AutoScout24 · Strategic AI Design Leadership

AI Visibility, Review, and Control

Teams were being asked to ship AI before we had a shared idea of how AI should behave in the product. I built a practical Visual AI Framework to turn that ambiguity into decisions about visibility, confirmation, correction, recovery, and user control.

I then used that guidance in reviews and hands-on Apps design for Conversational AI Search, where natural-language requests need to stay connected to filters, listings, and a reliable path back to search.

Role: Principal Product Designer, AI Design Lead
Problem: Teams needed shared rules for how AI should behave in product
Owned: Visual AI Framework, review model, and UX education paths
Current product work: Apps Conversational AI Search patterns for confirmation, correction, recovery, and control
Focus: When AI should ask, explain, invite review, support correction, or require confirmation

Full-bleed laptop mockup showing an AI communication framework.

Why This Work Existed

AI work was moving faster than shared behavior standards

The design risk was fragmentation: different teams could ship different versions of AI, with different confidence cues, correction paths, and levels of user control. I turned that into a shared way to discuss product behavior.

Leadership wanted AI products moving faster, but teams had little precedent for how AI should behave inside the product.
Without shared rules, every team could invent its own visibility, confidence cues, correction paths, and level of user control.
Teams needed practical review criteria while they were already designing, not a detached framework after the fact.

Role

Ownership and influence

I owned the reusable standards work and the AI design guidance behind it. I also applied that guidance in product work, designing how AI-assisted search should capture what the user is asking for, confirm understanding, support correction, and return control to the user. Across teams, I helped shape key flows, principles, and tradeoffs alongside other designers and cross-functional partners.

Direct ownership

The Visual AI Framework
AI design guidance and education paths for the UX team
Hands-on Apps design for Conversational AI Search patterns around review, correction, recovery, and control
Cross-functional capability work around AI-assisted product-development workflows
AI prototyping capability across the broader product organisation

Shaped through influence

Senior design direction on early AI-assisted product concepts
Alignment with senior product, design, and technology leaders on AI-assisted workflow change
Key flows and experience principles on early AI product work
Guidance that shaped decisions beyond direct reporting lines

Visual AI Framework

The framework made AI behavior discussable

The framework translated broad AI ambition into product questions: what did AI change, should it act or confirm, how can the user correct it, and when does control return?

Visibility scaled with user impact

I set the rule that teams should increase AI visibility as its effect on user understanding, decisions, and control increased.

Different AI behaviours needed different signals

I distinguished between low-visibility assistance and cases where AI generated, summarised, recommended, personalised in a non-obvious way, or acted on the user's behalf.

Controls increased with stakes

As AI moved closer to decision-shaping or action-taking, teams needed stronger review, consent, editability, override, and exit paths.

I rebuilt the model below from internal framework work with generalized examples to protect confidential product strategy.

Subtle assistive behaviour could stay quiet. Generated summaries, recommendations, and non-obvious personalisation needed clearer labelling and a stronger review path.

I combined a shared AI visibility model with reusable signals such as labels, badges, gradients, and icons, plus guidance on review and override. I used it to clarify what needed explicit signalling, what could remain lightweight, and when stronger review or control was necessary.

Quiet assistance

Formatting, cleanup, or small suggestions

Signal: No persistent label needed

Control: Normal edit or undo

Generated or rewritten content

A draft, summary, or suggested wording

Signal: Lightweight AI label

Control: Edit, regenerate, or dismiss

Recommendation

A ranked suggestion or next-best action

Signal: Explain why it appears

Control: Compare, override, or ignore

Decision-shaping summary

A synthesis that may affect user judgement

Signal: Clear disclosure and source access

Control: Review source, reject, or correct

Action on the user's behalf

Changing state, sending, publishing, or committing

Signal: Explicit confirmation

Control: Consent, undo, and audit trail

Decision Guidance

Decisions I clarified with the framework

I used the framework to make AI concept reviews more specific before patterns hardened. I connected AI behaviour to user-facing signals, controls, and review paths.

Low-impact assistance could stay quiet

Formatting, cleanup, and small suggestions could rely on normal edit or undo controls when the risk stayed low.

Generated output needed review paths

I treated drafts, summaries, and suggested wording as outputs that needed visible AI signals, editability, regeneration, and dismissal.

Recommendations needed reasons

I paired ranked suggestions and next-best actions with a short reason, plus a way for users to compare, override, or ignore them.

Decision-shaping summaries needed sources

For AI synthesis that could affect user judgement, I pushed for source access, correction paths, and stronger review.

AI actions needed confirmation

For AI actions that changed state, sent, published, or committed something, I raised the bar to explicit consent, undo, and auditability.

Framework Example

How the framework handled decision-shaping summaries

This is a generalized pattern from the Visual AI Framework: AI that shapes interpretation needs clearer disclosure, source access, correction, and review than lightweight assistance.

Product question

When AI condenses information that may affect a user's judgement, the interface needs to show that AI shaped the output.

Design risk

A neutral-looking summary could make users over-trust the output or miss the source material behind the recommendation.

Design direction

I use the framework to push for visible AI signalling, source access, correction paths, and review before the user acts on the output.

Tradeoff

Routine assistance stayed lightweight. AI that shaped interpretation needed stronger disclosure and control because the consequence for user judgement was higher.

Use In Practice

From guidance to product behavior

The useful part of the work was not the framework alone. It was making reviews and product decisions more concrete while teams were already designing AI-assisted experiences.

Feature-team guidance

I used the model with Conversational AI Search and Lead Assistant teams, turning rules for visibility, confidence, correction, recovery, and control into concept-review criteria before patterns hardened.

Conversational AI Search

I now apply the same guidance hands-on in Apps, where natural-language requests need to stay connected to filters, listings, confirmation, return-to-results, and the classic search path.

UX education and enablement

I turned the guidance into education paths and examples for designers working with AI in research, synthesis, content work, workshop planning, and prototype exploration.

Senior stakeholder alignment

I used the same decision language with senior product, design, and technology leaders when discussing how AI-assisted workflows changed design practice and product development.

Working Artifacts

Making the framework usable in product reviews

I translated principles into artifact guidance and review prompts that teams could use during concept reviews and prototype discussions.

AI presence scale

I mapped AI behaviours from quiet assistance to action-taking so teams could judge how much signalling and control a concept needed.

Visual signal guidance

I shaped guidance for labels, badges, gradients, and iconography so AI communication stayed clear without turning every interaction into AI theatre.

Prototype review prompts

I translated the guidance into review questions teams could apply to concept flows before investing in detailed UI or implementation.

Input

Identify the data, source material, and what the user is asking the system to do.

Output

Check what the user sees and whether AI output could read like neutral product copy.

Uncertainty

Decide where the design needs source access, confidence cues, alternatives, or correction.

Control

Confirm that users can edit, reject, undo, override, or stop the AI at the right moment.

Failure

Define the path when AI is wrong, incomplete, overconfident, or acting on weak context.

Leverage

The value was in standards, product craft, and organisational capability

I used this work to create a clearer basis for AI product decisions and more practical ways to design AI-assisted experiences.

Early product direction

I contributed senior design direction to early AI-assisted product concepts and applied the same rules to hands-on Apps search design. I focused on what the system should reveal, when it should ask or confirm, how users correct it, and where control returns.

Internal capability building

I led AI guidance for UX, education paths for AI-assisted workflows, and prototyping capability across the broader product organisation. I also led cross-functional capability work with senior product, design, and technology leaders on changes to design practice and product development.

Practical UX education

I translated AI research and tool exploration into examples designers could use for discovery, synthesis, content work, workshop planning, and prototype exploration.

Prototype-ready review criteria

I helped early AI ideas move toward testable flows with review points for capability, limits, user control, and failure recovery.

I keep confidential product detail and outcome metrics out of this case. The public evidence here is the framework, review model, capability work, and the way I apply those rules to hands-on AI feature design: request capture, confirmation, correction, recovery, and user control.