An Adviser AI reads, reasons, and recommends. It never acts without being asked, and it never creates something new from scratch. Think of it as the most well-read colleague you’ve ever had — one who analyses anything you put in front of it and gives you a clear judgment. The key characteristic: the human is always in full control, and the human always carries the risk of acting on the advice.
Products that work this way
Perplexity AI — answers your questions with cited sources. It reasons over the web and gives you a judgment. You decide whether to trust it.
ChatGPT in Q&A mode — you ask, it answers. No action taken, nothing created unless you specifically ask for it.
Harvey (legal AI) — reads contracts, flags risks, recommends positions. The lawyer decides what advice to give the client.
UX patterns this type must get right
TransparencyShow the AI’s reasoning process, limitations, and confidence level — not just the answer.ExplainabilityAllow the user to ask “why did you say that?” and get a meaningful, human-readable answer.Source footprintsEvery factual claim traces back to a specific, verifiable source — so the human can check the AI’s work.Confidence signalsSignal when the AI is certain vs. uncertain — so users calibrate how much to trust the output.
Interaction Model 2
Generator
“It makes. You choose.”
A Generator AI produces new things — text, images, code, sound, video — from a brief. It doesn’t copy; it composes. The creative act is shared: you describe what you want, it produces multiple options, you judge and select. The output is raw material, not finished work. The human who accepts and publishes a Generator’s output becomes responsible for it — including any errors, IP issues, or misrepresentations.
Products that work this way
Midjourney — you describe an image in a prompt; it generates four variations. You choose which to develop.
GitHub Copilot — suggests code completions as you type. You accept, modify, or reject each suggestion.
Adobe Firefly — generates images inside Creative Suite from text prompts. Designed for commercial use with IP-safer training data.
UX patterns this type must get right
Edit & refineLet users modify specific parts of the output — not just accept or reject the whole thing.Multiple optionsAlways offer variations — not a single answer. Selection is how the human exercises creative judgment.Consent to useMake clear that using generated content means the human accepts responsibility for it.Provenance signalsIndicate what the AI was trained on, so users can assess IP and copyright risk before publishing.
Interaction Model 3
Agent
“It acts. You supervise.”
An Agent AI doesn’t just think or make — it does. Given a goal, it breaks the work into steps, uses tools, browses, writes, sends, and delivers — without being prompted at each turn. This is where AI stops being a tool and starts being a collaborator with initiative. The critical design question for any Agent is: what happens when it’s wrong, and can the human undo it? The Flemisch H-Metaphor calls this the “loose rein” principle — you need enough control to stop the horse.
Products that work this way
Salesforce Agentforce — handles customer queries end-to-end, updates CRM records, escalates when needed. Operates autonomously within guardrails.
Clay — researches prospects, enriches CRM data, and triggers outreach sequences without human involvement at each step.
Perplexity Deep Research — given a question, it runs dozens of searches, synthesises findings, and delivers a full report autonomously.
UX patterns this type must get right
Human approvalBefore any significant or irreversible action, the agent presents its plan and waits for explicit sign-off.Undo / rollbackEvery action the agent takes should be reversible — or have a checkpoint before it can’t be undone.Progress trackingShow the agent’s steps as they unfold — not just the final result — so humans can intervene mid-task.Escalation pathWhen the agent hits something beyond its authority or confidence, it stops and hands control back to a human.FootprintsA full log of every action the agent took — what it did, when, and why — so outcomes are auditable.
Interaction Model 4
Orchestrator
“It leads. Others do.”
An Orchestrator AI manages a team of specialist agents. It holds the mission, assigns tasks, arbitrates when agents conflict, decides when work is good enough, and reports outcomes upward. It doesn’t execute the work itself — it decides who does it and when. This is the most powerful and the most complex interaction model, and it’s where governance matters most: the humans are furthest from the action, yet still fully accountable for everything the team does.
Products that work this way
CrewAI — a framework for building teams of AI agents, each with a defined role. One “crew manager” orchestrates the others toward a shared goal.
AutoGen (Microsoft) — multi-agent conversations where agents specialise and a coordinator routes between them to complete complex tasks.
Claude Projects with sub-agents — a lead model coordinating specialist Claude instances for research, writing, fact-checking, and synthesis.
UX patterns this type must get right
System transparencyUsers should be able to see which agent did what — not experience the team as a single black box.Audit trailEvery decision in the pipeline is logged — including which agent made it and what inputs it acted on.Human overrideAt any point, a human must be able to stop the orchestrator, override a decision, or take manual control.Failure escalationWhen one agent in the team fails or conflicts with another, the orchestrator escalates to a human rather than guessing.RACI clarityFor every action the team takes, it must be clear which human is Accountable. Accountability never sits with an AI.
AI Type, Risk & UX Assessment
What kind of AI is this?
Four archetypes. UX patterns. Risk. The automation ladder. Hover or tap each type to learn before you start.
2 pts first try · 1 pt after a hint · +1 pt bonus follow-up
Adviser
Reads and recommends. Never acts without you.
Tap to learn more →
Generator
Makes something new. You judge and choose.
Tap to learn more →
Agent
Given a goal, it acts. You supervise.
Tap to learn more →
Orchestrator
Leads a team of agents. It directs — others do.
Tap to learn more →
Also covered in this assessment
Automation LadderLevels 0–5 of AI autonomy
Five levels from “AI only suggests” to “AI acts fully autonomously.” The higher the level, the more critical the UX control patterns become. Coined as the H-Metaphor by Flemisch, Adams et al. (NASA, 2003).
H-MetaphorHorse & rider, Flemisch et al. 2003
A framework from NASA research (Flemisch, Adams, Conway, Goodrich, Palmer & Schutte, 2003) that models human-AI cooperation as a horse and rider. The key finding: “a loose rein beats no rein” — some human control is always better than none. Applied here to agentic AI design.
UX Risk Patterns8 patterns from Patricia Reiners
Eight design patterns that determine how safe and trustworthy an AI product is: Transparency, Explainability, Source Footprints, Progress Tracking, Human Approval, Undo/Rollback, Escalation, and Edit & Refine. From “Beyond the Copilot” — Patricia Reiners, UX on the Beach, Venice 2026.
