Irreplaceable › Guides › AI design frameworks
Frameworks · AI product design
Frameworks for AI product design decisions
AI can make the screen. It can't decide how much an agent should do alone, which metric would hide a failure, or where friction belongs. These 6 short frameworks help you make those calls.
By Sarit Elisha, founder of Irreplaceable · Updated October 10, 2026
The frameworks with the checklist, AI failures and senior takes, searchable. Free.
01 · AI behavior design
The autonomy dial
Decide what an agent does alone, with undo, after asking, or never.
- Cost: what does a mistake cost the user, in money, time or face?
- Reversibility: can it be undone completely, and for how long?
- Who else is affected: does the action reach anyone besides the user?
- Set the level: low on all three → just do it; reversible → do it with undo; costly or touches others → ask first; irreversible, legal or medical → never.
Worked examples: what should the agent do on its own?
| Action | Level | Why |
|---|---|---|
| Email agent: Archive newsletters you never open | Just do it | Low stakes and easy to find again. |
| Email agent: Unsubscribe you from mailing lists | Do it + undo | Reversible, but annoying if wrong. Do it and offer an undo. |
| Email agent: Draft replies to customers | Just do it | A draft is only a suggestion; nothing leaves your inbox. |
| Finance agent: Categorize this month's expenses | Just do it | Pure organizing, trivially fixable. |
| Finance agent: Pay the electricity bill, same as last month | Do it + undo | Routine and expected. Pay with a cancel window. |
| Finance agent: Send the quarterly report to your accountant | Ask first | Sharing financial data outside. Confirm the recipient. |
02 · Owning outcomes
Metric pairs
Never ship an AI feature with an adoption metric alone.
- Adoption (used, accepted, clicked) is paired with quality (edited after, undone, reopened, returned next week).
- Speed (handle time, time to reply) is paired with resolution (solved, no repeat contact).
- Engagement (searches, session length) is paired with goal reached (purchase, task done).
- Add a guardrail: the number that must not get worse, agreed before launch.
03 · Trust
Friction scaled to stakes
Uniform friction is ignored. Put it only where a wrong answer hurts.
- Sort answer types by harm: trivia, preferences, money, health, legal, safety.
- Low stakes: clean answer, no disclaimer.
- High stakes: sources, a clear limit, a human or professional route.
- Measure the behavior: how often people verify high-stakes answers.
04 · Problem framing
The reframe canvas
Turn a solution request into a problem worth solving.
- Write the ask in their words, and the fear behind it.
- Find one fact you already have that points to the user's real problem.
- Pick a metric that measures the problem, not the feature.
- Choose the cheapest first move that could prove you wrong this week.
05 · Influence
Win the room
How seniors stop bad ideas without becoming the blocker.
- Speak their metric first: deadline, revenue, demo.
- Offer a safer scope that still gets them a yes.
- Bring one concrete example, not a principle.
- Attach a guardrail metric you'll own after launch.
06 · Judging AI output
Reviewing AI output
A five-pass check before anything AI-made ships.
- Constraints: did it respect what the user actually asked for?
- Claims: is every fact sourced, and is uncertainty shown?
- Actions: did it act before asking on anything costly or shared?
- Exits: can the user undo, edit, or reach a human?
- Harm: who could this hurt, and does it hide that in fine print?
From framework to reflex
A framework helps when you remember to use it.
Drills give you the decision under a little time pressure, until the framework is how you think.