The right severity, the wrong pattern: why that’s still a governance failure

This is the second in a series on design governance in high-stakes AI products. Read the first article here.

In my last article, I wrote about why error classification is a governance decision and introduced the severity model I built for our healthcare AI product: S0 through S3, each level encoding a different risk threshold and a different expectation of the user.

But classification is only half the work.

You can get the severity exactly right and still fail the user. Because the severity tells you how serious something is. It doesn’t tell you how to show it.

That’s where pattern selection comes in. And it’s where I see the second most common governance failure in AI products: not misclassifying errors, but mismatching them to the wrong interaction pattern.

When the right call gets the wrong pattern

Consider this scenario: a clinician is reviewing treatment recommendations and the system detects duplicate chemotherapy orders.

That’s an S0. Imminent risk. Potentially irreversible harm. Every part of the severity model points to one thing: stop everything, make the user acknowledge this before they do anything else.

Now imagine that alert surfaces as a toast notification in the bottom-right corner of the screen. Five seconds. Auto-dismisses. Gone.

The classification was correct. The pattern was wrong. And in a clinical environment, that mismatch isn’t a UX problem, it’s a liability transfer. The system did its job technically. The clinician missed it. Who absorbs the risk?

This is why pattern selection is a governance decision, not a design preference.

The Core Principle: disruption should match stakes

Every interaction pattern sits somewhere on a spectrum from highly disruptive to nearly invisible. A blocking modal stops everything. A toast barely registers. In between: persistent banners, inline alerts, contextual warnings.

The governance principle is simple: the level of disruption should match the level of stakes.

Over-disrupt and you create alarm fatigue: users start dismissing modals without reading them because everything feels urgent. Under-disrupt and critical information gets missed. Both outcomes shift accountability from the system to the human.

Getting this calibration right is the job.

Walking through the patterns

S0 (Critical)— Blocking modal

The most disruptive pattern in the toolkit, and the only appropriate one for S0. It centers on screen, traps focus, and requires explicit acknowledgment before the user can continue. No close “X.” No snooze. No auto-dismiss.

This is not about being aggressive with the user. It’s about the system taking responsibility for what it knows. If the AI has detected something that could cause irreversible harm, the interface has an obligation to make that impossible to miss. Anything less is a design choice that quietly hands risk to the clinician.

S1 (High/Error) — Prominent banner

The primary task is blocked, but the situation is recoverable. A prominent banner at the top of the viewport, persistent until resolved or dismissed, with an inline action (“Fix now” or “View details”) that doesn’t require navigation away from the current context.

A toast on creation is appropriate here to signal that something has appeared, but the banner does the actual work. It stays. It waits. It doesn’t let the user forget.

S1 also has an escalation path: if the condition worsens or the user hasn’t interacted within a defined threshold, it escalates to S0.

S2 (Medium/Warning) — Inline alert or non-blocking banner

The user can continue, but something is degraded. In healthcare AI, this is where low-confidence outputs live: the system generated a summary, but it’s flagging uncertainty. The user should know. They don’t need to stop.

Placement matters here: near the problem area, not necesarily at the top of the page. A banner for a low-confidence recommendation that appears at the top of the viewport feels often disproportionate. An inline indicator next to the specific output respects the user’s flow while keeping the information visible.

Snooze is appropriate at S2. Dismiss is appropriate. The user is in control, the system has done its job by surfacing the information.

S3 (Low/Info) — Toast

Informational only. No action required. Five to seven seconds, bottom-right, gone. The user should be able to pause it on hover if they want to read it, but it shouldn’t demand anything.

The Matrix

The Anti-patterns worth naming

A few specific mismatches I see repeatedly, and why they matter:

Modals for S2/S3. Every unnecessary modal desensitizes users to the ones that actually require action. Reserve blocking interactions for blocking situations.

Auto-closing S0 or S1. If the condition still exists, the alert should still exist. Auto-dismissing a persistent problem because the timer ran out doesn’t resolve anything, it just hides it.

Alerts without actions. Every alert should either tell the user what to do or tell them explicitly that no action is needed. An alert that surfaces a problem with no path forward just adds cognitive load without adding value.

Aggregation failure. If the same S2 alert fires fifteen times in an hour, don’t show fifteen banners. Collapse them. Show one, with a counter. Flooding the interface with repetition is its own form of alarm fatigue.

Classification gets you to the door. Pattern selection gets you through it.

The severity model tells you what something means. The pattern selection matrix tells you how to say it: how loudly, how persistently, how much of the user’s attention it’s entitled to demand.

Both are governance decisions. Both have consequences when they’re wrong.

In the next article in this series, I’ll go one layer deeper: color semantics and accessibility: how the visual language of your alerts either reinforces or undermines the severity model, and why accessible design in high-stakes AI isn’t a compliance checkbox. It’s a governance decision.

I’m a Senior AI Product Designer with 9+ years of experience, currently building AI systems for healthcare. I write about design governance, risk, and accountability in high-stakes products. Follow along, there’s more coming.