We just dropped 5 podcast clips from our conversation with Menno Cramer, whose simple but powerful User Experience map has been sparking some great thinking across our team.
In our conversation, Menno talks about why he created the map. His goal was to make sense of a rapidly shifting field. Traditional UX is still essential, but we’re being asked to stretch our skills. As AI systems and intelligent agents become part of our products, designers need to think beyond usability. We now need to consider explainability, agent intent, conversational tone, emotional intelligence, and much more.
We shared the image on LinkedIn with this quote from Menno:
“Honestly just trying to make sense of things… we need to rethink and reshuffle our existing understanding of UX to ensure that we can give AI its rightful place within our world.”
Here are the short clips to watch:
His chart organize user experience into three layers:
Traditional UX includes usability, research, wireframes, and metrics.
AI in UX adds adaptive design, HITL, fairness, and trust.
Agent UX explores agent personas, tone of voice, cognitive load, and emotional cues.
Discussion
How is your team approaching this shift? Are you starting to test things like trust, confidence, and agent behavior?
We’d love to hear how you’re reshaping your UX practices. Have you started using UX metrics like intent clarity, trust, or feedback perception in your AI-related work?
This is so cool. Thanks for sharing this here! Would love to hear thoughts from others on this. Especially on the “Designers just simply copy trends”. On one side I feel we need a scientific study of HITL, HATL, HOTL, HBTL, and on the other hand I guess well just learn as we go…
Build, learn, iterate, repeat…
Yeah! I could see this becoming super important. I’m guessing that things will more than likely diverge depending on how people want things, aka more variety.
I can see cheap, fully automated restaurants becoming a thing, and expensive but high quality human serviced restaurants being on the other end of the spectrum.
I think it would be helpful to hear what HITL actually looks like in practice, and how that might help us understand the “Designers just simply copy trends” hunch. Could you maybe explain what all is involved across H*TL methods?
To cut a long story short. The difference is based on the extend to which the human in present. Do they approve, do they monitor, do they get a dashboard, or do they get called upon when there is an issue…
In more detail.
HITL Human-in-the-Loop
Summary:
Humans are actively involved in the training, testing, and feedback loop of the AI system. They validate, correct, or fine-tune outputs before the AI acts on them. This is often used in high-stakes, ambiguous, or evolving environments.
UX Patterns & Behaviors:
Active validation UI: Users are prompted to confirm or adjust AI suggestions before proceeding.
Thumbs up/down feedback: Common for quick sentiment input on individual results.
Editable suggestions: AI provides a draft, and humans make final changes.
Confidence indicators: Show how sure the AI is, encouraging user scrutiny in low-confidence scenarios.
Interruptible flows: Humans can override or halt actions easily.
HATL Human-Above-the-Loop
Summary:
Humans oversee the AI system from a supervisory position. The AI acts autonomously, but humans define the rules, monitor outcomes, and intervene as needed.
UX Patterns & Behaviors:
Dashboard-based monitoring: Users interact with summary views showing system performance and potential flags.
Audit logs: Traceable histories of AI actions to support transparency and governance.
Threshold settings: Users can tune decision thresholds, parameters, or escalation rules.
Periodic interventions: Users engage primarily in exception handling or system updates.
Alerts and notifications: Passive until anomalies arise.
HOTL Human-On-the-Loop
Summary:
Humans remain available during AI operations and can intervene, but are not required to act unless there’s a problem. AI has more autonomy than in HITL.
UX Patterns & Behaviors:
Real-time status feeds: Users track the process in progress, like a live operations panel.
Intervention tools: UI includes options to pause, stop, or adjust processes if anomalies are detected.
Confidence thresholds with color coding: Visual indicators help prioritize what might need attention.
“Set it and watch it” behavior: Users trust the system but want assurance it’s working correctly.
HBTL Human-Below-the-Loop
Summary:
The AI operates entirely autonomously, and humans only review or assess the results after the fact. This is typical in automated environments where human review is rare or only needed for auditing.
UX Patterns & Behaviors:
Post-hoc reporting interfaces: Users engage with analytics dashboards or outcome summaries.
Quality sampling UI: Random or targeted review of outputs with flagging options.
Batch review workflows: Evaluate AI performance over time or across large data sets.
Detached interaction: Users engage only periodically, often through performance or compliance lenses.
Another followup AI post, Sean J. Savage does a good job explaining how to build AI products people can trust over time. He suggests you should treat them as layers that change at different speeds.
Instead of thinking of AI as one big feature, he breaks it into parts, like quick user interactions, the interface, the AI system itself, and long-term strategy. The real challenges (and opportunities) show up where these layers connect.
Such a good overview here @menno - Thank you for getting this conversation into a larger audience. I am working diligently in the HITL space; more specifically in the philosophical category, and it has been an interesting and wide discussion taking me across the world so far.
Here’s how I see your model functioning based upon existing and current AI frameworks, High-level models, and direct conversations with AI ‘workers’ (Profs at Universities, Data Scientists, Engineers, etc. - only those with whom I have direct communication, or access to products)
Traditional UX : Utilizes usability, research, wireframes, and metrics in a continuous loop which drives…
AI in UX : (In this model ) HITL = fairness + trust [for access to literal ontologies and objective reality] which then allows for correct adaptive design
Agent UX : tone of voice + cognitive load + emotional cues = agent personas. Can only be correctly implemented if AI in UX stage is implemented correctly.
In this observation, the potential success/failure point for ALL AI implementation is the AI in UX layer and how correctly this layer will be constructed. It is the center point of success or failure.
I’m wondering whether this is the case or more foundational to principles. Is the MoltBook experiment really about a UX problem, or something more foundational? Seems to be what you have been highlighting @schuboxaz recently. Curious about your thoughts @menno.
Discussion
) HITL = fairness + trust [for access to literal ontologies and objective reality] which then allows for correct adaptive design