove, will you remove them?

The book's claim that weak sponsorship dooms well-designed transformations now has quantitative support: where the chief executive is personally accountable for decisions based on AI outputs, confidence in AI strategy runs at 60 per cent versus 22 per cent, meaningful business value at 57 per cent versus 21 per cent, and established returns at 14 per cent versus 4 per cent.

When a check fails

A failed check is useful information: it says what must be fixed before the transition can work, and proceeding anyway is how the model makes things worse. The remedies map to the failed check:

• Readiness failed: build psychological safety and decision space first. Model curiosity about misses, democratise access to performance data, and delegate method decisions within a defined scope. Sophisticated measurement of an unaddressed dysfunction produces dashboards and gaming, not outcomes.
• Data maturity failed: either invest in data capture as a structural requirement before the transformation, run a deliberately small pilot on manual extraction while automating, or accept a documented proxy temporarily. Whichever path, capture the baseline before deploying anything.
• Sponsorship failed: do not compensate with methodology — no framework substitutes for a sponsor unwilling to spend political capital. Either renegotiate the mandate or shrink the scope to what the available authority can actually clear.

Phase 0 — foundations

1. Run the Strategy Test. Write the diagnosis: what specific obstacle is AI meant to overcome, in which one to three business domains? If no obstacle can be named, stop — any spend at this point is fear-of-missing-out spend.
2. Run the readiness and data-maturity checks. If blame culture would poison incident reporting, build safety first. Map candidate domains against the data-maturity levels; where a process has no baseline data, instrumenting it is the first AI investment, because a pilot without a baseline is unfalsifiable.
3. Ask the sponsor question directly and record the answer: when this transformation creates conflict with other priorities, will you personally engage to resolve it? The KPMG accountability data shows why this is not a formality — chief-executive accountability roughly triples the rate of realised value.
4. Deploy Tier 1 controls immediately: a written acceptable-use policy naming permitted tools and prohibited data classes; an approved-tool list with a sanctioned, convenient default; business-tier subscriptions with no-training defaults; a staff briefing built on the two incidents that make the risks concrete (Samsung engineers pasting source code into a consumer chatbot, and the deepfake video call that extracted US$25.6 million from the engineering firm Arup); and out-of-band verification with dual authorisation for payment instructions. These cost little and close the largest immediate exposures — and they honour the finding that bans fail while sanctioned channels work.

Phase 1 — diagnose, baseline, and design

1. Select one to three lighthouse domains — the small number of business areas where the approach will be proved end to end — prioritising workflows that are high-volume, well-bounded, and instrumented, the task shape where the experimental literature shows AI reliably helps.
2. Capture roughly 90 days of baseline data on the exact metrics the AI is meant to move.
3. Build the driver set before the pilot: two to four drivers per initiative, each passing the four tests, including at least one counter-driver and one vitality driver, with commitment levels assigned. Run the Gaming Check on every driver: how might someone hit this number without improving the outcome?
4. Write the pilot charter: success criteria as measurable business metrics against the baseline; kill criteria set in advance ("if the AI absorbs less than 15 per cent of the baselined work by day 80, or exception rates exceed the current error rate, we shut it down"); adversarial testing included; the human escalation path designed as a feature; one named owner.
5. Clear the hinge: list the blockers, identify the critical one, and have the senior leader remove it on a deadline before execution begins.

Phase 2 — pilot to production

1. Run production-intent pilots with real users in a controlled slice, at the assist-first autonomy level, long enough to validate quality before scaling — DBS ran nine months; the NHS ran a nine-site, 17,000-encounter trial before committing to a regional rollout.
2. Redesign the workflow around what the pilot shows. This is the Value Action — the book's term for the one action that directly delivers the value and cannot be skipped — and the strongest predictor of eventual financial impact; fund the people-and-process work at something approaching the 70 per cent of effort the evidence says it needs.
3. Review drivers weekly, outcomes at the gate. Where the counter-driver diverges — cost falling while quality falls — treat it as gaming or quality erosion and investigate before scaling. Kill weak pilots on the pre-agreed criteria and say so plainly; a disciplined exit is a success of method.
4. Stand up Tier 2 controls and the governance rhythm: the use-case registry with risk tiering, vendor assessments with model-change and retention clauses, human sign-off for high-stakes outputs, and the cross-functional committee cadence.
5. Channel bottom-up energy onto the sanctioned platform: open safe tooling to employees, celebrate the internal builders, and let the shadow economy surface into governance.

Phase 3 — scale and sustain

1. Scale what validated, through structure. Convert pilot practices into embedded processes: a standing dashboard in place of periodic reminders, a defined role in place of one heroic champion, a curriculum in place of one-off workshops. Test every element by asking what would keep working if the AI champion left tomorrow.
2. Manage AI as a portfolio rather than as single projects: a stage-gated pipeline mixing quick wins, process transformations, and a few strategic bets; a single benefits register per workforce population to prevent double-counting; honest reporting of the adoption ramp.
3. Report outcomes so they would survive adversarial review: against the public baseline and target, validated with a counterfactual, alongside the vitality drivers. Announce nothing as achieved that is only projected.
4. Keep the Watch List on the leadership agenda — the book's list of conditions no structure can install: trust, fear, psychological safety around error reporting, the judgment to override the AI, and alertness to the moment when a control itself has become the obstacle. These require sustained leadership attention, because they erode when attention moves elsewhere.
5. Track the regulatory trajectory in every operating jurisdiction, aligning with voluntary frameworks before they harden — the consistent pattern across Singapore, South Korea, Australia, and the European Union, and the cheapest compliance strategy on record.

The SME variant — the same sequence in weeks

Every link stays, but each is kept short. The owner writes the one-paragraph diagnosis of the costliest bottleneck. A funded trial — in Singapore, through the GenAI Sandbox or the Productivity Solutions Grant — tests one pre-approved tool against one workflow. The driver set is one operational number plus one counter-driver. Tier 1 controls are the whole security programme: business-tier subscriptions, a one-page acceptable-use policy, and the payment-verification procedure. The continuation decision follows the measured result inside the trial window.

The evidence that the compressed discipline works: Singapore's first sandbox cohort enrolled more than 150 SMEs and approximately 80 per cent continued using their solution after the funded period ended, and the worked example — Far East Flora's pre-approved support chatbot — reported a 67 per cent reduction in staff hours spent answering queries. Among Singapore's AI-using firms, 84 per cent rely on off-the-shelf tools: the SME on-ramp is buy rather than build, one workflow at a time, with one honest metric.

Threat: shadow AI

Employees using unapproved AI tools is the single most consistent empirical finding of the period, and the behavioural baseline every control programme has to assume. More than 80 per cent of employees report using unauthorised AI tools, and 68 per cent of security leaders admit doing so themselves. The financial consequence is quantified: breaches involving shadow AI cost an average of US$4.63 million — about US$670,000 more than breaches without it — and among organisations suffering AI-related incidents, 97 per cent lacked proper AI access controls and 63 per cent had no AI governance policy.

The design conclusion: shadow AI is evidence that employees perceive real value and the organisation failed to provide a sanctioned path. Roughly 35 per cent of workers say they would keep using unauthorised tools under an explicit ban. Every control in the tiers below is therefore paired with a legitimate, convenient alternative.

Threat: prompt injection

Content that manipulates a model into ignoring its instructions is ranked the number-one risk in the OWASP (Open Web Application Security Project) Top 10 for large-language-model applications, and it is structural to current models — a standing property, not an edge case.

The lethal trifecta, and the Agents Rule of Two

Security researcher Simon Willison named the combination that makes an AI agent exploitable: access to private data, exposure to untrusted content, and the ability to communicate externally. An agent holding all three can be steered by a poisoned input into exfiltrating whatever it can read — no software vulnerability required, because the vulnerability is the design itself. EchoLeak is the trifecta operating exactly as described: Copilot could read tenant data, ingested untrusted email, and could emit content externally, and a single crafted message connected the three.

Prompt injection remains an unsolved problem — models cannot yet reliably distinguish instructions from data — so the defence is architectural rather than instructional: remove one of the three properties from any agent that does not need all of them, instead of writing ever-sterner system prompts. The Replit incident, in which an autonomous coding agent deleted a production database despite direct instructions not to touch it, is the same lesson in miniature: the instruction failed to constrain the agent, and only removing the write credential would have.

The Rule of Two as a checkable gate

Meta's security team operationalised the trifecta as a design rule: within a session, an agent should satisfy no more than two of three properties — it can process untrustworthy inputs, it can access sensitive systems or private data, it can change state or communicate externally. If a task genuinely requires all three without a fresh session, the agent should not operate autonomously; it requires supervision through human-in-the-loop approval or another reliable validation. The rule becomes a one-line question at use-case registration: which two of the three properties does this agent hold, and who approved the third?

Two caveats belong in any briefing. Meta positions the rule as a supplement to, not a substitute for, least privilege and defence in depth — designs satisfying it can still fail against other threat vectors. And the rule bounds the blast radius of prompt injection rather than preventing the injection itself, which is precisely why it is honest: it accepts that the attack cannot yet be stopped and removes what the attack could reach. In Performance Bridge terms it is a pure structure control — a permission boundary that persists without ongoing attention, where a system-prompt instruction is a reminder that fails silently.

Threat: agentic compromise

The shift from chatbots to agents moved the attack surface from answers to actions. Every AI integration holding standing credentials into core systems is a new trust relationship that inherits none of the perimeter's controls.

Controls: least-privilege, short-lived credentials for every agent integration; environment separation and human approval gates for destructive actions, because the Replit incident showed that instructions alone do not constrain an agent; and Singapore's Cyber Security Agency has published a dedicated addendum on securing agentic AI.

Threat: deepfake fraud

The control that failed was a payment process, and the fix is also a process. The payment-authorisation procedure accepted a video call as identity verification. Out-of-band callback verification and dual authorisation for large transfers — ordinary process controls that involve no AI technology — would have defeated the fraud. Voice cloning now requires only seconds of audio, which every executive who has spoken at a conference or on an earnings call has already supplied. This belongs in Tier 1 for every company regardless of size.

Threat: intellectual-property risk, in both directions

Direction one: company IP leaking into third-party models

The commercial tier and the signed contract are the control. Business and enterprise tiers of the major providers do not train on customer content by default; consumer tiers commonly do. Two caveats belong in every policy briefing: litigation can override retention promises — a court order compelled production of 20 million anonymised consumer conversation logs in the New York Times case — and the protections are contractual, not architectural: the same model reached through a personal account carries none of them.

Direction two: the IP status of AI outputs

The US Copyright Office confirmed that human authorship is required and that purely AI-generated output is not copyrightable — so fully AI-generated marketing assets or content libraries may be unprotectable against copying by competitors. On training data, the industry pattern is settlement and licensing rather than definitive rulings (the authors' class action against Anthropic settled for US$1.5 billion), which leaves residual uncertainty with user-side companies and argues for vendors offering indemnification. For AI-generated code, indemnity is conditional — Microsoft's commitment requires duplicate-detection filters kept enabled — and suggestions match training-set code in roughly 1 per cent of cases, a real surface for open-source licence contamination at enterprise scale.

Tier 1 — foundational

• A written acceptable-use policy naming permitted tools and prohibited data classes.
• An approved-tool list with a sanctioned, convenient default — so employees have a legitimate path as easy as the shadow one.
• Business-tier subscriptions with no-training defaults; no confidential data in consumer tools.
• A staff briefing built on the two incidents that make the risks concrete: the Samsung engineers who pasted source code into a consumer chatbot, and the Arup finance employee who transferred US$25.6 million after a deepfake video call.
• Out-of-band verification and dual authorisation for payment instructions, regardless of how convincing the requester appears.

Risks addressed: shadow AI, casual data leakage, deepfake payment fraud. For a small or medium enterprise, this tier plus vendor attestations is the entire security programme — and it addresses the highest-loss incident types of the period.

Tier 2 — managed

• Enterprise agreements with negotiated retention, subprocessor, and exit-deletion clauses.
• Data-classification gates specifying which classes may enter which tools, enforced by data-loss-prevention tooling tuned for AI destinations.
• An AI inventory and registration process covering vendor-embedded AI — the estate you did not know you had.
• Vendor risk assessment extending to the vendor's own AI integrations, because the Salesloft Drift breach spread through exactly that route — with model-change notification clauses, because the National Eating Disorders Association's vendor changed an approved bot's behaviour without telling it.
• Human sign-off for high-stakes outputs — public content, legal, financial, clinical — reflecting the Air Canada liability principle that the company owns what its AI says.

Risks addressed: contractual exposure, an unknown AI estate, third-party AI supply chain, output liability.

Tier 3 — advanced

• Private endpoints or virtual-private-cloud deployment for sensitive workloads; AI gateways as a policy-enforcing control plane.
• Systematic red-teaming and evaluations against OWASP and MITRE ATLAS scenarios before and after every deployment and update.
• An AI-specific incident-response playbook: model rollback, guardrail-regression detection, prompt-injection triage.
• Least-privilege, short-lived credentials for every agent integration; environment separation; approval gates for destructive actions.
• The Agents Rule of Two applied as a design gate at use-case registration: no agent session holds untrusted inputs, sensitive data access, and external actions at once without human-in-the-loop supervision (the lethal trifecta panel).
• NIST AI Risk Management Framework alignment internally, and ISO/IEC 42001 certification where customers demand attestation.

Risks addressed: prompt injection, agentic compromise, regulatory attestation. Two design principles govern all three tiers: every control is paired with a sanctioned path, because prohibition alone demonstrably fails; and the highest-loss incidents of the period (Arup, Salesloft Drift) were or would have been defeated by process controls, not AI-specific technology.

The framework stack

Four voluntary frameworks plus two binding regimes cover most of what a transitioning company needs to align with. The common pattern is NIST for internal discipline and ISO for external attestation.

Sector overlays: in healthcare, every ambient-documentation vendor is a business associate requiring signed agreements, consent is now in litigation, and an AI that informs clinical decisions is likely regulated as a medical device; Singapore's refreshed AI in Healthcare Guidelines (AIHGle 2.0, March 2026) map developer, deployer, and user responsibilities. In financial services, the Monetary Authority of Singapore has moved from voluntary principles to a consultation on binding AI risk-management guidelines for all financial institutions.

Two kinds of outcome, one pairing rule

The book distinguishes performance outcomes — the results of the period — from vitality outcomes: the organisation's sustained capability to keep producing results, carried in trust, skills, energy, and the willingness to surface problems. A transition that hits its performance targets while burning vitality has borrowed its results rather than earned them, and "burning vitality for performance" is one of the six mistakes that break the bridge.

The operating rule that makes this measurable is pairing: every performance driver in an AI programme gets a vitality counter-driver that makes its gaming, or its collateral damage, visible. The payments company Klarna illustrates the stakes: its cost per contact fell while customer satisfaction and escalation behaviour deteriorated, and because only cost was on the dashboard, the divergence stayed invisible for a year until it forced a public reversal. A paired dashboard would have shown it within weeks.

Trust in AI outputs

The global baseline is poor, which is why a programme that never measures trust is assuming a condition the evidence says is absent: a 48,000-respondent study across 47 countries found 66 per cent of people using AI regularly but only 46 per cent willing to trust it, and 48 per cent of employees admitting to using AI in ways that contravene company policy.

How to measure it. A monthly pulse with specific behavioural questions — "Would you send an AI draft to a client without rewriting it?" — rather than an annual engagement survey. The annual instrument fails the book's WATCH test: it arrives months after the damage, far too late to steer.

Displacement fear

The Asia-Pacific paradox makes this the region's central change-management driver: 53 per cent of frontline workers fear job loss — against 36 per cent globally — in the same survey that shows the region leading the world in adoption, with Singapore, South Korea, and Thailand reporting the highest concern. In Singapore, only 15 per cent of employees feel confident about job security.

How to measure it. The same anonymous pulse, plus attrition and internal-transfer patterns in the teams the programme touches. What answers the fear is a demonstrated redeployment path rather than reassurance: IKEA retrained 8,500 call-centre workers into a measured design-advisory channel, and the Italian bank Intesa Sanpaolo negotiated its workforce transition with its unions. A town-hall promise is only a projection, and the workforce knows it.

Shadow-usage rate

More than 80 per cent of employees report using unapproved AI tools at work, 68 per cent of security leaders admit doing so themselves, and roughly 35 per cent of workers say they would keep using unauthorised tools under an explicit ban. Read correctly, a high shadow-usage rate tells you the sanctioned tools are less useful or less convenient than the unsanctioned ones — and it also measures a workforce appetite that the successful programmes put to work rather than suppressed.

How to measure it. The share of observed AI traffic flowing through sanctioned tools — a structural metric — plus a periodic anonymous self-report. The successful cases show the driver moving when the sanctioned offer improves: employees of the Spanish bank BBVA built more than 2,900 custom assistants within months of receiving a governed platform, and staff at DBS, Singapore's largest bank, have built roughly 26,000 personal agents on its internal platform.

Incident and near-miss reporting rate

Every incident-response plan, model registry, and governance committee depends on one behaviour: a person voluntarily reporting that the AI got something wrong. No procedure can produce that behaviour; it appears only where people believe reporting is safe. Practitioners across Asia consistently report that employees rarely volunteer bad news unless leaders make it explicitly safe, and healthcare research links inclusive leadership, through psychological safety, to error reporting.

How to measure it. The count of voluntarily reported AI errors and near-misses, read against usage volume — and read with the inversion rule: early in a transition, a rising count usually means the reporting channel works. The alarming signal is an empty incident register beside heavy usage, because that combination measures suppressed reporting rather than safe operation. De-personalise error — attribute the failure to the system rather than to the person who surfaced it — and visibly reward the report.

Human-override rate

The share of AI outputs overridden or heavily edited by the responsible human is one of the few drivers that must be read two-sided. Persistently high override rates mean the tool is not trusted or not good enough for the workflow it sits in. Near-zero override on consequential decisions signals automation complacency — the condition that let a chatbot's invented bereavement policy reach an Air Canada customer, and an invented licensing policy reach Cursor's users, unchecked.

The third reading is saturation: when alert volume is high enough, override becomes reflexive dismissal. The Epic sepsis model — a widely deployed hospital early-warning system — showed this at scale: external validation found it missed most sepsis cases while alerting so constantly that clinicians learned to dismiss it. The driver only works when someone owns the question of which of the three readings applies.

Rework burden on recipients

"Workslop" — AI-generated content that masquerades as good work but lacks the substance to advance a task — is the named version of this driver's failure state: 40 per cent of surveyed workers had received it in the previous month, each incident cost recipients nearly two hours of rework, the estimated invisible tax runs to roughly US$186 per employee per month, and recipients marked the sender down as less capable and less trustworthy. The survey was self-selected, so treat the numbers as suggestive; the concept is the useful part.

The code version is quantified more cleanly: duplicated code blocks rose eightfold in 2024 and code churn — lines reverted or substantially revised within two weeks — rose from 3.1 to 5.7 per cent of changed lines. More AI-assisted output shipped; more of it came straight back.

How to measure it. Sampled downstream ratings of AI-assisted work and rework-time tracking, paired as the counter-driver to any output-volume metric.

Redeployment of freed time

Time saved is a legitimate leading indicator, but it is a proxy — value materialises only when the freed time is deliberately redeployed, and the book's surrogation warning says teams forget proxies are proxies and optimise the number itself. The evidence says the forgetting is the norm: in a 3,200-respondent study, companies were more likely to reinvest AI time savings in more technology (39 per cent) than in employee development (30 per cent), 32 per cent simply increased workload, and nearly four in every ten gained hours were lost to reworking AI output.

How to measure it. Track where the freed hours verifiably went — development, redesigned work, or simply more volume. An organisation reporting "hours saved" as value, without a redeployment mechanism and a destination measurement, is optimising a proxy that has decoupled from the outcome.

Skill progression

MIT's GenAI Divide report — the study behind the widely quoted finding that 95 per cent of generative-AI pilots showed no measurable profit impact — diagnosed the cause as a learning gap: organisations stall because neither their tools nor their people advance beyond the first plateau. Training hours completed will not catch this, because hours accumulate without learning. The driver that does is the number of trained employees observed applying the skill in their own workflow, tracked over time, with internal certification progression as the long-run signal.

How to read vitality drivers

Rule one: early in a transition, rising incident reports are good news. A climbing count in the first months usually means the reporting channel works and people believe using it is safe. The untrained reading — "incidents are up, the programme is in trouble" — punishes exactly the behaviour the programme needs, and in face-conscious cultures it will silence reporting permanently. State the inversion explicitly to every leader who will see the dashboard.

Rule two: vitality drivers are the counter-drivers of the performance set. Pairing each performance driver with a vitality driver is what makes the Klarna pattern — cost falling while satisfaction, escalations, and complaints deteriorate — detectable by design: a paired dashboard shows the divergence in weeks, where a cost-only dashboard concealed it until the company had to reverse in public.

Rule three: treat silence as a measurement gap, never as consent. A workforce can be anxious, adopting, and silent all at once — 58 per cent of Asia-Pacific respondents would use AI without company approval while half fear for their jobs. Where hierarchy or face dynamics suppress individual voice, make the pulse survey anonymous and the questions specific, and use the institutional channels — Singapore's Company Training Committees, European works councils, unions — as the collective route for what individuals will not say alone.

The vitality bankruptcy pattern

The clearest change-management failures of 2024–2026 share one shape: the performance number was real and loudly announced, the vitality account drained unmeasured, and the debt was called in publicly.

Mandates and bans are the same mistake in opposite directions. Usage mandates — Shopify made AI use a performance-review criterion, and Meta announced "AI-driven impact" as a core performance expectation — buy visible adoption at the price of performative usage where training and redesign are absent, because once usage is graded, the graded quantity is what employees inflate. Bans buy visible compliance at the price of hidden usage: Samsung banned generative AI after its source-code leaks, and roughly 35 per cent of workers say they would defy an explicit ban. Both approaches substitute an instruction for the work of enablement and trust, and both optimise a compliance metric while the underlying vitality drains.

What working change management looks like

Sanctioned capability before any demand (Structure). Panasonic Connect built a secure internal assistant that removed the compliance ambiguity around usage before asking anything of employees, paired it with visible top-management commitment, and measured honestly: 788,000 hours saved in fiscal 2025 against a stated long-term goal. DBS built its internal tools and training curriculum on the same logic — the safe path existed before usage was expected.

Channel the energy that already exists (Actions). The shadow-AI statistics, read correctly, measure appetite. The 2,900 assistants built by BBVA's employees and the 26,000 personal agents built by DBS staff did not come from persuasion campaigns; both companies provided a governed outlet, and demand that already existed flowed into it.

Answer displacement fear with a demonstrated path (Outcomes). IKEA's reskilled call-centre workers and the union-negotiated transition at the Italian bank Intesa Sanpaolo could point to named colleagues whose work changed and survived — which is evidence in the Outcomes sense, where a promise is only a projection. The message must also fit the labour context: in Japan, where the workforce problem is shortage, the credible framing is that AI covers the work there are no longer people for; in Singapore, where displacement fear is among the world's highest, the message must address that fear directly.

Leaders model what no policy can install (the Watch List — the book's list of conditions structure cannot solve). DBS describes the hardest part of its transformation as cultural: moving a banking culture built on flawless execution to one that experiments, framed internally as winning "heads and hearts". The leadership behaviours are the lever for every driver structure cannot move: visibly using the tools, receiving bad news about the AI well, and thanking the person who reports the near-miss.

The change-management plan as a bridge artefact

1. A named owner and a budget line. The 70 per cent of effort that belongs in people and process gets an owner and funded time; a plan without one restates the inverse allocation — 80 per cent of budget on platforms — as intent.
2. Enablement sequenced before any demand. Sanctioned tool, training with protected time to learn, and at least one redesigned workflow exist before usage appears in anyone's performance expectations.
3. Two or three vitality drivers with commitment levels, paired against the initiative's performance drivers and reviewed in the same governance rhythm as the security controls, so the operating review cannot ignore them. Using the book's three commitment levels — GUARDRAIL for what must never be breached, COMMITTED for what the team is expected to hit, STRETCH for genuine ambitions — incident-reporting suppression sits at GUARDRAIL, trust and redeployment targets at COMMITTED, and cultural transformation of the DBS depth is honestly a STRETCH.
4. The interpretive rules briefed to every dashboard reader — in particular the inversion on incident reporting — before the first review meeting rather than after the first misreading.
5. A demonstrated redeployment path for the roles the programme touches, stated in Outcomes terms: named roles, destination work, and a validation check.

1 · A strategic diagnosis with a quantified counterfactual

The case opens with the business problem — the same "what's in the way?" that opens the Strategy link — and a quantified do-nothing scenario. Without a counterfactual, the funding request is a technology purchase; with one, it is a comparison a chief financial officer can actually make. The failure statistics of the period describe programmes funded without this component: pilots whose success could never be demonstrated because nobody stated what failure would have looked like.

2 · A measured baseline

Roughly 90 days of pre-deployment data on the exact metrics the AI is meant to move. A pilot in a process with no baseline is unfalsifiable — precisely the condition under which the MIT GenAI Divide report found pilots claiming success nobody could demonstrate. The gold standard is an experimental design: the economists Brynjolfsson, Li, and Raymond used a staggered rollout across 5,179 support agents as a natural control group, measured issues resolved per hour — a business metric rather than a self-report — and found productivity up 14 per cent on average and 34 per cent for novice agents. Even a simple matched comparison group is far better than no counterfactual at all.

3 · A full total-cost-of-ownership model

Gartner has estimated real generative-AI deployment costs at US$5 million to US$20 million, and 96 per cent of deploying organisations faced higher-than-expected costs. The model must itemise: licences and seats; token and inference consumption at projected volumes (agentic workflows consume 5 to 30 times more tokens per task than a chatbot query — the cost line is a function of usage, never a fixed fee); integration engineering; data preparation; security and compliance review; monitoring and evaluation infrastructure; training; and sustained change management. The hidden 70 per cent is the people-and-process line the 10-20-70 rule demands — and vendor-embedded AI has produced cost uplifts of around 30 per cent on existing tools without upfront disclosure. On sourcing: 76 per cent of enterprises now buy rather than build, and bought solutions convert pilot to production at 47 per cent against 25 per cent; the working rule is to buy the platform, build only the differentiating layer, and treat integration as the main engineering effort.

4 · Benefits stated in one named ledger

The single most common inflation in AI business cases is conflating the three time-value ledgers — utilisation, capacity, and headcount avoidance (see the teal row below for the full treatment). The case must name which claim it is making and the conversion mechanism: freed time redeployed to what, hiring avoided against which demand forecast, or which named workforce plan. Quality effects go on both sides: repeat-inquiry reductions are a credit; rework of AI output (nearly two hours per workslop incident; repeat contacts at double cost behind inflated deflection rates) is a debit most cases omit.

5 · Scenario discipline

A conservative case at partial benefit realisation, with assumptions documented per scenario. The empirical justification is the skew of AI returns: Deloitte's 2024 wave found about 20 per cent of organisations achieving returns above 30 per cent on their most advanced initiative while the median struggled — a distribution that punishes single-point estimates. The skew is also the argument for funding a portfolio (quick wins, process transformations, a few strategic bets) rather than one flagship: a single bet maximises variance, not expected value.

6 · An adoption ramp

A case that books full benefit in the first quarter is contradicted by the entire survey base: 91 per cent of finance organisations report only low-to-moderate initial impact, with impact compounding over years of adoption maturity — organisations further along the curve are two to three times more likely to report moderate or high impact. Most executives expect two to four years to satisfactory returns; only 6 per cent achieve payback in under a year. Model the ramp, and count the early capability building — data pipelines, evaluation infrastructure, AI-literate staff — as option value that lowers the cost of every subsequent use case.

7 · Pre-agreed success metrics, kill criteria, and stage gates

Success criteria defined before the pilot begins, as measurable business metrics against the baseline — "reduce average handle time from 8.2 minutes to 5.5 minutes", not a satisfaction score. Kill criteria set in advance are what make the funding request credible: "if the AI absorbs less than 15 per cent of the baselined work by day 80, or exception rates exceed the current error rate, we shut it down." A disciplined exit is a success of method — McDonald's two-year, hundred-restaurant drive-thru test ended cleanly when accuracy never reached the threshold at which removing the human paid. Stage gates run quick wins, then pilots with go and no-go decisions, then scale, with weak pilots actively retired at each gate.

8 · A named accountable owner

The empirical justification for refusing to fund orphan initiatives: in organisations where the chief executive is personally accountable for decisions based on AI outputs, confidence in AI strategy runs at 60 per cent versus 22 per cent elsewhere, meaningful business value at 57 per cent versus 21 per cent, and established returns at 14 per cent versus 4 per cent. Leaders with strong cost visibility are five times more likely to achieve returns (15 per cent versus 3 per cent) — which also argues for a single AI investment register, since 42 per cent of leaders report only partial visibility of AI spending and the same productivity gain is routinely double-counted across initiatives.

The three time-value ledgers

Time saved multiplied by loaded labour cost is the dominant benefit claim in AI business cases, and it needs three separate ledgers that must not be conflated, because they are different claims with different evidence requirements.

The wider calibration: workers using AI save on average around 5 per cent of weekly hours; self-reports overstate the savings (in one randomised trial, developers believed they were 20 per cent faster while measuring 19 per cent slower); and a Danish study matching 25,000 workers to administrative records found time savings of about 3 per cent of hours with no measurable effect on earnings. Task-level gains become firm-level gains only when someone deliberately redesigns what the freed time is spent on — which is what Verizon did by redirecting it into retention and sales conversations.

Two worked examples

Service operations: customer-service deflection economics

The unit economics: human-handled tickets average US$8–12 (US$25–35 in business-to-business software support) against roughly US$0.50–1.05 for AI-handled tickets, with realistic containment of 40–60 per cent for well-documented tier-one query types. The critical measurement distinction is deflection against containment: deflection counts sessions that ended in the bot; containment counts contacts fully resolved with no follow-up on any channel within 24 hours. True containment typically runs 15–25 percentage points below reported deflection, and the customers who call back cost roughly double. The credible case prices benefits on containment, nets off repeat contacts, and measures by category before scaling — as the Singapore telecommunications company Singtel did, committing to scale only after measuring about 73 per cent containment on troubleshooting queries and 76 per cent on roaming sign-ups.

Healthcare: ambient clinical documentation

The evidence is deliberately two-sided and the honest case reflects it. Time savings are real but modest per encounter: Kaiser Permanente's 15,791 hours across roughly 2.5 million encounters, and 16 minutes of documentation per eight hours of care in the five-centre academic study. Burnout effects are the strongest measured outcomes — a 21.2 per cent reduction in burnout prevalence at Mass General Brigham. But the Peterson Health Technology Institute's multi-system evaluation concluded that clear financial returns have not yet been demonstrated, with licence costs of roughly US$100–600 per provider per month. The honest business case in this domain is therefore a retention-and-capacity case with option value — clinician time, burnout, and turnover-cost avoidance modelled explicitly as assumptions — and not a payroll-savings case.

The six failure patterns

1. Benefits claimed on gross time savings. Multiplying minutes saved by salary assumes 100 per cent of freed time converts to value; the telemetry and survey evidence shows it is reinvested diffusely, and self-reports overstate the savings to begin with.
2. The adoption ramp ignored. Full benefit booked in quarter one against a survey base showing years-long compounding.
3. Rework and verification costs omitted. Nearly two hours per workslop incident; repeat contacts at double cost behind inflated deflection rates.
4. Double-counting across initiatives. The same "10 per cent productivity gain" for the same population claimed by multiple projects; the control is a single benefits register per workforce population.
5. No counterfactual. Pre-and-post comparisons without a control group attribute secular trends — volume changes, seasonality, other process changes — to the AI.
6. Benefits priced on vendor-defined metrics. Where the vendor bills per "resolution" or "outcome", the buyer must independently audit that billable outcomes match business outcomes — the deflection-versus-containment problem applies directly to invoices.

Singapore and Asia-Pacific

The most institutionally supported environment

Singapore is the case where the Structure link is substantially built by the state: the National AI Strategy 2.0 (December 2023) with more than S$1 billion committed at Budget 2024; the IMDA (Infocomm Media Development Authority) Model AI Governance Framework for Generative AI with its nine dimensions; the AI Verify Foundation's testing tools including Project Moonshot; sector guidance from MAS (Monetary Authority of Singapore) for finance — on a visible trajectory from voluntary principles to binding risk-management guidelines — and AIHGle 2.0 for healthcare; and an SME grant architecture (Productivity Solutions Grant, GenAI Sandbox, Digital Leaders) whose first sandbox cohort saw roughly 80 per cent of participating SMEs continue after funding ended. The caveat: subsidy has not closed the SME divide (14.5 per cent against 62.5 per cent large-firm adoption), because the binding constraint is absorptive capacity, not cost.

The central change-management fact: the adoption-anxiety paradox

APAC leads the world simultaneously in adoption and in fear: 78 per cent of employees use AI at least weekly (versus 72 per cent globally), yet 53 per cent of frontline workers fear job loss (versus 36 per cent), with Singapore, South Korea, and Thailand highest — and 58 per cent would use AI without company approval. Only 15 per cent of Singapore employees feel confident about job security. The change-management task this creates is unusual: the enthusiasm already exists, so the work is converting existing, anxious, partly ungoverned energy into governed adoption.

Culture-aware design rules

• Senior sponsorship carries more weight where hierarchy is accepted — and so does its liability: pair endorsement with an explicit, repeated invitation of dissent.
• Face-saving suppresses the error reporting AI governance depends on; de-personalise error and reward the report.
• Kiasu — Singapore's fear of losing out — plausibly drives both fast adoption and reluctance to be the visible first failure; treat it as an interpretive lens and make early pilots low-stakes.
• Institutional voice channels compensate for suppressed individual voice: NTUC's Company Training Committees tie AI adoption to negotiated wage and career outcomes, converting individual anxiety into collective voice.
• Japan is the mirror image — lowest adoption (51 per cent) but lowest job fear (40 per cent) — because labour institutions make displacement unlikely; the credible message there is that AI covers work there are no longer people for, while in Singapore the message must address displacement directly.

Europe and the United Kingdom

The adoption story is better than its reputation

On official all-firm surveys, EU enterprise adoption (20.0 per cent in 2025, up 6.5 points in a single year — the fastest rise recorded) matches or exceeds the comparable US Census figure, and Denmark (42 per cent) exceeds anything measured in the US on the same instrument. Where Europe genuinely lags is capital deployed and the share of firms scaling beyond pilots: 48 per cent of Europe's largest companies had scaled a transformational generative-AI initiative against 31 per cent of smaller ones — and Europe's economy is small-firm-heavy.

What a deployer must actually do

The EU AI Act's near-term duties for a typical deploying enterprise are AI literacy, prohibited-use screening, and supplier due diligence; the heavier high-risk conformity work (employment screening, credit scoring) arrives on the staged timeline. The UK remains principles-based through sector regulators, with legislation repeatedly deferred. The deeper structural difference is labour institutions: German works councils hold mandatory co-determination over AI systems capable of monitoring performance, so rollouts are negotiated, slower, and redeployment-heavy.

North America

The capital lead, and the regulatory whipsaw

US private AI investment runs roughly 23 times China's, and US generative-AI investment exceeds China and Europe combined — but on official all-firm surveys US adoption (17–20 per cent) no longer leads the EU. The federal direction reversed twice in two years: a January 2025 executive order revoked the prior administration's AI order, and a December 2025 order created a Department of Justice task force to sue states over their AI laws, while Colorado gutted and delayed its landmark state act in May 2026. The operative constraint for a deployer is litigation risk and a volatile state patchwork — employment discrimination (the iTutorGroup settlement; the Workday collective action reaching the vendor itself), consumer protection, and product liability — rather than a single rule.

The reference deployments, and the claim inflation

The strongest governed rollouts are here — JPMorgan's single secure gateway for 200,000 users, Morgan Stanley's evaluation-first assistant, Kaiser Permanente's published ambient-scribe study, Cleveland Clinic's year-long vendor bake-off — alongside the sharpest natural experiment: Providence deployed comparable scribe technology and reached only about 8 per cent active clinician use. This is also the home of the unaudited chief-executive productivity claim: Amazon's "4,500 developer-years saved" and Salesforce's "AI does 30 to 50 per cent of the work" should be discounted against third-party-evaluated results; Accenture's US$5.9 billion in generative-AI bookings against US$2.7 billion recognised revenue is the cleaner demand indicator — client commitments run ahead of delivered work.

Middle East

The Gulf pattern is state-led and infrastructure-first, with the state simultaneously investor, customer, and regulator. Saudi Arabia's HUMAIN carries a reported US$77 billion infrastructure plan targeting 6.6 gigawatts of data-centre capacity by 2034; the UAE's Stargate project targets 1 gigawatt; and a November 2025 US export authorisation for up to 70,000 advanced chips unblocked frozen capital. Regulation runs through procurement and certification rather than statute: Dubai's AI Seal — a six-tier certification with 325 corporate applicants by mid-2025 — is positioned to become the credential for bidding on government AI work, and the UAE Cabinet has announced that half of government services will move to autonomous AI systems within two years. Corporate deployments follow the sovereignty logic: Aramco built its industrial large language model in-house on decades of proprietary drilling data, and ADNOC signed a US$340 million agentic-AI contract for its upstream value chain. For a vendor or partner, certification readiness is a market-entry requirement, not an afterthought.

Latin America

The region shows high grassroots usage against a severe investment deficit: 14 per cent of global visits to AI tools with only 11 per cent of global internet users, yet just 1.12 per cent of global AI investment against 6.6 per cent of global output. Brazil's PL 2338 — an EU-style risk-based bill with fines up to R$50 million and a contested training-data remuneration requirement — passed the Senate in December 2024 and remains before the Chamber of Deputies. The flagship corporate deployments run on imported foundation models but are genuinely instructive: Nubank cut chat response time 70 per cent with roughly 55 per cent of tier-one enquiries resolved by its assistant across more than 2 million monthly chats, and Mercado Libre built an internal AI platform ("Verdi") that lets its 17,000 developers compose AI workflows — a platform-layer answer to the sprawl of unconnected AI agents.

What travels, and what is context-dependent

Four patterns recur in every region

1. Staged rollouts with measured gates beat big-bang deployments. BBVA's ramp — 3,300 licences, then 11,000, then all 120,000 employees, each stage gated on measured usage — is the counter-model to Klarna-style substitution: adoption first, redesign second.
2. The rollout design determines adoption, because the technology is broadly comparable. The US health systems Kaiser Permanente and Providence bought similar ambient-scribe products; one reached 7,260 actively using physicians, the other roughly 8 per cent of its clinicians.
3. Two use cases have replicated, quantified outcomes across regions: ambient clinical documentation and first-line customer service — everywhere from California to London to São Paulo.
4. Vendor and chief-executive claims systematically outrun independent evaluation. Weight peer-reviewed and third-party-evaluated results; treat the rest as marketing until validated.

The multinational minimum

An EU AI Act compliance baseline applied group-wide (the strictest broadly applicable regime often becomes the de facto global standard); country-level workforce processes wherever works councils or sector unions exist; US deployments reviewed for state-law exposure and litigation risk; and certification and evaluation-registry readiness for government business in the Gulf or the UK. For Singapore and APAC companies expanding outward, the asymmetry runs in their favour on speed but against them on compliance depth — Europe's staged obligations and negotiated-labour environment are the two systems most unlike their home experience.

The cross-case analysis: where the chain breaks

Every case in the library below carries a five-segment verdict strip: for each link of the Performance Bridge, the public record either shows the case ran that link well, shows the chain broke there, or offers no evidence either way. Counting those verdicts across all 43 cases produces the table below, and the heat map above the library shows the same tallies as bars. Three findings stand out.

First, success is a complete chain, not a strong link. Fourteen of the twenty-one validated successes ran all five links well — the winning profile is not excellence at one link but the absence of a broken one. The successes share the same construction: structure built before demands were made, workflows redesigned around the tool, a small honest driver set, and outcomes validated against a baseline someone else could check.

Second, failures compound. Thirteen of the failure and caution cases broke at two or more links, because a break upstream travels: a missing diagnosis leaves nothing to validate, and missing structure turns ordinary employee enthusiasm into leakage. The links that broke most often are Structure and Drivers — eleven cases each — which are exactly the two links the book identifies as the seat of leadership leverage, and the two that most programmes neglect.

Third, the most common condition at Outcomes is silence. Outcomes shows fewer outright breaks than Structure or Drivers, but it is the link most often marked "not evidenced": in nineteen of forty-three cases, the public record contains no validated outcome at all — no baseline, no audited result, no independent check. An organisation that cannot show its outcome has not necessarily failed, but it cannot demonstrate success either, and the aggregate statistics of the period suggest most of that silence is not modesty.

The verdicts are editorial judgments applied to the published record, case by case; the reasoning for each is in the case's expandable reading below, and the "Diagnose it yourself" mode hides these verdicts so you can form your own before comparing.