When AI Decisions Refuse to Close

The Edge · Leadership · Friday, 28 August 2026

Researched by an agentic pipeline · reviewed and gated by the author

Decision architecture is the machinery that turns specialist challenge into a decision that can close, be acted upon and later be revised.

The meeting that never quite ends

A business unit proposes an AI-assisted claims workflow. The economics are credible. The model performs well enough for a controlled release. The data team has resolved the most serious quality gaps, and operations has people ready to redesign the work.

Then the decision starts to circulate.

Technology wants a narrower integration boundary. Risk asks for a different evidence threshold. Legal will advise but will not own the residual decision. Operations believes the remaining uncertainty belongs with the business sponsor. The sponsor assumes the executive committee must decide. The executive committee asks the team to return with alignment.

No one has actually rejected the use case. No decisive technical obstacle has appeared. Yet three weeks later the same issue returns, accompanied by a larger slide deck and fewer clear choices.

This composite situation is recognisable because AI decisions cross functions that were designed to control different kinds of risk. Each function may be acting rationally. The failure lies in the architecture that should convert their separate judgements into a bounded enterprise decision.

That does not mean decision rights have replaced skills as the central AI bottleneck. Current evidence does not support that claim. Skills, data, infrastructure, regulation and business-case quality remain substantial independent constraints. [S1] [S2] The narrower and more useful conclusion is conditional: once those foundations are adequate, cross-functional decision architecture can become a distinct scaling bottleneck.

Diagnose the layer before redesigning authority

AI use is broad, but operational depth is uneven. Stanford’s 2026 AI Index reports widespread use in at least one business function while agent deployment remains in single digits across almost every function. [S3] That gap is real, but it does not identify a single cause.

The OECD’s evidence is an essential counterweight to fashionable leadership explanations. Its research continues to identify specialised skills as a major adoption barrier, particularly among non-adopting firms and smaller businesses. [S1] [S2] Deloitte’s 2026 enterprise survey similarly identifies worker skills as a leading obstacle to integrating AI into workflows, while organisations report weaker readiness in data, infrastructure, risk and talent than in strategy. [S5] These are mostly survey findings rather than causal proof, but they are too consistent to dismiss.

Leadership teams therefore need a layered diagnosis. Before changing committees, decision matrices or executive roles, test whether the use case has passed four thresholds:

• Economics: Is there a sufficiently credible value case, including implementation and operating cost?
• Technology: Does the system perform reliably enough for the proposed boundary and risk level?
• Data: Are access, quality, provenance and permitted use adequate?
• Capability: Can the workforce implement, operate, challenge and improve the new process?

If one of those thresholds fails, delay may be evidence of a weak proposition rather than leadership drift. A request for more data can be procrastination, but it can also be sound judgement when model performance is unstable. An escalation can reveal confused authority, but it can also expose a risk that genuinely belongs at a higher level.

Only after the foundations are credible should the organisation test the fifth layer: can the leadership system close the cross-functional decision?

What decision architecture actually does

Decision architecture is often reduced to a RACI chart or the appointment of a chief AI officer. Those devices can help, but they are not the operating mechanism.

Decision architecture is the machinery that turns specialist challenge into a decision that can close, be acted upon and later be revised.

It has five practical elements:

1. Bounded ownership. One role owns the outcome and the decision within an explicit boundary. Ownership includes consequences; it is not merely responsibility for convening meetings.
2. Defined authority. Each function knows whether it decides, advises, supplies evidence or holds a veto over a named class of harm.
3. Evidence thresholds. The team agrees in advance what is sufficient to proceed, pause or stop. “More assurance” is replaced by a testable condition.
4. Escalation rules. Escalation is triggered by the nature or magnitude of an unresolved trade-off, not by discomfort or lack of consensus.
5. Review and expiry. Reversible decisions have a review date, monitoring conditions and a route to correction. Temporary restrictions do not silently become permanent.

MIT CISR’s earlier transformation research found that changing decision rights is difficult: nearly two-thirds of represented firms were no better than moderately effective at it. Its proposed response was guardrails that allow teams to act while remaining aligned with enterprise interests. [S6] The lesson predates generative AI, which is precisely why it matters. AI has intensified an established organisational problem; it has not invented it.

Research on top-management-team structure reinforces the point that formal titles are only part of the system. Outcomes depend on how formal roles, informal influence and the overall bundle of responsibilities interact. [S7] A new executive role can clarify accountability, or it can add another participant to an already diffuse decision.

Three operating models

Most leadership drift in AI appears in one of three forms.

Diffuse consensus

Every affected function participates, and no one can close the decision without universal comfort. The apparent virtue is inclusion. The operating consequence is that advisory concerns behave like vetoes, evidence requirements expand during the process and reversible choices are treated as irreversible commitments.

The CEO bottleneck

Ambiguity is solved by pushing material decisions upward. This produces closure, but at the price of queueing, thin attention and weak local ownership. It can also suppress specialist challenge: executives learn that the safest route is to package a recommendation for approval rather than expose a genuine trade-off.

Bounded delegation

The business owner decides within a documented envelope. Technology, data, risk, legal and people functions have explicit roles. A limited set of conditions triggers escalation. Evidence thresholds and review dates allow the organisation to proceed without pretending uncertainty has disappeared.

Model	Closure mechanism	Typical failure
Diffuse consensus	Everyone aligns	Decisions cycle because advice becomes an implicit veto
CEO bottleneck	Senior authority decides	Throughput collapses and operating ownership weakens
Bounded delegation	Named owner acts within guardrails	Fails if boundaries, evidence or escalation rules are vague

Bounded delegation is not a recipe for speed at any cost. Its purpose is to make the relationship between authority and evidence visible. A high-impact, difficult-to-reverse deployment may still require board or executive approval. A reversible pilot within agreed limits should not wait for the same forum.

The strongest sceptical explanation

There is a credible reason to resist the leadership-drift story: many teams may appear indecisive because the underlying use case is not ready.

Model performance can be adequate in a demonstration and inadequate in a live process. Benefits can disappear once exception handling, supervision and integration are costed. Data access can remain contested. Regulatory expectations can be unsettled. Workers can lack the domain or technical capability to absorb the change. In those conditions, forcing closure through clearer decision rights may simply accelerate a bad investment.

Governance activity can also be mistaken for governance effectiveness. Stanford reports growth in AI-specific governance roles while knowledge, budget and regulatory uncertainty remain prominent obstacles. [S4] The creation of a council, policy or executive title is therefore weak evidence that the organisation can make better decisions.

This sceptical case changes the prescription. The objective is not to shorten every decision. It is to distinguish legitimate uncertainty from organisational drift. A disciplined team should be able to state which threshold has not been met, what evidence would meet it, who decides when it has been met and when the question will return. Drift begins when those answers remain implicit and the same unresolved issue changes owners without changing evidence.

Measure closure, not fluency

Boards often test whether management can explain an AI strategy, describe the technology and discuss responsible use. Executive fluency is useful, but it reveals little about the organisation’s ability to close decisions under uncertainty.

A better review examines the path from credible use case to production:

• Median time from approval to a named accountable owner.
• Decision latency for data, risk, workflow and workforce choices.
• Number of reversals and escalations per material decision.
• Share of decisions with explicit evidence thresholds and review dates.
• Production conversion after controlling for business-case quality, data readiness and technical maturity.
• Difference between the formal decision map and the people who actually influence outcomes.

The controls matter because raw speed can be misleading. A centralised organisation may decide quickly while creating brittle commitments. A participatory organisation may take longer initially but surface better evidence. The aim is neither unanimity nor unilateral command. It is a repeatable path by which expert disagreement becomes an accountable choice.

Leadership development should reflect that reality. Simulations should ask executives to separate reversible from irreversible choices, assign authority without silencing challenge, state evidence thresholds and design review conditions. The exercise is not to produce consensus. It is to reveal whether the team knows how a decision ends.

A test worth running now

Choose one AI use case that has credible economics, workable technology, adequate data and a capable operating team, yet has stalled across functions. Reconstruct the last three decision cycles.

Identify what new evidence appeared, which role believed it could decide, which concerns behaved as vetoes, why escalation occurred and whether any option had an expiry or review date. If each cycle produced meaningful evidence against deployment, the system may be working. If the same trade-off moved among forums without a change in evidence, decision architecture is the likely constraint.

The main uncertainty cannot be wished away: there is no robust cross-sector causal evidence separating leadership-team drift from skills, data, infrastructure, economics and regulation. The diagnosis must remain conditional. If outcomes improve when technical readiness or capability improves, without a change in decision rights, leadership drift was secondary.

But once the other layers are strong enough, continued drift deserves to be named. At that point the board’s most useful question is not whether executives understand AI. It is whether the organisation can show, for one consequential decision, who could close it, on what evidence, within what boundary and with what route to revision.

Sources

1. OECD — The Adoption of Artificial Intelligence in Firms — 2025 — https://www.oecd.org/en/publications/the-adoption-of-artificial-intelligence-in-firms_f9ef33c3-en.html
2. OECD — AI and Skills — 2026 — https://www.oecd.org/en/publications/ai-and-skills_f843b352-en/full-report.html
3. Stanford Institute for Human-Centered AI — 2026 AI Index Report: Economy — 2026 — https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
4. Stanford Institute for Human-Centered AI — 2026 AI Index Report: Responsible AI — 2026 — https://hai.stanford.edu/ai-index/2026-ai-index-report/responsible-ai
5. Deloitte — The State of AI in the Enterprise — 2026 — https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html
6. MIT CISR — Decision Rights Guardrails to Empower Teams and Drive Company Performance — 2020 — https://cisr.mit.edu/publication/2020_0801_DecisionRights_Meulen
7. Strategic Management Society — How Should Top Management Teams be Structured? — 2022 — https://www.strategicmanagement.net/publications-resources/strategic-management-explorer/how-should-top-management-teams-be-structured/