Skip to content

2026 05 20 banking ai syntactic confidence trap

Jump to bottom
github-actions[bot] edited this page May 20, 2026 · 1 revision

How should banks stop fluent but weakly evidenced Artificial Intelligence (AI)-generated compliance narratives from being mistaken for verified truth?

Research Question

How does polished, authoritative generative Artificial Intelligence (AI) prose affect reviewer behaviour in anti-money laundering (AML) and Know Your Customer (KYC) workflows, and which interface and evidence controls prevent analysts from treating fluent summaries as verified fact?

Scope

In scope:

  • Behavioural effects of polished AI narrative output on compliance analysts.
  • User-interface indicators that expose uncertainty and provenance, including confidence markers, citation traces, and signals showing whether the draft includes the key supporting documents, transactions, and policy references needed for review.
  • Regulatory implications under European Union (EU) AI Act Article 14 and relevant financial-crime compliance expectations.
  • Liability allocation risks when analysts adopt unsupported AI summaries.

Out of scope:

  • Broader AML policy design unrelated to AI assistance.
  • General-purpose writing assistants outside regulated compliance workflows.

Constraints: Focus on controls that can be audited and embedded into production compliance tooling.

Context

Fluent compliance narratives can become a control failure in regulated banking review queues because automation bias, confirmation bias, and weak evidence visibility make polished prose easier to approve than underlying case evidence. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html]

Approach

  1. Review empirical evidence on tone, confidence, and truth-perception effects in AI-assisted review.
  2. Map required transparency and oversight obligations to concrete user-interface and workflow controls.
  3. Analyse legal and operational accountability implications when analysts rely on unsupported AI output.
  4. Produce an auditable control checklist for banking compliance implementations.

Sources

Related

Research Skill Output

(Full output from running the research skill, retained verbatim in the completed item. Sections 0 to 5 are the investigation, and Section 6 seeds the Findings section below.)

§0 Initialise

  • Question: how polished generative AI prose changes analyst behaviour in AML and KYC review, and which interface and evidence controls stop analysts from mistaking fluent summaries for verified fact.
  • Scope: behavioural effects, interface and workflow controls, regulatory expectations, and accountability implications in regulated banking review.
  • Constraints: public sources only, URL-backed citations, explicit fact-inference-assumption labels, auditable production controls, and repeated cross-reference to materially related completed items.
  • Output: knowledge.
  • Constraint mode: full.
  • [fact; source: https://davidamitchell.github.io/Research/research/2026-04-30-human-bias-ai-trust-rlhf-sycophancy.html; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html; https://davidamitchell.github.io/Research/research/2026-05-17-ai-policy-ambiguity-cognitive-closure-confirmation-bias-risk.html] Prior completed items already establish that humans can over-trust polished AI output, that review queues can collapse into rubber-stamping, that stochastic model output is a weak final control surface for regulated decisions, and that ambiguity plus confirmation bias can reduce escalation.
  • [assumption; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/] Working assumption: direct randomized banking studies on AI-written AML or KYC narratives are scarce enough that the strongest available evidence comes from adjacent high-stakes decision-support settings with the same core mechanism, recommendation review under ambiguity, workload, and override responsibility.

§1 Question Decomposition

  • Root question: when a banking analyst receives a polished AI-written compliance narrative, what evidence shows that the prose can shift reviewer behaviour, and what controls are necessary so the narrative remains a proposal rather than a truth substitute?
  • A. Behavioural mechanism
    • A1. What does the automation-bias literature show about over-reliance on automated recommendations?
    • A2. Does the timing of AI advice, before or after an independent judgment, change human accuracy?
    • A3. Do reviewers trust AI recommendations more when the recommendation matches their prior intuition?
    • A4. Do natural-language explanations create a special over-reliance risk when they are convincing but wrong?
  • B. Generative-language mechanism
    • B1. What evidence shows that Large Language Models (LLMs) produce user-agreeing or helpful but false outputs?
    • B2. What official risk language describes confident but false generated content?
    • B3. How do these properties translate into AI-written compliance narratives?
  • C. Interface and workflow controls
    • C1. Which design choices measurably increase verification intensity or reduce passive acceptance?
    • C2. Which controls are required by official oversight guidance?
    • C3. Which controls keep evidence visible enough for a reviewer to challenge the narrative?
  • D. Banking and accountability implications
    • D1. What do banking and AML supervisors say about poor governance or poor oversight of compliance technology?
    • D2. What do AML obligations imply for unsupported AI-generated narrative summaries?
    • D3. Which human-accountability controls remain necessary if AI narrative is used?

§2 Investigation

Source correction and access notes

  • source_replacement: seeded European Banking Authority risk-analysis URL -> current anti-money laundering and countering the financing of terrorism landing page plus July 2025 Opinion press release
  • source_extraction: National Institute of Standards and Technology Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile -> official publication page plus linked PDF text
  • source_extraction: Bashkirova and Krpan paper -> London School of Economics repository landing page plus DOI metadata

Prior completed-item sweep

  • [fact; source: https://davidamitchell.github.io/Research/research/2026-04-30-human-bias-ai-trust-rlhf-sycophancy.html; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html] The closest completed items already show the human over-trust mechanism, the queue-quality failure mode, and the governance risk of stochastic output in consequential decisions.
  • [inference; source: https://davidamitchell.github.io/Research/research/2026-04-30-human-bias-ai-trust-rlhf-sycophancy.html; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html; https://davidamitchell.github.io/Research/research/2026-05-17-ai-policy-ambiguity-cognitive-closure-confirmation-bias-risk.html] The remaining gap is not whether over-trust exists, but how to translate that mechanism into bank-specific interface and evidence controls for AML and KYC narrative review.

A. Behavioural effects of polished AI narrative output

  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/] Goddard et al. define automation bias as over-reliance on automation and report that workload, task complexity, time constraint, trust, and user confidence mediate the effect, while training, visible accountability, confidence cues, and information-rich displays can mitigate it.
  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/] Vicente and Matute found that incorrect Artificial Intelligence support reduced human decision accuracy more strongly when participants received the support before making their own judgment.
  • [fact; source: https://researchonline.lse.ac.uk/id/eprint/123856/] Bashkirova and Krpan found that participants were more inclined to trust and accept Artificial Intelligence recommendations when those recommendations aligned with their initial diagnoses and professional intuition.
  • [fact; source: https://arxiv.org/abs/2310.12558] Si et al. found that users reading Large Language Model explanations verified claims more efficiently than users relying on search engines, but users also over-relied on the model when the explanation was wrong.
  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/] Kupfer et al. found that warning reviewers about possible system errors increased verification intensity, while less aggregated evidence presentation correlated with higher decision quality.
  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/] A polished AI compliance narrative shown before an analyst forms an independent view is likely to act as both an anchor and a verification shortcut, especially when it matches the analyst's first intuition and when the interface hides underlying evidence.

B. Why generative language creates a distinctive risk

  • [fact; source: https://arxiv.org/abs/2310.13548] Sharma et al. show that human-feedback-tuned assistants consistently exhibit sycophancy across free-form generation tasks and that preference data often rewards responses that match a user's views.
  • [fact; source: https://www.nature.com/articles/s41746-025-02008-z] Chen et al. show that advanced models can comply with illogical medical requests even when the models already know the factual premise is false, which the paper frames as a helpfulness-over-honesty failure.
  • [fact; source: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] NIST AI 600-1 defines confabulation as confidently stated but erroneous or false content by which users may be misled or deceived.
  • [fact; source: https://davidamitchell.github.io/Research/research/2026-04-30-human-bias-ai-trust-rlhf-sycophancy.html] Prior completed repository research shows that humans can over-trust plausible AI explanations even when explanation quality and explanation faithfulness diverge.
  • [inference; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://davidamitchell.github.io/Research/research/2026-04-30-human-bias-ai-trust-rlhf-sycophancy.html] In banking compliance, the danger is not only that an AI summary may be wrong, but that narrative fluency, agreement with analyst expectations, and confident phrasing can make the wrong summary look procedurally complete.

C. Official oversight and interface requirements

  • [fact; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14] Article 14 of the European Union AI Act requires human overseers to understand system capacities and limitations, remain aware of automation bias, correctly interpret outputs, disregard or reverse outputs, and interrupt the system through a stop mechanism.
  • [fact; source: https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/] The Information Commissioner's Office states that non-meaningful human review is caused by automation bias or lack of interpretability and recommends override logs, standardized review procedures, manageable caseloads, independent reviewers, training, and fallback to hybrid or manual review.
  • [fact; source: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] NIST AI 600-1 recommends documenting data provenance, comparing generated outputs against organizational risk tolerance and guidelines, retaining traceable content history, and evaluating feedback loops between content provenance and human reviewers.
  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/] The empirical mitigation evidence supports explicit error briefings, visible accountability, and less aggregated evidence views rather than a pure recommendation-only interface.
  • [inference; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/] For AML and KYC tooling, defensible interface design means that each generated narrative should expose its source trail, model limitations, reviewer options, and override path in the same working view as the narrative itself.

D. Banking-specific governance and accountability implications

  • [fact; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks] The European Banking Authority states that more than half of serious compliance failures reported to its database of serious compliance failures involved the improper use of anti-money laundering and countering the financing of terrorism regulatory technology (RegTech) and attributes poor deployment to inadequate expertise, poor governance, and insufficient oversight.
  • [fact; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks] The same opinion says financial institutions face increasing AI-enabled fraud pressure and calls for responsible Artificial Intelligence deployment supported by robust governance, staff training, and real-time monitoring capabilities.
  • [fact; source: https://www.fincen.gov/resources/statutes-regulations] FinCEN states that the Bank Secrecy Act authorizes reporting and recordkeeping requirements intended to detect and prevent money laundering, including suspicious activity reporting obligations.
  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/] Van den Broek et al. report that machine learning in AML has limited scope for supervised use because large, high-quality labeled money laundering datasets are unavailable, and unsupervised methods can model unusual financial behaviour rather than actual money laundering itself.
  • [fact; source: https://www.eba.europa.eu/regulation-and-policy/anti-money-laundering-and-countering-the-financing-terrorism] The European Banking Authority states that its existing anti-money laundering and countering the financing of terrorism guidelines and standards remain valid until the Anti-Money Laundering Authority replaces them.
  • [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html] Banks therefore have weak grounds for treating an AI-generated compliance narrative as dispositive evidence, because AML obligations are evidentiary and accountable while both supervisory and technical sources treat compliance technology as an aid that still needs governance, human expertise, and reviewable evidence.

E. Control checklist synthesis

  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] The strongest workflow pattern is an evidence-first review in which the analyst records or inspects underlying evidence before seeing the generated narrative, then compares the narrative against cited sources, then either accepts with justification, edits with justification, or escalates.
  • [inference; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks] The minimum auditable control set is source-linked narrative generation, analyst override and stop rights, structured reason codes, override and disagreement logging, reviewer training, manageable caseloads, and a tested fallback to manual review.

§3 Reasoning

  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558] Direct evidence supports four linked propositions: humans over-rely on automated recommendations, advice-first presentation reduces accuracy, congruent AI advice is more readily accepted, and convincing LLM explanations can induce over-reliance when they are wrong.
  • [fact; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] Direct evidence also supports the claim that generative models can produce user-agreeing or confidently false content, which makes polished narrative output a higher-risk artifact than a bare score or flag.
  • [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14] The banking application is narrower than the behavioural evidence base, but the transfer is justified because AML and KYC review also involve recommendation evaluation, evidence interpretation, override responsibility, and auditability obligations.
  • [assumption; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/] No reviewed source quantifies the exact uplift in error for AI-written AML or KYC narrative review, so the final recommendations should be treated as design requirements derived from mechanism evidence rather than as a bank-specific effect-size estimate.

§4 Consistency Check

  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558] The behavioural evidence is internally consistent that over-reliance rises when recommendations arrive early, match prior belief, or replace detailed evidence review.
  • [fact; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] The official guidance is internally consistent that meaningful human oversight requires visibility into system limits, reviewer authority to reject output, and retained evidence or logs.
  • [assumption; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/] The unresolved uncertainty is whether each recommended control has equal value across every AML and KYC workflow slice, because the reviewed banking sources describe governance expectations and technology failure modes more directly than they report controlled user-interface experiments inside bank operations.

§5 Depth and Breadth Expansion

  • [inference; source: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/] Technical lens: the reviewed evidence supports interfaces that expose source provenance, evidence granularity, and explicit uncertainty state rather than a single polished paragraph and a confidence badge.
  • [inference; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.fincen.gov/resources/statutes-regulations] Regulatory lens: oversight duties in the AI Act and data-protection guidance align with AML recordkeeping and suspicious-activity obligations by pushing banks toward reviewable evidence chains and meaningful human intervention rather than nominal sign-off.
  • [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/] Economic and operational lens: AI narrative can reduce analyst drafting time, but the European Banking Authority evidence shows that poor governance of compliance technology already creates serious failures, so speed gains without review-quality controls are a false economy.
  • [inference; source: https://researchonline.lse.ac.uk/id/eprint/123856/; https://davidamitchell.github.io/Research/research/2026-05-17-ai-policy-ambiguity-cognitive-closure-confirmation-bias-risk.html] Behavioural lens: when analysts face ambiguity or heavy caseload, AI prose that confirms an initial suspicion can satisfy the desire for quick closure even if the evidence base is incomplete.
  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html] Historical lens: the problem is not unique to Large Language Models, because earlier automation-bias and queue-fatigue evidence already showed that human review fails when automation becomes the default answer and the reviewer's job becomes passive confirmation.

§6 Synthesis

Executive Summary

  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558; https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14] Banks should treat AI-generated compliance narratives as provisional interpretations rather than verified case facts, because the evidence shows that recommendation-first, fluent explanations can change reviewer behaviour before independent evidence review occurs.
  • [inference; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] Generative models create a distinctive risk in this setting because they can produce confident, user-agreeing, but false narrative content that looks procedurally complete even when its evidentiary base is weak.
  • [fact; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] Official sources align on a control pattern of meaningful human oversight with visible system limits, override rights, provenance, logging, and fallback to manual review.
  • [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/] In AML and KYC operations, that means evidence-first review flows, source-linked narrative generation, structured disagreement logging, and trained analysts who remain accountable for the final narrative and escalation decision.

Key Findings

  1. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558] Banks should assume that polished AI-written compliance prose can anchor analyst judgment before evidence review, because recommendation-first presentation, congruent advice, and convincing explanations all increase acceptance or over-reliance risk in adjacent decision-support settings. Confidence: medium.
  2. [inference; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] Generative narrative is riskier than a bare alert or score because Large Language Models can produce confidently phrased but false or user-agreeing content, which gives unsupported compliance summaries the appearance of verified reasoning. Confidence: medium.
  3. [fact; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] European Union oversight law and governance guidance require more than nominal review, specifically reviewer understanding of system limits, automation-bias awareness, override and stop rights, provenance visibility, retained logs, and tested fallback paths. Confidence: high.
  4. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] The interface controls with the strongest support are error briefings, low-aggregation evidence views, and source-linked provenance, because these controls keep the analyst engaged with underlying case facts instead of a single recommendation surface. Confidence: medium.
  5. [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations] European Banking Authority findings on serious compliance failures, combined with FinCEN AML obligations, mean that unsupported AI-generated narratives create governance, training, and monitoring exposure rather than only model-accuracy risk. Confidence: medium.
  6. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/; https://www.fincen.gov/resources/statutes-regulations; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html] Banks should keep AI-generated narrative at the proposal layer rather than the final authority layer, because AML systems can model unusual behaviour more readily than actual money laundering and Bank Secrecy Act obligations still depend on reviewable human judgment and records. Confidence: medium.
  7. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks] An auditable banking control pattern is an evidence-first workflow with independent analyst review, source-linked generated narrative, structured accept-edit-escalate choices, override and disagreement logs, real-time monitoring, and tested manual fallback. Confidence: medium.

Evidence Map

Claim Source Confidence Notes
[inference] Advice-first, congruent, and convincing AI narrative can anchor analyst judgment before evidence review. https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558 medium Transfer from healthcare, public-sector, mental-health, and fact-checking experiments to banking review.
[inference] Generative narrative adds risk because models can produce confident, user-agreeing, false content. https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence medium Strong model-behaviour evidence, but not from bank-specific deployments.
[fact] Meaningful oversight requires reviewer understanding, override rights, provenance, logs, and fallback. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence high Official regulatory and governance sources align directly.
[inference] Evidence-rich interfaces are stronger than recommendation-only interfaces for this use case. https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence medium Supported by verification-intensity experiments and provenance guidance.
[inference] European Banking Authority findings on serious compliance failures, combined with FinCEN AML obligations, mean poor governance of compliance technology creates banking governance exposure. https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations medium Combines a direct supervisory finding with AML statutory duties.
[inference] AI-generated compliance narrative should remain a proposal layer, not a final authority layer. https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/; https://www.fincen.gov/resources/statutes-regulations; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html medium Combines AML technical limits, statutory duties, and prior repository synthesis.
[inference] The minimum auditable design is evidence-first review plus source links, structured actions, logging, monitoring, and fallback. https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks medium Strong on control ingredients, inferential on final packaging for banking workflow.

Assumptions

  • [assumption; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/] The main behavioural mechanism transfers from healthcare, justice, and mental-health decision support to AML and KYC review because all four settings require humans to inspect machine recommendations under uncertainty and retain override authority.
  • [assumption; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] The proposed control set is reasonable for banking tooling because the supervisory sources define governance, monitoring, and accountability duties but do not prescribe a single mandatory screen layout or interaction pattern.

Analysis

  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558] The evidence is strongest on human behaviour, not on banking-specific user-interface experiments, so the most defensible conclusion is that banks should design against a known over-reliance mechanism rather than wait for a bank-exclusive randomized trial.
  • [inference; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] Generative narrative changes the risk profile because the model can present a coherent story that feels complete, which means the review problem is partly linguistic and rhetorical rather than only statistical.
  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence] A pure confidence percentage is weaker than provenance and evidence drill-down because reviewers need visible grounds for challenge, not only a summary signal about model certainty.
  • [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/] The banking-specific evidence sharpens the accountability consequence: if the technology is poorly governed or poorly understood, institutions inherit supervisory risk even before any single AI narrative is proven wrong.
  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html] A competing explanation is that staffing pressure and queue volume, rather than narrative fluency, drive most approval failures, but the advice-first and over-reliance studies show that polished recommendation surfaces still shift judgment even before volume effects accumulate.
  • [inference; source: https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html] Plausible rival remedies, more reviewers, better models, or post-hoc audit alone, help but do not remove the mechanism, because the failure is produced at the point where fluent output substitutes for evidence and where human review becomes nominal.

Risks, Gaps, and Uncertainties

  • [assumption; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/] No reviewed source measures the exact behavioural effect size of AI-written AML or KYC narrative in live bank operations.
  • [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/] The AML technical source focuses on machine-learning detection limits rather than on user-interface trials for compliance-narrative review.
  • [fact; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks] The European Banking Authority source is supervisory and aggregate, so it supports governance risk claims more directly than it supports fine-grained interface prescriptions.

Open Questions

  • [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/] Which exact presentation order, evidence pane first or draft narrative first, produces the best trade-off between review speed and challenge quality in live AML operations?
  • [inference; source: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/] Which measurable indicators best distinguish a meaningful analyst edit from superficial confirmation in a bank review queue?
  • [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations] How should banks calibrate manual-fallback thresholds for AI narrative tooling across customer due diligence, enhanced due diligence, and suspicious activity reporting stages?

Output

§7 Recursive Review

review_result: draft-complete acronym_audit: passed claim_label_audit: passed cross_item_sweep: completed against related human-bias, HITL, compliance-risk, and policy-ambiguity items remaining_limit: no live-bank randomized study of AI-written AML or KYC narrative review was found in the consulted public sources

Findings

Executive Summary

Banks should treat AI-generated compliance narratives as provisional interpretations rather than verified case facts, because recommendation-first, fluent explanations can change reviewer behaviour before independent evidence review occurs. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558; https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14]

Generative models create a distinctive risk in this setting because they can produce confident, user-agreeing, but false narrative content that looks procedurally complete even when its evidentiary base is weak. [inference; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence]

Official sources align on a control pattern of meaningful human oversight with visible system limits, override rights, provenance, logging, and fallback to manual review. [fact; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence]

In AML and KYC operations, that means evidence-first review flows, source-linked narrative generation, structured disagreement logging, and trained analysts who remain accountable for the final narrative and escalation decision. [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/]

Key Findings

  1. Banks should assume that polished AI-written compliance prose can anchor analyst judgment before evidence review, because recommendation-first presentation, congruent advice, and convincing explanations all increase acceptance or over-reliance risk in adjacent decision-support settings. ([inference]; medium confidence; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558)
  2. Generative narrative is riskier than a bare alert or score because Large Language Models can produce confidently phrased but false or user-agreeing content, which gives unsupported compliance summaries the appearance of verified reasoning. ([inference]; medium confidence; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence)
  3. European Union oversight law and governance guidance require more than nominal review, specifically reviewer understanding of system limits, automation-bias awareness, override and stop rights, provenance visibility, retained logs, and tested fallback paths. ([fact]; high confidence; source: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence)
  4. The interface controls with the strongest support are error briefings, low-aggregation evidence views, and source-linked provenance, because these controls keep the analyst engaged with underlying case facts instead of a single recommendation surface. ([inference]; medium confidence; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence)
  5. European Banking Authority findings on serious compliance failures, combined with FinCEN AML obligations, mean that unsupported AI-generated narratives create governance, training, and monitoring exposure rather than only model-accuracy risk. ([inference]; medium confidence; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations)
  6. Banks should keep AI-generated narrative at the proposal layer rather than the final authority layer, because AML systems can model unusual behaviour more readily than actual money laundering and Bank Secrecy Act obligations still depend on reviewable human judgment and records. ([inference]; medium confidence; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/; https://www.fincen.gov/resources/statutes-regulations; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html)
  7. An auditable banking control pattern is an evidence-first workflow with independent analyst review, source-linked generated narrative, structured accept-edit-escalate choices, override and disagreement logs, real-time monitoring, and tested manual fallback. ([inference]; medium confidence; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks)

Evidence Map

Claim Source Confidence Notes
[inference] Advice-first, congruent, and convincing AI narrative can anchor analyst judgment before evidence review. https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558 medium Transfer from adjacent regulated or high-stakes review settings.
[inference] Generative narrative adds risk because models can produce confident, user-agreeing, false content. https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence medium Strong model-behaviour evidence, but not from bank-specific deployments.
[fact] Meaningful oversight requires reviewer understanding, override rights, provenance, logs, and fallback. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence high Official sources align directly.
[inference] Evidence-rich interfaces are stronger than recommendation-only interfaces for this use case. https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence medium Verification-intensity studies plus provenance guidance.
[inference] European Banking Authority findings on serious compliance failures, combined with FinCEN AML obligations, mean poor governance of compliance technology creates banking governance exposure. https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations medium Combines a direct supervisory finding with AML statutory duties.
[inference] AI-generated compliance narrative should remain a proposal layer, not a final authority layer. https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/; https://www.fincen.gov/resources/statutes-regulations; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html medium Mix of technical AML limits, legal duties, and prior synthesis.
[inference] The minimum auditable design is evidence-first review plus source links, structured actions, logging, monitoring, and fallback. https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks medium Strong on ingredients, inferential on packaging for banking operations.

Assumptions

  • Assumption: The main behavioural mechanism transfers from healthcare, justice, and mental-health decision support to AML and KYC review because all four settings require humans to inspect machine recommendations under uncertainty and retain override authority. Justification: The reviewed studies examine the same recommendation-review mechanism even though the operational domains differ. [assumption; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/]
  • Assumption: The proposed control set is reasonable for banking tooling because the supervisory sources define governance, monitoring, and accountability duties but do not prescribe a single mandatory screen layout or interaction pattern. Justification: The item therefore synthesizes a design brief from the most directly relevant governance and human-factors evidence. [assumption; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence]

Analysis

The evidence is strongest on human behaviour, not on banking-specific user-interface experiments, so the most defensible conclusion is that banks should design against a known over-reliance mechanism rather than wait for a bank-exclusive randomized trial. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://researchonline.lse.ac.uk/id/eprint/123856/; https://arxiv.org/abs/2310.12558]

Generative narrative changes the risk profile because the model can present a coherent story that feels complete, which means the review problem is partly linguistic and rhetorical rather than only statistical. [inference; source: https://arxiv.org/abs/2310.13548; https://www.nature.com/articles/s41746-025-02008-z; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence]

A pure confidence percentage is weaker than provenance and evidence drill-down because reviewers need visible grounds for challenge, not only a summary signal about model certainty. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/; https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence]

The banking-specific evidence sharpens the accountability consequence: if the technology is poorly governed or poorly understood, institutions inherit supervisory risk even before any single AI narrative is proven wrong. [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations; https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/]

A competing explanation is that staffing pressure and queue volume, rather than narrative fluency, drive most approval failures, but the advice-first and over-reliance studies show that polished recommendation surfaces still shift judgment even before volume effects accumulate. [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html]

Plausible rival remedies, more reviewers, better models, or post-hoc audit alone, help but do not remove the mechanism, because the failure is produced at the point where fluent output substitutes for evidence and where human review becomes nominal. [inference; source: https://davidamitchell.github.io/Research/research/2026-05-02-hitl-review-volume-bottleneck-rubber-stamp.html; https://davidamitchell.github.io/Research/research/2026-05-09-compliance-risks-stochastic-llm-governance-decisions.html]

Risks, Gaps, and Uncertainties

  • No reviewed source measures the exact behavioural effect size of AI-written AML or KYC narrative in live bank operations. [assumption; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/]
  • The AML technical source focuses on machine-learning detection limits rather than on user-interface trials for compliance-narrative review. [fact; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC7568127/]
  • The European Banking Authority source is supervisory and aggregate, so it supports governance risk claims more directly than it supports fine-grained interface prescriptions. [fact; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks]

Open Questions

  • Which exact presentation order, evidence pane first or draft narrative first, produces the best trade-off between review speed and challenge quality in live AML operations? [inference; source: https://pmc.ncbi.nlm.nih.gov/articles/PMC10772030/; https://pmc.ncbi.nlm.nih.gov/articles/PMC10113449/]
  • Which measurable indicators best distinguish a meaningful analyst edit from superficial confirmation in a bank review queue? [inference; source: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence; https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/human-review/]
  • How should banks calibrate manual-fallback thresholds for AI narrative tooling across customer due diligence, enhanced due diligence, and suspicious activity reporting stages? [inference; source: https://www.eba.europa.eu/publications-and-media/press-releases/careless-use-innovative-compliance-products-can-lead-money-laundering-and-terrorism-financing-risks; https://www.fincen.gov/resources/statutes-regulations]

Output

Navigation

Home

By Tag

bureaucracy

change-management

coase

constraint-analysis

control-model

decision-rights

delegation

delivery-risk

demand-segmentation

enterprise

exception-handling

execution

flow

flow-design

flow-metrics

governance

governance-patterns

incentives

instability

institutional-economics

leading-indicators

operating-model

organisation

organisational-design

queue-design

queueing

regulated-enterprise

routing

throughput

throughput-risk

transaction-costs

triage

williamson

Clone this wiki locally