Introduction
Most designers think the next evolution is smart design tools…
The truth?
The evolution is that the designer is no longer just drawing screens — but writing the rules of trust and the boundaries of autonomy for AI agents that act on their own.
📊 According to Deloitte (2026):
Only 11% of companies have AI agents with mature governance — the rest are all pilot experiments or chaos.
This shift does not mean “user experience for robots.”
The topic is bigger:
Who owns the decision when the product acts on its own?
Who is responsible if AI exceeds the user's intent?
This article is not about design tools…
It is about how the designer is becoming a trust systems engineer, and how the role of each individual in the product team changes when AI becomes a real player.
The difference is no longer in visual skill.
The difference is in who understands how the system is enforced…
And who manages the chaos when the screen is no longer the centre of the game.
Table of Contents
1. What is AX really? And why is it not just "UX for robots"? 2. Why is this shift happening now? (Three forces that triggered the sequence) 3. The six trust patterns: The new design framework for the age of agents 4. QA in the age of uncertainty: From "button" testing to "decision" testing 5. The PM problem: Who really owns the AI agent? 6. The new designer: Skills that are no longer optional 7. AI-generated interfaces: Impossible testing and invisible techniques 8. What does the real AX Design Process look like? (And why no one has the recipe yet) 9. Where does the product team stand now? 10. Key Insights 11. FAQ
1. What is AX really? And why is it not just “UX for robots”?
The main idea: AX means designing the rules and boundaries of trust for agents — not the interfaces.
Everyone thinks AX is an extension of UX.
Fatal error.
UX = the user presses a button, the system responds as expected
AX = the agent acts alone… the outcome is not guaranteed… the designer sets the “rules” not the “steps”
The system has become probabilistic (Probabilistic) rather than deterministic (Deterministic).
An AI agent may book a ticket, change a payment card, or decide to escalate without any human interaction.
The designer here specifies:
- When does the agent act alone?
- When does it need permission?
- When does it stop completely?
The designer has become a “lawmaker”... not a “screen designer.”
2. Why is this shift happening now? (Three forces have triggered the sequence)
The main idea: three forces have made AX a necessity, not an option.
1️⃣ AI agents have become capable of “execution” rather than just suggestion
Generative AI suggests… Agentic AI executes and acts.
The difference? Real responsibility on the designer to set boundaries for action.
2️⃣ Model Context Protocol (MCP):
📊 Study (2026):
Improving the description of tools in MCP increased task success rates by 6%, and completion of partial goals by over 15%.
The power here is not in intelligence… but in the quality of the “description” and “rules” written by the designer.
3️⃣ Generative UI:
The interface is no longer pre-designed — it is generated by the agent in real-time according to context.
SAP, Google, and Figma now rely on it, and the test is no longer about “Does the button work?” but “Does the whole system understand the boundaries of action?”
All this has raised the question: who owns the decision when the agent makes a mistake or oversteps?
3. The six trust patterns: the new design framework for the age of agents
The main idea: there is no AX without six fundamental trust patterns.
1️⃣ Intent Preview
Agents must preview their intent before any irreversible action
Example: instead of “Your trip has been booked” it should show: “I found a trip like this... should I proceed?”
📊 FuseLab Creative:
Most cases of lost trust are due to “What just happened?” — the absence of prior review.
2️⃣ Autonomy Dial
The user determines the level of autonomy of the agent (for example: automatically books under $1000, requests permission above that).
The designer sets the assumptions and ranges.
3️⃣ Confidence Signal
Every decision comes with a confidence indicator
But...
📊 Research in LLM calibration:
90% confidence means only 75% accuracy — and each additional step decreases confidence further.
👉 Therefore: never trust a confidence signal alone — it must be linked to escalation or human approval.
4️⃣ Explainable Rationale
The agent explains why it made this decision (“because it’s the cheapest and matches your previous pattern...”) — no need for a full explanation… just enough logic to reveal errors.
5️⃣ Action Audit & Undo
A timeline for each action + the ability to undo
Example: the user sees which card was used, and can undo immediately.
6️⃣ Escalation Pathway
When does the agent stop and request human intervention?
Without this: it either fails silently, or retries endlessly.
These patterns are not theoretical — any product with autonomous behaviour needs them… otherwise the system will collapse at the first mistake.
4. QA in the age of uncertainty: from testing a “button” to testing a “decision”
The main idea: testing agentic systems is no longer about testing fixed outcomes… but testing boundaries and decisions.
Testing the AI agent is impossible in the traditional way:
One input = many outputs
Katalon (2026):
Traditional testing = following a recipe
Agentic QA = a chef who understands the goal and innovates
What is needed now:
- Boundary testing (Does the agent respect the spending cap?)
- Sequential error testing (First misunderstanding = all subsequent decisions are wrong)
- Escalation testing (Is the escalation real and not just a drawing on Figma?)
- Change testing after the Model update (Model update ≠ code diff)
📊 Salesforce:
They created the Agent Development Lifecycle (ADLC):
Complete tracking of the decision path, classifying user intents, and alerting when the agent's behaviour changes.
QA has become strategic: designing adversarial scenarios, monitoring the system, and deciding when to halt updates.
5. The PM's problem: Who really owns the AI agent?
Main idea: In AX, responsibility is no longer linear — it is distributed across the entire system.
An AI agent makes a mistake — who bears the fault?
Did the PM write the Acceptance criteria?
Did the designer set the rules?
Did QA not test the rare scenario?
The answer:
Everyone is responsible… and no one has complete control.
📊 Salesforce, Microsoft, Deloitte:
They all confirm — we need to redefine “ownership” and “quality” and “escalation” in agent systems.
The PM sets operational goals for the agent, not just for the user.
There is no longer a single owner of the decision… but a system that continuously monitors itself.
6. The new designer: Skills are no longer optional
Main idea: The successful designer now is a “trust and scenarios engineer” rather than just an interface maker.
Jakob Nielsen (2026):
The designer has become an “agent choreographer”
Salesforce:
From “interface architect” to “experience coordinator”
Smart products require:
- Permission models (What actions require permission?)
- Escalation rules
- Classification of trust layers
- Tool descriptions in a language the agent understands
📊 Figma (2026):
72% of designers use Generative AI
Those who intensified their usage reported a 25% increase in job satisfaction compared to others.
New skills:
- Understanding LLM behaviour
- Designing for probabilistic systems (variable outcomes)
- Writing smart Prompts (rules, personality, boundaries… before any interaction)
It is no longer enough to write Microcopy or draw Wireframes… you must design how the entire system “thinks”.
7. AI-generated interfaces: impossible testing and invisible techniques
Main idea: the interface is no longer static… it is generated as needed — and testing it has become a new challenge.
SAP, Google, Figma:
Generative UI has become a reality — the agent assembles components according to the user's intent.
The designer must:
- Build accurately described component libraries (when and why to use them, not just how they look)
- Write clear generation rules
- Prepare the system for failure (what happens if generation fails?)
QA testing here:
- Testing rules not results
- Testing failure cases
- Defining what is “correct” — is it enough for the interface to meet the intent? Does it comply with the Design System? Is it accessible?
For the first time: the designer and QA are testing a self-generating system… not all scenarios can be captured.
8. What does the real AX Design Process look like? (And why does no one have the recipe yet)
Main idea: there is no unified process — only principles that crystallise in the field.
Pixelmojo:
Start with an AX readiness audit (transparency, independence, rollback security)
Then build Trust Architecture (classifying actions for trust layers, mapping consent boundaries)
Then design Mediation Interfaces (review cards, trust indicators, action log)
And finally: failure scenarios — testing how the system behaves during misunderstandings or sequential errors.
The truth?
Everything is still being built with diligence… no ready-made templates.
The entire industry is writing its practical guide now… and most teams are learning by paying the price of mistakes.
9. Where does the product team stand now?
The main idea: AX is not a “design” problem — but a problem for the entire “product team.”
Boundaries, rules, scenario testing, protocol integration:
If any party is absent (designer, QA, PM, engineer)… the system collapses.
📊 Deloitte:
77% of companies have not yet established clear AX patterns.
The entire industry is living a moment of “what if the system acts on its own?” — and all solutions are being invented room by room.
10. Key Insights
- The successful designer in 2026 writes rules and boundaries for agents, not just screens.
- Trust is not a feeling… but a system: six patterns that must be built into every product.
- QA is no longer a testing function… but a strategic quality engineer for probabilistic systems.
- There is no single decision owner in AX — responsibility is collective and variable.
- AX is not an extension of UX… it is a redefinition of the relationship between humans and products and artificial intelligence.
- The entire industry is building the practical guide now… and those who delay will reinvent the wheel later.
11.FAQ
Will AX kill the role of the traditional designer?
No — it will make it more strategic. The screen has not disappeared, but those who set the rules have become more important than those who design the shape.
Does every product need AX?
Only products that rely on self-action from AI… if your product is just a display interface, you will remain in the world of UX.
How does QA test a system that does not provide consistent results?
It tests boundaries, rules, and failure cases — not literal results. It observes the system as a whole, not each button individually.
Who holds the decision if the AI agent fails?
No one alone — the whole system is responsible: the designer, the PM, QA, the engineer… they must share in building and monitoring the boundaries.
What is the one skill that AI cannot take from the designer now?
Judgment in ambiguous situations, and designing for complex systems.
About Echo Media
Echo Media is a company specialised in digital growth strategies and AI systems,
helping businesses build sustainable growth engines through marketing, sales, and operations.
We focus on transforming AI from experimental tools into real operational systems
that support decision-making, build scalable digital assets, and help businesses grow
independently of the individual effort of the founder.
Our expertise includes:
• AI strategies for businesses
• Building scalable growth systems (Growth Systems)
• Product design and digital experience (UX)
• Data-driven content and SEO strategies
Learn more: