Map Product Opportunities Across 5 Core AI Capabilities
Accurately name the 5 foundational AI capabilities (Classification, Ranking, Generation, Extraction, Agentic), contrast technical maturities, and trace compounding errors.
Map Product Opportunities Across 5 Core AI Capabilities
Before conducting a Feasibility Assessment, the fundamental responsibility of a Product Manager is to accurately name the specific AI capability required. Each technical capability represents a distinct technological maturity level, requires different data structures, and carries unique failure modes. Misclassifying the core problem archetype at the start is the most common reason an AI initiative looks feasible on paper but collapses during engineering execution.
Running Example: E-commerce and logistics platform FinTrack Logistics - evaluating various AI initiatives: automated refund fraud detection, personalized seller recommendations, invoice extraction, and autonomous customer resolution agents.
Flawed Approach: Generic 'Use AI' Label
Production Approach: 5 Core Capabilities
1. Name the Exact Problem Archetype Instead of Generic "AI" Labeling
The term "AI" is far too broad for product roadmapping. Modern AI product engineering recognizes 5 core capability archetypes:
- Classification / Prediction: Transforms raw inputs into discrete category labels or probability scores. Example: "Is this outbound fund transfer fraudulent?"
- Ranking / Recommendation: Orders a set of candidate items based on contextual relevance for a specific user. Example: "Which suppliers should be featured on this merchant's dashboard?"
- Generation: Synthesizes net-new multi-modal content (text, image, audio, code) conditioned on instructions. Example: "Generate compelling product descriptions from raw catalog images and specs."
- Extraction / Retrieval: Locates and pulls exact structured fields from large unstructured corpuses without fabricating new facts. Example: "Extract penalty clauses and indemnity limits from an 80-page supplier contract."
- Agentic / Orchestration: Executes multi-step, goal-driven action workflows in external environments, making autonomous branching decisions. Example: "Autonomously resolve end-to-end customer return tickets: parse email, query ERP database, issue refund API call, and dispatch notification."
2. Compare Technological Maturity and Failure Modes Across 5 Capabilities
Each capability archetype carries distinct engineering maturity and operational risk profiles:
| Capability Archetype | Technology Maturity | Mandatory Input Data | Typical Failure Mode |
|---|---|---|---|
| Classification | Highly Mature (Decades) | Large, clean historical labeled dataset | Label error (False Positive / False Negative) - Quantifiable and tuneable |
| Ranking | Mature | User interaction and clickstream logs | Rank drift - Hard to detect immediately, silently degrades conversion |
| Generation | Rapidly Evolving | Domain context and system prompts | Hallucination - Generating fabricated facts with high confidence |
| Extraction | Mature | Structured/unstructured source corpus | Extraction misses or out-of-context parsing |
| Agentic | Early Stage | All 4 above combined + Tool access APIs | Compounding execution errors - Single step failure breaks whole chain |
Organizational Analogy: Mapping capabilities is like hiring specialized staff:
- Classification is a gatekeeper: making binary In/Out decisions.
- Ranking is a merchandising manager: organizing store shelves for optimal visibility.
- Generation is a copywriter: drafting marketing assets from core product briefs.
- Extraction is a paralegal: scanning files to extract exact legal clauses.
- Agentic is an autonomous operations associate: executing entire workflows without step-by-step supervision. As you move down the list, failure modes multiply and root-cause tracing becomes exponentially harder.
Map Product Opportunities Across 5 Core AI Capabilities
1. Pattern Recognition & Classification
Benchmark: Precision & Recall ≥ 98%Categorizes raw unstructured inputs into defined labels
Tags raw bank transactions ('GrabFood $4') into [Dining & Food].
3. Trace Error Propagation Boundaries from Gatekeeper to Autonomous Agent
When evaluating feasibility, PMs must account for Compounding Error Rates.
In a standalone Classification task operating at 95% accuracy, your risk surface is 5%. However, in an Agentic system requiring 5 consecutive steps (Triage ticket → Extract ID → Retrieve policy → Draft reply → Execute payment API), if each step maintains 95% reliability, overall system reliability drops to (0.95^5 approx 77.4%). Over 22% of production runs will fail without human approval gates.
4. Deconstruct Real Product Scenarios into Distinct Capability Layers
Exercise 5.1: Analyze the following 4 FinTrack Logistics product initiatives, identify the core Capability archetype (or combination), and highlight the primary failure mode to guard against:
- Refund Fraud Early-Warning System: Automatically flags suspicious claims before funds are disbursed to buyers.
- Dynamic Carrier Dispatch Recommender: Evaluates hundreds of partner fleets to recommend the top 3 carriers by cost and transit time.
- Legal Precedent Research Assistant: Lawyers input dispute terms; AI searches and extracts matching appellate rulings from judicial databases.
- Autonomous Credit Limit Approval Bot: Ingests dealer credit requests, queries banking APIs, computes risk scores, and autonomously executes credit upgrades in accounting software.