The Consensus That Never Happened
A board pack arrives. Twelve pages have become three. The summary is clear, the tone is confident, and the decision path looks obvious. There is just one problem: the dissent that shaped the original discussion has vanished. The minority view that forced a critical concession is now a footnote, if it appears at all. The uncertainty that led to a conditional approval has been rendered as full endorsement. The summary is not wrong in the sense of containing factual errors—it is wrong because it has flattened the narrative into a shape that never existed in the room.
This is narrative flattening, and it is the governance risk that most AI summary workflows are not designed to detect, let alone prevent. The risk is not hallucination in the traditional sense—fabricated facts or invented sources. It is the systematic smoothing of complexity, the editorial bias towards coherence, and the loss of decision-useful texture that occurs when generative models optimise for readability rather than fidelity. For DACH mid-market firms deploying AI summarisation in board reporting, works council documentation, HR case files, customer complaints, and compliance records, the consequences are not abstract. They are fiduciary, legal, and reputational.
The challenge is that narrative flattening does not announce itself. It does not trigger an error message. It produces output that looks professional, sounds authoritative, and meets the brief—until someone asks to see the source, or until a regulator questions why a contested decision appears uncontested in the record, or until an employee tribunal reveals that the summary omitted the very uncertainty that would have changed the outcome. By then, the governance failure is not in the model. It is in the workflow that trusted the model without anchoring it to source integrity, dissent preservation, and review rules.
Why Summaries Smooth Away What Matters Most
Most generative models used for summarization are autoregressive LLMs trained to predict the next token based on patterns in their training data. While this creates fluent text, it can compress competing viewpoints into a single narrative arc, resolve ambiguity in favour of the majority position, and edit out the procedural hesitations that signal uncertainty. Production systems often layer additional controls—fine-tuning, instruction prompts, or retrieval-augmented pipelines—but the underlying tendency towards consensus, clarity, and coherence remains. This is not a flaw in the model—it is a feature of the statistical objective. The model is doing exactly what it was trained to do: produce fluent, readable text that feels like a summary.
The problem is that governance documents are not optimised for fluency. They are optimised for accountability. A board minute that records a split vote, a conditional approval, or a decision deferred pending further analysis is not poorly written—it is precisely written. It preserves the texture of the decision process because that texture is what makes the record defensible. When a summary flattens that texture, it does not just lose detail. It loses the evidence that the board exercised due care, that management disclosed material risks, and that dissenting directors had their objections recorded. In a fiduciary context, that loss is not editorial—it is substantive.
Works council minutes present the same risk in a different register. German labour law requires that employee representatives' objections, alternative proposals, and procedural concerns be documented. A summary that renders a contested restructuring plan as "approved after discussion" may be technically accurate—there was discussion, and there was approval—but it obscures the fact that the discussion included formal dissent, that the approval was conditional, and that the works council reserved the right to revisit the decision if certain conditions were not met. When that summary becomes the record, the legal protection that the original minutes provided evaporates. The council's ability to demonstrate that it raised concerns in good time is compromised, and the employer's ability to show that it consulted in good faith is weakened. Both parties lose, and the model that created the summary has no way of knowing.
HR case files amplify the problem further. A performance improvement plan that documents an employee's initial resistance, subsequent engagement, and partial success is a record of process. A summary that describes the employee as "underperforming" or "non-compliant" may be drawing on language in the file, but it is also making an editorial choice about which moments in the timeline matter most. If that summary becomes the basis for a termination decision, and if the employee later challenges the dismissal, the tribunal will ask to see the source. If the source shows a more nuanced picture than the summary conveyed, the employer's case is weakened—not because the summary lied, but because it flattened the narrative in a way that removed the evidence of procedural fairness.
The Governance Gap in Summary Workflows
Most organisations deploying AI summarisation today are not treating it as a governance-critical workflow. They are treating it as a productivity tool. The model is embedded in a document management system, a collaboration platform, or a custom application. Users paste in a long document, the model returns a short summary, and the summary is filed, shared, or presented. There is no systematic review of whether the summary preserved dissent, no audit trail linking summary claims back to source paragraphs, and no policy governing when a summary is sufficient and when the full source must be retained and reviewed.
This is not because organisations are careless. It is because the risk is not yet visible. Narrative flattening does not show up in a penetration test. It does not trigger a data breach notification. It does not appear in a compliance dashboard. It emerges slowly, in the gap between what the summary says and what the source said, and it becomes visible only when someone with standing—a regulator, a tribunal, an auditor, a journalist—asks to see the original and compares it to the record.
The Financial Stability Board's AI governance framework for financial institutions makes the point explicitly: boards and senior management are responsible for aligning AI adoption with risk appetite and ensuring that AI lifecycle management includes explainability, performance monitoring, and human oversight. The framework does not single out summarisation, but the principle applies directly. If a bank's board pack is summarised by a generative model, and if that summary omits a risk disclosure that appeared in the full report, the board's ability to demonstrate that it was adequately informed is compromised. The model did not lie—it simply optimised for brevity in a way that removed the very information the board needed to exercise its fiduciary duty.
Industry surveys indicate growing C-suite concern about technological risk, with AI governance maturity often limited to access controls, usage policies, and vendor risk assessments. It rarely extends to workflow-level controls that ensure decision-useful fidelity in AI-generated outputs.
What Source Anchoring and Dissent Preservation Look Like in Practice
A governance-grade summary workflow does not eliminate AI summarisation. It anchors it. Every summary claim is linked back to the source paragraph or section that supports it. If the summary states that the board approved a proposal, the link shows whether the approval was unanimous, conditional, or accompanied by dissent. If the summary describes a customer complaint as resolved, the link shows whether the resolution was accepted by the customer, contested, or still under review. The summary remains short, but the accountability remains intact.
Dissent preservation is a specific design requirement. If the source document contains a minority view, a procedural objection, or a reservation, the summary must either include it or flag its omission. This is not a natural behaviour for a generative model—models are trained to synthesise, not to highlight divergence—so it must be engineered into the workflow. One illustrative approach uses a two-stage process with dissent-marker detection: the model generates a draft summary, and a second pass checks the source for dissent markers (phrases like "however," "in contrast," "the minority view," "subject to," "conditional on") and flags any that do not appear in the draft. However, keyword-based dissent detection is illustrative only and production systems require semantic models, context-aware classifiers, or human-in-the-loop validation to avoid false positives and negatives. If dissent is detected and omitted, the workflow either auto-inserts a reference or routes the summary to human review before it is finalised.
Review rules are the third pillar. Not every summary requires the same level of scrutiny. A summary of a routine operational update may be low-risk. A summary of a board decision, a works council negotiation, a disciplinary case, or a regulatory filing is high-risk. The workflow should classify documents by governance sensitivity and apply review rules accordingly. High-risk summaries are reviewed by a human with domain authority before they are filed or shared. The reviewer's role is not to rewrite the summary—it is to verify that the summary preserves the decision-useful texture of the source, that dissent is recorded, and that uncertainty is not rendered as certainty.
Audit trails close the loop. Every summary is versioned, and every version is linked to the source document, the model that generated it, the prompt that was used, and the reviewer who approved it (if applicable). If a summary is later challenged, the audit trail makes it possible to reconstruct the workflow, identify where fidelity was lost, and demonstrate whether the loss was a model limitation, a prompt design flaw, or a review failure. This is not paranoia—it is the same discipline that organisations already apply to financial reporting, clinical documentation, and legal discovery. The fact that the output is generated by a model rather than typed by a human does not reduce the governance standard; it raises it.
The Board's Role in Setting the Standard
Narrative flattening will not be solved by the IT department. It will be solved by the board and senior management setting a clear expectation: AI-generated summaries of governance-critical documents must preserve dissent, anchor claims to sources, and be subject to review rules that match the document's fiduciary sensitivity. That expectation does not require the board to understand how the model works. It requires the board to understand what the organisation is at risk of losing if summaries are treated as editorial convenience rather than governance artefacts.
Governance experts emphasise that AI's impacts on strategy, talent, and risk make it essential for boards to adapt their oversight approaches and embed AI in governance. The point is not that boards should micromanage AI tools—it is that boards should ensure that AI adoption does not degrade the integrity of the decision-making record. If the board pack is summarised, the board should ask: who reviewed the summary, what rules governed the review, and how do we know that dissent was preserved? If the answer is that no one reviewed it and no rules exist, the governance gap is visible.
The challenge for DACH mid-market firms is that narrative flattening is not yet on the risk register. Shadow AI adoption, data leakage, and vendor lock-in are visible risks, and they are being addressed—sometimes well, sometimes poorly, but they are being addressed. Narrative flattening is invisible until it is too late, and by the time it becomes visible, the damage is done. The works council minute that omitted dissent is already filed. The board pack that rendered a conditional approval as unconditional is already in the archive. The HR case summary that flattened a nuanced performance record into a binary judgment is already the basis for a termination decision. The governance failure is not that the model made a mistake—it is that the workflow allowed the model to make editorial choices without accountability.
A newsletter subscription delivers practitioner-grade analysis of AI governance risks like narrative flattening—before they become compliance failures.
References: Financial Stability Board AI governance framework; governance commentary on board oversight of AI strategy and risk.