Interaction Patterns

Context

Your AI feature works technically. Accuracy is solid, the model delivers. Then user feedback rolls in: “I don’t know what it’s doing.” “Why did it just do that?” “How do I undo this?”

The problem isn’t the AI. The problem is the interaction design. As a PM, you don’t just decide what the AI does — you decide how much control the user has, when the AI acts, and what happens when it gets it wrong.

Concept

Human-in-the-Loop — Three Levels

Not every AI interaction needs the same degree of human control. There are three levels:

Human-in-the-Loop: AI proposes, human must explicitly approve before anything happens. Highest safety, lowest speed. Example: doctor confirms AI diagnosis.
Human-on-the-Loop: AI acts, human monitors and can intervene. Moderate trade-off. Example: AI categorizes tickets, team reviews dashboard.
Human-out-of-the-Loop: AI acts fully autonomously. Only for low-risk, validated tasks. Example: spam filter.

AI Suggestions vs. AI Actions

The interaction between user and AI lives on a spectrum:

Pattern	AI does	User does	Example
Suggestion	Proposes	Decides	GitHub Copilot ghost text
Auto-complete	Fills in	Can override	Gmail Smart Compose
Action with preview	Acts, shows preview	Confirms or rejects	Notion AI diff (gray/blue)
Autonomous action	Acts without approval	Gets informed	Linear auto-apply
Autonomous with undo	Acts without approval	Can reverse	Email categorization

PM implication: The further right on the spectrum, the more you invest in undo, audit, and explainability.

Pre / In / Post — Three Phases of Every AI Action

Every AI interaction has three phases you need to design deliberately:

Pre-action: Intent preview (what will the AI do?), Autonomy dial (how much control does the user have?)
In-action: Explainable rationale (why is it doing this?), Confidence signal (how certain is it?)
Post-action: Action audit and undo (what did it do, how do I reverse it?), Escalation pathway (what happens when things go wrong?)

Escalation Patterns

When the AI hits its limits, it needs clear escalation paths:

Confidence-based: Below a threshold, automatically hand off to humans
Risk-based: High-stakes decisions always require humans
Expertise routing: AI learns which humans are best for which questions
Channel-based: Slack/email/dashboard based on urgency

Feedback Loops

Users need to give the AI feedback — but the mechanism matters:

Blocking feedback: AI pauses and waits for input. Good for critical decisions.
Parallel feedback: AI continues working, feedback arrives async. Good for volume tasks.

Apple distinguishes three feedback types: Implicit (behavior), Explicit (thumbs up/down), Corrections (fixing mistakes). Key principle: corrections are no substitute for poor results. Guided corrections (choosing from options) beat freeform corrections (open text field).

Service Recovery Paradox

Successful recovery after failure creates more loyalty than flawless performance. This means: invest more UX budget in error states and undo than in the happy path.

Framework

Interaction Pattern Decision Matrix — choose the right pattern based on risk and frequency:

	Low frequency	High frequency
High risk	Human-in-the-loop + action with preview	Human-on-the-loop + confidence escalation
Low risk	Suggestion or auto-complete	Autonomous action with undo

Checklist for every AI interaction:

Phase	Must have	Example
Pre	Intent preview	”I will summarize 3 paragraphs”
Pre	Autonomy dial	User can adjust autonomy level
In	Confidence signal	Color coding, percentage display
In	Explanation	”Based on the last 5 tickets”
Post	Undo	One-click reversal
Post	Escalation pathway	”Hand off to human” button

Scenario

You’re a PM at a B2B contract management tool. Your AI will automatically review contracts and flag risks. 3,000 contracts per month: 40% standard NDAs, 35% service agreements, 25% complex custom contracts.

Your options:

Suggestion-only: AI flags risks, lawyer reviews every contract manually
Tiered autonomy: Standard NDAs autonomous with undo, service agreements with preview, complex contracts suggestion-only
Fully autonomous: AI reviews all contracts, lawyers only review escalations

The numbers:

AI accuracy: 96% on NDAs, 88% on service agreements, 74% on complex contracts
Average cost of missed clause: NDA $800, service $4,200, complex $18,000
Current manual review: 45 min per contract, 8 lawyers

Decide

How would you design the interaction model?

The best decision: Option 2 — tiered autonomy.

Why:

NDAs at 96% accuracy and $800 error cost: autonomous action with undo is justified. This saves ~45 min x 1,200 contracts = 900 hours/month.
Service agreements at 88% and $4,200 error cost: action with preview — AI flags risks, shows diff, lawyer confirms. Human-on-the-loop.
Complex contracts at 74% and $18,000 error cost: suggestion-only. AI provides analysis, lawyer decides. Human-in-the-loop.
Invest in post-action: undo for NDAs, audit trail for everything, escalation pathway when confidence drops below 85%.

Service recovery: Build excellent error states. When the AI misses a clause and the lawyer catches it — show what happened, learn from it, report it transparently. This builds more trust than flawless demos.

Reflect

Autonomy is a dial, not a switch. Different tasks in the same product need different interaction patterns. Risk and error cost determine the level.
Design all three phases. Pre-action (what will happen?), in-action (why?), post-action (how to reverse?). Most products only invest in the middle.
Undo beats approval. For low-risk, high-frequency tasks, undo is more efficient than approval flows — and users trust the system more because they retain control.
Invest in failures, not just successes. The service recovery paradox shows: good error handling creates more loyalty than perfect performance.

Sources: Apple Human Interface Guidelines — Machine Learning (2024), Smashing Magazine “Designing AI-Powered Interfaces” (2026), GitHub Copilot Interaction Design, Linear Product Updates, Notion AI UX Patterns, Gmail Smart Compose Research (Google, 2023)