← All insights
Intelligence · 8 min read · Thursday 5 November 2026

When AI Becomes the Interface: Designing for Delegation, Not Clicks

AI agents are moving from answering questions to acting on our behalf. The evidence shows trust is earned through visible plans, user checkpoints and competence — not personality.

The interface is shifting from screens to intent

For thirty years, digital product design assumed a person would navigate: tap, scroll, compare, confirm. Agentic AI breaks that assumption. A user states an outcome — book the flight, reconcile the invoice, pay the supplier — and a system plans and executes the steps.

Nielsen Norman Group describes AI agents as new actors inside the service ecosystem: personal assistants that coordinate across services, and organisation-built agents that serve customers (and increasingly, customers' own agents). Service metrics, they argue, will have to change with them.

What the research says about trust in agents

A 2025 CHI study of 248 participants tested LLM agents as daily assistants across six tasks of varying risk, including flight booking and credit card payments. The finding is a design brief in itself: agents performed well when they produced a high-quality plan and kept users involved at execution — but users easily mistrusted agents, or over-trusted flawed plans, when that structure was missing.

Usability testing by Carnegie Mellon researchers on commercial agents (Operator and Manus) found participants were impressed, yet struggled with capabilities that didn't match their mental models and with agents that lacked the self-awareness to collaborate effectively.

And on the question of tone: NN/G's review of a 410-participant study found people trust AI more when it appears intelligent rather than emotional. Simulated feelings can actually reduce reliance on factual, task-oriented advice.

Design principles for delegated experiences

Taken together, the evidence points to a practical pattern product teams can adopt now:

  • Show the plan before acting. Let users inspect, edit and approve steps — especially for money, identity or irreversible actions.
  • Scale checkpoints to risk. Low-risk tasks can run autonomously; payments and data sharing need explicit confirmation.
  • Signal competence, not sentience. Clear reasoning, sources and limits beat a warm persona.
  • Design the recovery path. Undo, audit trails and human hand-off are part of the experience, not edge cases.
  • Measure delegation success: task completion without correction, override rates and time-to-trust — not clicks.
Figure 1.1
State machine for agentic UX delegation
Intentuser states outcomePlansteps made visibleCheckpointscaled to riskExecuteagent actsVerifyaudit trailuser edits planRecoveryundo · explain · human hand-offLow risk: auto-approveMoney / identity: explicit confirmIrreversible: confirm + cooling-off
Figure 1.1. Delegation flows from intent to a visible plan, through a risk-scaled user checkpoint, into execution and verification. Every state has a recovery path back to the user.Source: Synthesised from ACM CHI 2025 (Plan-Then-Execute) and NN/G 2025. Diagram: DfRW.

Why this matters in high-risk markets

In markets where social engineering already drives most digital fraud, an agent that can move money is also a new attack surface. The design question isn't only 'can the agent do it?' but 'can the user tell, at every step, what is being done in their name?'. That is the core of LAB 01.

Key takeaways

  • Trust in agents follows visible plans and user checkpoints.
  • Competence signals build reliance; anthropomorphism can erode it.
  • Success metrics move from engagement to delegation quality.

Sources

  1. How Service Design Will Evolve with AI Agents — Nielsen Norman Group, 2025
  2. Plan-Then-Execute: User Trust and Team Performance with LLM Agents — ACM CHI 2025
  3. Why Johnny Can't Use Agents — Carnegie Mellon University, ACM CAIS 2025
  4. Prioritize Smarts over Sentience to Increase Trust with AI — Nielsen Norman Group, 2025
  5. Understanding User Experiences of Computer Use Agents — Apple / CMU, arXiv
More insights