Design Without Governance Is Decoration

McKinsey’s latest Re:think newsletter makes a claim that is directionally correct and structurally incomplete. Chris Smith, a partner in their design practice, argues that AI’s scaling problem is not a technology problem. It is a design problem. Organizations are stuck in pilot purgatory because they bolted chat boxes onto pre-AI workflows and expected transformation.

The diagnosis is half right. The prescription is where it breaks down.

The Four Principles That Need a Foundation

McKinsey identifies four design characteristics that AI experiences need: clarity, continuity, depth, and collaboration.

Clarity means systems should reveal how conclusions are reached and be transparent about uncertainty. Continuity means AI should recognize progress across users and tasks instead of treating every request as a blank slate. Depth means automating entire workflows rather than providing isolated answers. Collaboration means human-AI interaction that goes beyond correcting the system after the fact.

These are not wrong. They are incomplete. Each principle describes a desirable surface behavior while skipping the infrastructure required to produce it.

Clarity Without Audit Trails Is Theater

Telling a system to “reveal how conclusions are reached” sounds like a design decision. It is not. It is an engineering and governance requirement that demands specific infrastructure.

Clarity requires explainability systems that trace which data sources informed a decision, which model version produced the output, and what confidence level the system assigned. It requires audit logging that captures not just what the system decided but what alternatives it considered. It requires decision tracing that connects outputs to inputs across multi-step workflows.

None of this is a design problem. It is a governance infrastructure problem. You can design a beautiful explanation panel that shows nothing useful because the system behind it has no tracing capability. As we explored in Governance Is a UX Problem, Apple and Carnegie Mellon’s research found that when agents take actions without explanation, trust collapses within a single interaction. But the research also showed that the explanation must be substantive. Users do not want the appearance of transparency. They want the mechanism of accountability.

McKinsey’s framing treats clarity as a design choice. In practice, clarity is the visible output of an explainability governance system. Without the system, the design is decoration.

Continuity Without Data Governance Is Liability

The continuity principle is the most revealing gap in McKinsey’s framework.

“Building an understanding of what came before so it can anticipate what comes next” is a description of persistent organizational memory. It means the AI system retains context across interactions, accumulates knowledge about the organization, and uses that accumulated context to inform future decisions.

This raises questions that design cannot answer. What data gets persisted? Who controls access to the organizational memory? How is outdated or incorrect information purged? What happens when an employee leaves and their context remains in the system? Who audits what the AI “remembers” about your organization?

These are data governance questions. They intersect with privacy regulations, data retention policies, and information security requirements. In regulated industries, they are compliance obligations with legal consequences.

Designing for continuity without answering these questions is building a system that remembers everything, forgets nothing, and has no accountability for what it carries forward. That is not better design. That is an unmanaged data liability.

Depth Without Boundaries Is Automation Risk

Depth means “automating entire workflows rather than just providing answers.” McKinsey describes AI that draws on multiple data sources and connects multistep processes.

This is where design thinking without governance thinking becomes actively dangerous.

An AI system that automates an entire workflow needs permission boundaries at every step. Which systems can it access? What data can it read? What actions can it take? When does it escalate to a human? What happens when it encounters an edge case outside its training distribution?

Workflow automation without boundary governance is how organizations get unauthorized data access, unintended actions at scale, and cascading failures across connected systems. The deeper the automation, the more critical the governance. As we discussed in The Governance Gap in AI Output Quality, even simple AI outputs converge toward mediocrity without structured constraints. Entire automated workflows without governance constraints converge toward risk.

McKinsey’s three case studies illustrate this gap without acknowledging it. A marketing campaign tool with 75% adoption and a “more than 2 percent” sales boost. A sales rep tool adopted by 90% of users. A hotel management tool where “nearly all” users deployed it. No organization named. No methodology disclosed. No discussion of what governance made these implementations safe. The case studies describe outcomes. They reveal nothing about the infrastructure that prevented adverse outcomes.

Collaboration Without Decision Governance Is Chaos

McKinsey correctly notes that “human in the loop” is insufficient. The goal, they argue, is for humans and AI to interact through steering, revising, and debating to drive superior outcomes.

This is the most interesting of the four principles, and the one most obviously requiring governance infrastructure.

If humans and AI are collaborating on decisions, who has authority? When the human and the AI disagree, which decision stands? How is that decision recorded? Who is accountable for the outcome?

Microsoft’s Magentic-UI project demonstrated that structured collaboration surfaces improve agent performance by 71%. But the improvement came not from better design. It came from six specific governance mechanisms: co-planning, co-tasking, action approval, answer verification, memory, and multi-tasking. Each mechanism is a governance surface that defines boundaries, creates checkpoints, and establishes accountability.

Design without these mechanisms produces collaboration theater. The AI presents options. The human clicks approve. Nobody tracks whether the approval was informed or reflexive. Nobody measures whether the collaboration improved outcomes or just distributed blame.

The Real Diagnosis

McKinsey’s article accidentally proves a point it does not intend to make.

The pilot-to-production gap is real. Multiple independent sources confirm it. Deloitte’s Q4 2024 report found 68% of organizations moved fewer than 30% of gen AI experiments to production. BCG’s AI Radar 2025 found only 26% of companies generate significant financial returns from AI. The NBER’s February 2026 survey of 6,000 executives found that over 80% report zero measurable productivity gains.

But the cause is not primarily design. BCG’s own research found the top barriers are data quality, talent gaps, unclear ROI measurement, security concerns, and organizational resistance. “Poor user experience design” does not appear in the top barriers.

This matters because the diagnosis determines the treatment. If the problem is design, the solution is design consulting. If the problem is governance infrastructure, the solution is building the systems that make design principles implementable.

We explored this distinction in McKinsey Measured the Wrong Thing. That earlier McKinsey study surveyed 300 executives about perceived AI productivity gains. The executives reported 16-45% improvements. METR’s controlled trial found developers were 19% slower with AI but believed they were 24% faster. The perception-measurement gap was not noise. It was missing infrastructure.

The pattern repeats here. McKinsey identifies what good AI should look like. They skip what makes good AI possible. The four principles are the interface. The governance infrastructure is the system. You cannot ship the interface without building the system.

What Actually Gets Organizations Past Pilot

Organizations that scale AI successfully do not just design better experiences. They build four governance capabilities that make better experiences possible.

Explainability infrastructure: Not just a panel that shows reasoning, but systems that trace decisions to data sources, log model versions, capture confidence scores, and create audit trails. The design surfaces this information. The governance infrastructure produces it.

Context governance: Not just persistent memory, but policies that define what gets retained, who can access it, how it gets updated, and when it gets purged. The design provides continuity. The governance makes continuity safe.

Workflow boundaries: Not just end-to-end automation, but permission systems, escalation rules, rollback capabilities, and error handling for every step in the chain. The design creates depth. The governance prevents depth from becoming uncontrolled scope.

Decision accountability: Not just collaboration interfaces, but clear authority models, documented decision points, and measured outcomes. The design enables collaboration. The governance makes collaboration governable.

The Commercial Circularity

It is worth noting what McKinsey is selling here. McKinsey acquired LUNAR Design in 2012 and Veryday in 2015 to build their design consulting practice. Chris Smith writes from that practice. The article diagnoses a design problem and prescribes design solutions. The circularity is structural: McKinsey’s research identifies problems that require McKinsey’s services.

This does not make the observations wrong. Many consulting firms produce genuinely useful analysis that happens to align with their service offerings. But it means the framing is commercially motivated. When McKinsey says the problem is design, they are saying the problem is something they sell. When we say the problem is governance infrastructure, we are also saying the problem is something we build. Intellectual honesty requires acknowledging both frames.

The difference is in what the evidence supports. The pilot-to-production gap data points to governance, data, and organizational barriers more consistently than it points to design. McKinsey’s own prior research supports this: their 2025 report found that only “end-to-end implementations” show results, which is another way of saying that isolated design improvements without systemic infrastructure produce isolated results.

The Integration

Design and governance are not alternatives. They are layers.

Governance without design is the policy PDF in the shared drive that nobody reads. We have known this for years. The policy exists. The compliance exists. The behavior does not change because the governance is invisible.

Design without governance is the beautiful interface that shows nothing useful. The explanation panel with no tracing system behind it. The collaboration surface with no accountability model underneath. The continuous experience with no data governance defining what persists.

Organizations that break through pilot purgatory will build both layers simultaneously. They will design AI experiences that surface governance information. And they will build governance infrastructure that produces information worth surfacing.

McKinsey got the surface right. The foundation is the part they left out.

This analysis responds to Chris Smith’s “The value of rich and meaningful AI interactions” (McKinsey Re:think, March 2026), drawing on Apple/CMU’s IUI’26 agent UX taxonomy, Microsoft’s Magentic-UI research, BCG’s AI Radar 2025, and the NBER February 2026 executive survey.

Victorino Group builds the governance infrastructure that makes AI design principles implementable. Let’s talk.