Case 02 / 04Governance · AI

AI rails, not AI drivers.

A 2×2 of appropriateness paired with severity controls. The framework a design org uses to decide where AI belongs — and where humans must remain authoritative.

Org
DesignOps SDK 1.0
Role
Author
Year
2025
Discipline
AI governance
Problem

AI doesn't get to default-on. It also doesn't get to default-off.

Every design organization is being pulled in two unproductive directions at once: a procurement function that wants AI everywhere, and a craft function that wants it nowhere. Both lose.

The work below is the structural alternative — a frame for placing each kind of DesignOps work into the quadrant where AI does the most good with the least risk, then operating it under severity controls that scale to the cost of being wrong.

Difficulty
HardEasy
High AI Appropriateness / Low Ease

Strategic Payoff, Real Work

  • Capacity modeling and forecasting
  • Scenario planning
  • Scenario tagging
  • Design-system consistency checks
  • Quality-drift detection
  • Portfolio-level prioritization signals
High AI Appropriateness / High Ease

Automation

  • Status reporting and rollups
  • Intake normalization and tagging
  • Meeting notes and management
  • Workflow state tracking
  • Retrospective synthesis
Low AI Appropriateness / Low Ease

Do Not Start Here

  • Strategy setting
  • Performance management
  • Career coaching
  • Conflict mediation
  • Cultural norm enforcement
Low AI Appropriateness / High Ease

Augmentation, Not Replacement

  • Heuristic evaluations (with human support)
  • Onboarding support arbitration
  • Change-comms drafting
  • Leadership comms (draft → human polish)
  • AI copilots, not drivers
High payoff
Low payoff
The 2×2: difficulty on the x-axis, strategic payoff on the y. Four quadrants — Automation, Augmentation, Strategic Payoff Real Work, and Do Not Start Here. Typeset by hand from the original AI-generated draft.
The four quadrants

What each quadrant decides.

Automation (high payoff, easy): status reporting, intake tagging, meeting notes, workflow state, retrospective synthesis. AI should do this. Humans should review the rollup.

Augmentation, not replacement (low payoff, easy): heuristic evaluations with human support, onboarding support, change comms drafting, leadership-comms drafts then human polish. AI as copilot, never driver.

Strategic payoff, real work (high payoff, hard): capacity modeling, scenario planning, design-system consistency checks, quality-drift detection, portfolio-level prioritization signals. The investment cases for serious AI work.

Do not start here (low payoff, hard): strategy setting, performance management, career coaching, conflict mediation, cultural norm enforcement. These are leadership work and must remain human-authoritative.

The rails behind the 2×2

Severity, auto-promotion, escalation.

The 2×2 names where AI belongs. The DesignOps SDK governs how it operates inside each box: a severity matrix (S1/S2/S3), auto-promotion logic that requires demonstrated reliability before AI moves from suggestion to action, and escalation paths for every category in which AI can be wrong.

The move
Severity matrix

S1: AI may act. S2: AI may draft, human approves. S3: human authors, AI may not draft. Every workflow gets a class.

The move
Auto-promotion logic

A workflow only moves to a less restrictive class after measured accuracy, bias review, and a defined trial period — not on vendor say-so.

The move
Recruitment pivot

Hiring for orchestration and critical evaluation, not for prompt fluency. The SDK rewrites the JD before it rewrites the workflow.

Outcome

Reality check, not theatre.

The SDK includes a "reality check" section — explicit acknowledgement that today's AI is brittle, that adoption is uneven, and that the right move in some places is to wait. Governance that refuses to oversell is the only kind of governance the practice can actually run on.

Leadership Principles

How this maps to leadership principles.

  • Are Right, A Lot: severity classes turn 'should AI do this?' into a structured, repeatable decision.
  • Insist on the Highest Standards: auto-promotion logic forces measured evidence before AI gains scope.
  • Earn Trust: explicit human-authoritative zones are how a design org earns trust from partners.
  • Frugality: refusing AI in the wrong boxes is the largest cost-avoidance move a design org can make.
← Previous · Case 01

The DesignOps Operating System

Next · Case 03

Living Clarity