AI and the Future of Human Decision-Making

A résumé-screening AI quietly learns and repeats a company's past hiring biases. A retail chatbot promises customers something the company never intended to deliver. A hospital's AI alert system speeds up treatment decisions but slowly dulls clinicians' own instinct for catching what the model wasn't trained to see. These aren't hypotheticals — Deloitte's latest Human Capital Trends research treats them as previews of what happens when AI gets folded into decision-making faster than organizations build the oversight to match it.

The gap between adoption and readiness

The numbers Deloitte cites are stark: in a 2023 Oracle study, the overwhelming majority of business leaders said they regretted or questioned past decisions, and most said data overload and distrust in their own data had stopped them from deciding anything at all. AI looks like a natural fix — 60% of executives in Deloitte's 2026 survey already use it to support decisions regularly, and Gartner projects that within a couple of years, half of all business decisions will involve AI in some augmented or automated capacity.

But adoption is outpacing governance. Deloitte's survey found that while most respondents see AI's role in decision-making as critical to near-term success, only a small sliver consider themselves genuinely ahead of the curve on managing it well.

Why AI can make bad decision-making worse, not better

The report is candid about a counterintuitive risk: AI often doesn't fix flawed decision processes, it amplifies them. Separate Deloitte research on decision intelligence found that a majority of organizations operate with low decision-making maturity — few teach decision-making as a skill or give teams real tools to support it. Layering AI on top of that kind of gap doesn't create better decisions; it just makes bad ones happen faster and at greater scale.

Underlying risks the report flags include: unclear chains of responsibility once "black box" algorithms are involved, people feeling less ownership over decisions AI helped make (and, per cited research, becoming somewhat more willing to act dishonestly when delegating a decision to AI), and managers who were never trained to supervise a machine collaborator the way they'd supervise a person.

Treat decision-making itself as a discipline

Deloitte's central recommendation is to stop treating decisions as informal by-products of meetings and dashboards, and instead design them deliberately. It points to Amazon's well-known distinction between "one-way door" decisions (hard or impossible to reverse, warranting real scrutiny) and "two-way door" decisions (easily reversible, safe to move on quickly) as a simple model for calibrating how much process a given decision actually deserves.

Two supporting practices stand out: surfacing which decisions actually matter enough to warrant explicit ownership and structure, and defining upfront what counts as good enough evidence before AI gets involved — rather than reverse-engineering justification after the fact.

Decision rights need to become dynamic, not fixed

A recurring theme is that classic frameworks for assigning decision ownership (like the traditional RACI model) assume authority is static — but AI keeps shifting who's effectively making the call. The report describes how Atlassian handled this: rather than writing a rigid rulebook for what AI handles versus what needs a human, the company treats the boundary as something to revisit regularly, which it credits with reducing bottlenecks and building more trust across teams.

At the board level, Deloitte's own research shows engagement lagging: in a 2024 survey, close to half of surveyed board members and executives said AI wasn't even on their meeting agenda. A follow-up in 2025 found a large share now admit AI is forcing them to rethink who even belongs on the board.

Skills and evaluation can't be an afterthought

The report highlights a kind of irony: many organizations are investing heavily in training AI to decide well, while assuming humans already know how. BAE Systems' response was a case-based leadership program that puts leaders through realistic, ambiguous scenarios to practice building and testing hypotheses under pressure — with early signs of better cross-team coordination and faster, more confident calls as a result.

Evaluating AI's own performance in decisions, the report argues, deserves the same rigor — not just accuracy metrics, but tracking whether human-AI interaction is actually building healthy judgment or quietly eroding it. Spotify is cited as an example of letting humans first define what "good" looks like (in its case, for podcast summaries) before building automated evaluation around that human-defined standard.

The real question isn't what the model says

One of the report's sharper framings, drawn from a Liberty Mutual Insurance example: once AI enters a workflow, the important question stops being what the model recommends, and becomes who has the standing — and the speed — to disagree with it. Liberty Mutual's claims adjusters can use AI to explore scenarios but retain override authority, which the report frames as a template for keeping real human agency intact rather than symbolic.

Trust, the report argues, follows a predictable pattern: people extend it to AI that shows reliability, capability, transparency, and something like humanity — and they're generally comfortable letting AI play a bigger role in analytical, high-stakes domains (fraud detection, forecasting, drug discovery) while wanting little or no AI involvement in more personal, value-laden calls.

The takeaway

Deloitte's closing argument is that organizations face a real choice: treat decision-making as a discipline worth designing deliberately — with clear ownership, evolving decision rights, real skill-building, and genuine human agency — or risk exactly the outcome AI is supposed to prevent: opaque choices, diluted accountability, and human judgment quietly eroding at the moment it matters most.