The deliberation engine

This is article 2 of 5 in The Broken Ballot, a series on AI, deliberation, and responsible corporate governance published every Thursday. Read article 1 here.

Before setting out what the architecture does, it is worth addressing an objection that attentive readers of Article 1 will have formed. If the argument is that the problem is not information — that investors already have access to scenario analyses, climate disclosures, and risk reports — then why does the proposed solution produce more information? Stakeholder maps, scenario sets, ethical trade-off matrices and post-vote audit trails are all information products. The answer is the distinction between information and structured engagement with information. Information that is generated but not required to be engaged with is noise. The innovation is not additional data generation: it is the mandatory sequencing of deliberative stages that forces genuine engagement with information that currently exists but is systematically not processed. This is why Stage 2 does not add scenarios to what companies already disclose — it requires institutional investors to engage with those scenarios before their vote is recorded. The stages are not information products; they are engagement requirements. That distinction is the architecture's core contribution.

Why the architecture must start with welfare, not value

If corporate governance exists to aggregate shareholder financial preferences and translate them efficiently into capital allocation decisions, then the AI framework is essentially an efficiency tool: a better mechanism for producing the same kind of output. But that is not the premise on which this series proceeds.

Oliver Hart and Luigi Zingales have argued that the appropriate objective of corporate governance is not shareholder value maximisation but shareholder welfare maximisation: the proposition that shareholders, as human beings with moral commitments as well as financial interests, may legitimately care about the social and environmental consequences of corporate decisions alongside their effect on share prices. Consider their own illustration: shareholders in a tobacco company might genuinely prefer lower profits if the company ceased marketing to teenagers. That preference is not irrational. It expresses a welfare choice (a human judgment about what kind of company they wish to own) that current governance architecture systematically suppresses in favour of a narrower financial calculus.

If what shareholders actually want to optimise for includes welfare considerations, then the AI governance architecture must be designed to surface and structure those preferences. Not to replace them with algorithmic outputs. The six-stage framework that follows is built on that premise. It is normative in design and neutral in outcome: it enforces the right questions without dictating the answers.

What the NEDM literature tells us

A 2026 review of 28 business-oriented Normative Ethical Decision-Making Models reveals a consistent and striking finding: most existing models prescribe process rather than outcomes. They structure how decisions should be made, not what decisions should be made. This apparently modest insight carries profound practical implications for governance design.

The review also identifies three structural limitations that run through the majority of existing NEDMs. First, incompleteness: most models systematically omit key decision stages, particularly reflection, option generation, and stakeholder balancing, the stages most critical to avoiding premature closure and groupthink. Second, single-path linearity: most models follow a fixed sequential structure that is adequate for routine choices but structurally insufficient for complex, contested decisions where material risk thresholds change the appropriate deliberative process itself. Third, weak pluralist ethical integration: most models incorporate a dominant ethical framework, typically utilitarian or deontological, without systematically integrating competing moral principles, producing a hidden normative hierarchy that appears neutral while privileging a particular ethical tradition.

These limitations map directly onto the structural weaknesses of shareholder governance identified in the first article in this series. The AI architecture proposed here addresses each limitation systematically — extending the NEDM literature rather than merely applying it.

The six-stage architecture

Stage 1: Automated stakeholder mapping

Before the vote opens, the AI system ingests available public and disclosed information and generates a structured stakeholder matrix. The system identifies all affected parties across the full value chain, direct shareholders and bondholders, proximate communities, legally cognisable third parties with active legal standing, and systemic actors such as insurance markets pricing stranded-asset risk. Silence on any material stakeholder tier is not permitted.

Stage 2: Scenario generation

The system produces probabilistic financial, regulatory, and climate scenarios with explicit assumption disclosure. Critically, the AI flags where the board's own demand assumptions diverge from independent analyst consensus, making the optimism bias that characterises most management projections visible and documentable. Where material divergence exists, the board must provide a written justification before the vote proceeds. This stage directly addresses the NEDM review's finding that option generation is systematically omitted from existing models.

Stage 3: Ethical trade-off display

Rather than allowing boards to present a single normative framing of a resolution, the system generates a structured trade-off matrix across four normative lenses simultaneously: utilitarian aggregate welfare analysis, rights-based procedural assessment, distributional justice analysis, and intergenerational impact evaluation. Investors cannot proceed to vote until they have reviewed all four lenses. The system does not select between these frameworks, it forces visibility.

"The AI framework is normative in process; neutral in outcome. It enforces the right questions without dictating the answers."

Stage 4: Bias detection and deliberative override

Drawing on dual-process theory, the system distinguishes between intuitive, fast-response decision-making and deliberative, structured reasoning. Current shareholder voting predominantly engages the former. The bias detection module flags herding around proxy advisory recommendations, authority bias in boards with founder concentration, and short-termism in institutional portfolios. A mandatory devil's advocate summary presents the strongest case against the position currently polling as majority, a systematic counterweight to groupthink dynamics that the NEDM literature identifies as endemic to high-status governance settings.

Stage 5: Branching logic gates

The most structurally significant contribution of the proposed framework is its extension of the NEDM model into multi-path conditional logic. Rather than applying a uniform deliberative sequence to every resolution, the system introduces conditional branching: if a litigation materiality threshold is crossed, additional independent analysis is required before the vote proceeds; if full value chain emissions are undisclosed, a climate disclosure gate is triggered; if the regulatory foundation of the resolution is itself under legal challenge, investors must affirmatively acknowledge that risk in their voting record. This converts the shareholder assembly from a single-path process into a dynamic, conditional governance instrument.

A note on the gates themselves, since they represent the most interventionist element of the framework and the most likely target for objection. If AI prevents voting until conditions are met, a critic may argue, AI is already shaping outcomes. The response is that the framework constrains process, not substantive choice — in precisely the same way that procedural due process constrains courts without dictating verdicts. A court that requires both parties to be heard before judgment is delivered is not directing the outcome; it is ensuring the preconditions for legitimate judgment are met. Stage 5 logic gates operate identically: they do not require a particular vote. They require that a material risk threshold has been acknowledged before the vote is cast. The investor remains free to vote in any direction. The gate ensures that direction is a deliberate choice rather than an oversight.

Stage 6: Post-vote reflection report

The AI system generates an automated governance audit report recording how each institutional investor engaged with the scenario modelling; the aggregate probability weights assigned to material risk scenarios; any divergence between board assumptions and independent forecasts; and a structured audit trail of the deliberative process. In the evolving climate litigation landscape, this post-vote audit trail is not merely good governance practice. It is increasingly a precondition of fiduciary defensibility.

Adversarial realism: what happens when actors game the engine

A governance architecture that assumes good faith is not a governance architecture: it is a governance aspiration.

The most likely failure mode of the AI deliberative framework is not technical malfunction. It is instrumental compliance: a board that produces scenario assumptions calibrated to minimise apparent transition risk while formally satisfying Stage 2's disclosure requirements; an institutional investor that processes the devil's advocate summary as a procedural step rather than a substantive input. The architecture does not require good faith, it requires a record. The post-vote audit trail documents what was considered, how scenarios were weighted, which risks were acknowledged. Legible bad faith is more actionable than invisible bad faith. The architecture's value is not that it prevents gaming, it is that it makes gaming visible.

Process without outcome: the non-negotiable distinction

The AI governance architecture honours this distinction at every stage. The logic gates require additional engagement, they do not redirect the vote. The ethical trade-off display makes competing principles visible, it does not weigh them. The bias detection module flags herding patterns, it does not override the investor's ultimate judgment.

This is not a pragmatic compromise. It is the only design that is consistent with democratic corporate governance. The moment AI begins to select outcomes rather than structure the conditions for choosing them, it ceases to be a governance tool and becomes a governance authority, a substitution of algorithmic judgment for fiduciary responsibility.

The third article confronts what architecture cannot do. Even well-designed deliberative systems fail when board culture suppresses dissent, when AI embeds invisible normative assumptions, and when power asymmetries defeat genuine deliberation. We map Sir A. Likierman's six elements of judgment against the six-stage framework to produce an honest audit of both its strengths and its limits.