Why Use AI Only When It Solves Real Problems: A Practical Scenario‑Analysis Guide for AI Product Managers

Scenario definition

Scenario = (time + physical environment + psychological state) + specific role + concrete goal. Example: a 75‑year‑old living alone who measures blood pressure at night versus the same need in a hospital setting. The former requires large fonts, loud voice prompts; the latter needs precise data capture and silent alerts to remote relatives.

Why scenario analysis matters

Concrete requirements & team alignment – storytelling unifies engineers and stakeholders.

Edge‑case discovery – anticipates network loss, bright light, one‑hand operation.

Interaction‑path optimization – voice feedback on a highway vs keyboard input in an office.

Data planning & instrumentation – early identification of needed permissions and future model‑training data.

Traditional 5‑step scenario‑analysis SOP

Step 1 Find the protagonist (define role boundaries)

Wrong: "20‑35‑year‑old urban white‑collar".

Right: "A junior programmer who works 996, lacks security, has a compulsive need to organize information, and spends an hour commuting daily."

Key tip: explicitly state who is NOT the user.

Step 2 Define goals (surface explicit & implicit needs)

Explicit: "I want to chat to pass time."

Implicit: "I feel lonely and need emotional acknowledgment."

Success metric: sending the first message or completing a 5‑minute conversation.

Step 3 Reconstruct the context (wear VR glasses)

Physical environment: dim lighting, noisy background, one‑hand operation.

Time & state: post‑work fatigue vs bedtime impulsivity; 5G vs weak elevator signal.

Step 4 Map the interaction path (microscopic action dissection)

Pre‑scene: what the user does a second before opening the app?

Granularity: record every click, swipe, input.

Path audit: identify redundant steps that cause churn.

Step 5 Find opportunities (output an action plan)

Example: users typing late at night dislike keyboards → provide voice input, one‑click quick replies, then log the need as a P0 item.

AI‑specific extensions – five higher‑order dimensions

Dimension 1 Capability‑Scenario Matching

Ask, "What can current AI actually achieve in this pain‑point?" Example: senior‑care health monitor uses silent data collection + backend alerts rather than a chatty assistant. Signals indicating a good AI fit include:

Positive ROI – token/computation cost lower than the value delivered.

Regular, learnable patterns (e.g., personalized recommendation, auto‑tag generation).

High manual cost or speed bottleneck (e.g., extracting minutes from a 3‑hour recording).

User tolerance for imperfect results (e.g., AI‑generated poetry).

Abundant, compliant training data.

Dimension 2 Failure‑Scenario Design (graceful degradation)

In high‑risk domains (medical advice, legal contracts) perform risk exposure assessment, error perception, and mandatory fallback to human review with clear disclaimer UI.

Risk exposure assessment – quantify actual loss if AI hallucinates.

Error perception mechanism – how does the user notice a mistake?

Graceful degradation – switch to traditional search or template when generation fails.

Human‑in‑the‑loop – provide edit, regenerate, or feedback buttons.

Confidence display – UI shows "AI result for reference only" or highlights uncertain facts.

Dimension 3 Human‑AI Boundary Definition

Define authority level per domain. Health monitoring: AI acts as a Copilot, final decision rests with doctor or family. Tarot‑reading app: AI can be Autopilot because hallucinations add charm.

Dimension 4 Data‑Scene Closed Loop

Plan data flow from the start:

Cold‑start solutions – purchase data, rule‑based guards, or gamified onboarding to collect initial signals.

Data flywheel – each user interaction (click, edit, dwell) feeds a knowledge base for Retrieval‑Augmented Generation, lightweight fine‑tuning, and dynamic prompt iteration.

If a scenario consumes model capacity without generating high‑quality data, the product lacks a defensible moat.

Dimension 5 Scenario Evolution

MVP scenario – current models only summarize text.

Expanded scenario – multimodal models generate PPT from summaries.

Deep scenario – AI auto‑assigns follow‑up tasks based on meeting minutes.

One‑page canvas for AI scenario analysis

Fields to fill:

Scenario name & trigger condition.

User state, physical environment, psychological pressure.

Expected outcome.

AI involvement mode (Autopilot, Assist, Human‑lead).

Required model capabilities (NLP extraction, CV recognition, multimodal reasoning).

Data source & compliance.

Fault tolerance & failure‑scenario preview.

Fallback plan (human edit, alternative UI).

Data feedback design (what behavior is logged to improve the model).

Core success metrics – DAU, task‑completion rate, token‑cost ROI.

Signals for AI‑fit vs anti‑fit

Fit signals – positive ROI, regular learnable patterns, high manual cost, tolerance for imperfect output, abundant compliant data.

Anti‑fit signals – zero‑tolerance domains (e.g., medical diagnosis, high‑value legal contracts), scarce or highly sensitive data, decisions requiring strong accountability.

Code example

-AI主导 (自动执行) /
-AI辅助 (提供建议) /
-人主导 (信息提供)