UX
Master's Thesis · UdK Berlin · 2025

UX is not
a feature.
It's governance.

What happens to UX designers when AI systems run the product? A research study across 3 organizations, 4 frameworks, and one uncomfortable truth about who actually steers AI behavior.

Author Uta Janzen
Institution Universität der Künste Berlin
Program Leadership in Digital Innovation
Date December 2025
Scroll

AI didn't replace UX.
It buried it.

When AI systems run 24/7, learn from user behavior, and make probabilistic decisions at scale — who decides what's acceptable? Who watches the outputs? Who sets the guardrails?

In theory: governance frameworks, management systems, risk owners. In practice: UX designers, creative directors, and art directors — doing it informally, without mandate, without recognition, often without even realizing it.

This research investigates that gap. Across three organizations and four major AI governance frameworks, it maps how AI-integrated business models reshape UX roles — and what happens when nobody officially owns the work that actually steers AI behavior.

74%
of companies struggle to scale AI beyond pilots — not because of technology, but because of organizational gaps. (BCG, 2024)
0%
Risk management coverage in one of the studied organizations — despite running a live AI platform with paying enterprise clients.
43%
of all coded interview segments were about AI curation work — the informal, unnamed governance layer that keeps AI systems functional.
Core Tensions

Six fault lines in
AI-integrated organizations

S — 01

Methods & Process Gaps

Classic UX frameworks (Design Thinking, Agile, Scrum) were built for discrete deliverables. AI systems are always-on, probabilistic, and continuously learning. The methods don't fit the material.

S — 02

Governance & Ownership

Who is responsible when an AI system fails? Frameworks define abstract roles like "Risk Owner" — but in practice, nobody is formally assigned post-deployment curation. The vacuum gets filled informally.

S — 03

Compliance & Transparency

The EU AI Act and ISO standards demand transparency and auditability. But UX teams — who design the moments where AI becomes visible to users — are barely mentioned in any compliance framework.

S — 04

Business Models & Metrics

Usage-based revenue, always-on services, and AI inference costs require new governance logic. 91% of US agencies use or explore generative AI (Forrester, 2024), yet no framework addresses this disconnect.

S — 05

Design Roles & Identity ⚠ Critical Gap

AI-specific design work — prompt engineering, guardrail design, behavioral blueprinting — is systematically invisible in governance standards. Three of four major frameworks score 0.5 or below on this dimension.

S — 06

Power & Fairness

AI systems trained on historical data reproduce historical inequalities. Who in your organization has the mandate to challenge the training data? Mostly: nobody. Or the UX team, informally, under pressure.

The Framework

The A×B×C×D Model

To analyze how AI business models shape UX work, this research developed a four-dimensional analytical grid connecting macro-economics to on-the-ground design practice.

A Lifecycle Position
AI System Phases
  • A1 — Planning
  • A2 — Data preparation
  • A3 — Model & experience design
  • A4 — Deployment
  • A5 — Monitoring & operations
B Governance Dimensions
Steering Mechanisms
  • B1 — Risk management
  • B2 — Transparency / docs
  • B3 — Roles & ownership
  • B4 — Processes
  • B5 — Ethics & fairness
  • B6 — Metrics
C UX Roles & Activities
What UX actually does
  • C1 — Strategy
  • C2 — Research
  • C3 — Design
  • C4 — Testing
  • C5 — Operations
  • C6 — AI Curation ★
D Outcome Dimensions
Measurable effects
  • D1 — Quality
  • D2 — Lead time
  • D3 — Incidents & risk
  • D4 — Cost & efficiency
  • D5 — Adoption
C6The invisible layer

C6 — AI Curation and Meta-Steering — was added inductively during the research. It describes work that classic UX models simply don't account for: deciding which AI outputs are acceptable, designing behavioral blueprints, defining guardrails, monitoring for bias, orchestrating model pipelines.

In every organization studied, this work was being done — by someone, somehow. In none of them was it formally assigned, budgeted, or recognized. C6 is the real governance layer. It just doesn't have a job title yet.

Prompt design Guardrail definition Output curation Bias monitoring Behavioral blueprinting Pipeline orchestration Failure-mode testing
How It Was Studied

Two sources,
one analytical grid

A qualitative, multi-perspective case study. Normative frameworks (the "should") and lived organizational practice (the "is") were coded against the same A×B×C×D grid — so gaps between the two become visible, not just asserted.

Source 01 — Normative

Framework analysis

Four major AI frameworks scored with an ordinal 0–2 scale (none / partial / explicit coverage) across 14 core A×B×C cells, then aggregated into the six tension areas. The result is the heatmap below.

NIST AI RMF ISO/IEC 42001 ISO/IEC 5338 PAIR Guidebook
Source 02 — Empirical

Expert interviews

Four guided interviews with six practitioners across three contrasting organizations (Nov 2025). Transcribed, anonymized (U1–U3), and segmented into 181 coded units along the same dimensions.

4 interviews 6 practitioners 181 segments 3 cases

Coding followed a structuring qualitative content analysis: deductive A×B×C×D codes plus inductive subcodes for AI-curation work. Framework "should" and practice "is" were then triangulated — a tension area counts as critical only when both the heatmap shows thin coverage and the cases report concrete problems there.

Framework Analysis

What governance standards
actually cover

Four major AI frameworks scored across six tension areas. Score 0–2: none / partial / explicit coverage.

Tension Area NIST AI RMF ISO/IEC 42001 ISO/IEC 5338 PAIR Guidebook
S1 — Methods & Processes 1.332.002.002.00
S2 — Governance & Ownership 1.671.671.670.33
S3 — Compliance & Transparency 1.001.001.002.00
S4 — Business Models & Metrics 1.002.002.001.50
S5 — Design Roles & Identity Critical Gap 0.500.000.502.00
S6 — Power & Fairness 2.002.002.001.00
Overall average 1.25 1.44 1.53 1.47

Green ≥1.5  ·  Amber 0.8–1.4  ·  Red <0.8  ·  ⚠ critical gap = significant structural absence

Empirical Findings

Three organizations.
Three very different answers
to the same problem.

U1 — AI-first Software Startup · ~10 people
The Speed Regime
Governance deliberately minimized. No formal risk management. No UX role. AI curation work happens invisibly inside hybrid roles. Fast prototypes ship in days — stability is fixed in production. "A risk owner would be a blocker."
C6 coverage: 18% of segments
Risk management (B1): 0%
Trade-off: Speed ↑ Quality/Risk ↓
U2 — Integrated Creative Agency · ~100 people
The Identity Regime
Governance eroded by margin pressure. No AI content owner, despite 1.5 years of AI use. Creative identity fills the gap: "it shouldn't look like AI" becomes the de-facto quality standard. One Art Director does the governance work of an entire team — informally, without budget.
C6 coverage: 42% of segments
S5 Identity: highest of all cases (25%)
Trade-off: Quality ↑ Cost/Wellbeing ↓
U3 — Communications Agency · ~45 people
The Control Regime
The only case with an explicit AI Engineer role. Governance emerging, driven by external legitimacy pressure (clients asking, AI Act looming). Internal AI server, ComfyUI pipelines, AI workflows in proposals. But still no formal metrics, no ethics process, no escalation paths.
C6 coverage: 43% of segments
B1 Risk: highest of all (28%)
Trade-off: Balance ↑ Measurability ↓

The Central Insight

"In every organization we studied, AI curation work was being done. In none of them was it formally assigned, budgeted, or recognized as governance."

The B-dimension (governance) is shifting into C6 (curation work) — without anyone acknowledging it in the organizational design. This isn't a bug. It's a structural consequence of accelerated GenAI adoption.

What This Means

Eight things to do
about it

For organizations

1

Couple AI strategy with governance architecture

Where AI intensity grows (A1→A5), governance mechanisms (B1–B6) must develop in parallel — not as an afterthought, but as a prerequisite.

2

Distribute AI literacy systematically

Transform C6 work from individual compensation to collective competence. This requires training, not just hiring one AI specialist.

3

Create AI-UX hybrid roles with real mandates

AI-UX Leads who combine curation expertise with standards knowledge and risk awareness. With budget. With decision authority.

4

Use the six tension areas as intervention points

S5 (role ambiguity) → clarify accountability. S3 (compliance) → map UX touchpoints to regulatory requirements. S4 (metrics) → make trade-offs explicit.

For UX professionals

5

Build competence in the C6 zone

Prompt design, guardrail definition, failure-mode diagnosis, monitoring UX, pipeline orchestration. These are design skills. Claim them.

6

Develop AI literacy beyond tool operation

Understanding model limits, failure modes, systematic biases, and governance implications — not just knowing which button to press in Midjourney.

7

Define and track C6-specific metrics

Prompt efficiency, output quality scores, guardrail effectiveness, bias reduction. If you can't measure it, you can't argue for the budget.

8

Position C6 as strategic design work

Not additional overhead. Not invisible labor. The work that determines whether AI systems behave in alignment with human values is design work. Communicate that.