Copilot vs. Agent Decision
The choice between copilot and agent is one of the most consequential product decisions in an AI feature. Get it wrong and you either frustrate users with an AI that makes them do all the work, or alarm them with an AI that acts without their understanding. This skill provides a structured decision framework, identifies the scenarios where each model excels, and defines the hybrid patterns that cover the space in between.
---
Context
The core distinction:
| Model | What the AI does | What the user does | Trust required |
|---|
| Copilot | Assists, suggests, generates — the user always decides | Review, approve, and execute each step | Low — user is always in control |
| Agent | Plans and executes autonomously — reports results | Define the goal, review the outcome | High — user delegates execution |
FULL COPILOT ──────────────────────────────────────────────── FULL AGENT
AI suggests AI drafts AI executes with AI plans and
what to do the action; confirmation per executes fully;
user decides irreversible step reports when done
Most well-designed AI features sit somewhere in the middle.
---
Step 1 — Define the decision context
Ask:
What is the user trying to accomplish?
What is the user's current workflow?
Which steps are high-judgment or high-stakes?
Which steps are repetitive or rule-based?
What is the cost of the AI making a wrong move undetected?
How much does the user trust AI for this task?
Step 2 — Run the decision framework
Six questions in order:
Q1: Does the task require user judgment at each step? → COPILOT
Q2: Is the task primarily repeatable, well-defined steps? → AGENT candidate
Q3: What is the cost of wrong action undetected? → COPILOT if irreversible
Q4: How established is user trust? → Start COPILOT if new
Q5: Does the task involve external communication? → AGENT WITH CONFIRMATION
Q6: Is the task daily and extremely predictable? → AGENT
Step 3 — Map to an autonomy tier
Five tiers:
Tier 1 — SUGGEST (Pure Copilot): AI generates options; user selects and executes
Tier 2 — DRAFT (Assisted Copilot): AI produces full draft; user reviews and confirms
Tier 3 — EXECUTE WITH CONFIRMATION (Hybrid): AI plans and executes; pauses at high-risk steps
Tier 4 — AUTONOMOUS WITH REPORTING (Supervised Agent): AI executes fully; reports results
Tier 5 — FULLY AUTONOMOUS (Full Agent): AI manages tasks with minimal user involvement
Step 4 — Define tier-specific product requirements
Each tier has specific UX requirements for how AI outputs are presented, confirmed, and monitored.
Step 5 — Plan the trust escalation path
Launch at Tier 1 or 2 (conservative)
Earned autonomy gate: users upgrade after N successful uses or explicit opt-in
Downgrade path: users can always reduce autonomy in settings
Quality check before delivering
Decision framework was followed — not just a gut call
Tier assignment includes reasoning
Trust escalation starts conservative
Irreversible actions are never fully autonomous
User downgrade path is defined
Suggested next step: Present the tier assignment to your engineering and design leads together. The confirmation UX at Tier 3 is where most teams underinvest — design it like you're asking the user to sign off on something real.