Chapter 0.1 — Model, product and system¶
🎯 Chapter objective¶
Distinguish three levels of abstraction that are often confused: model, product and system — and show why this distinction is the root of the most expensive mistakes in enterprise AI.
🧠 Core concept¶
- A model is a function that maps inputs to outputs. It has weights, an architecture, a competence boundary, assumptions about the data, and a validity window.
- A product is the application consumed by a person or another system. It has UX, API contracts, business rules, SLAs and context.
- A system is the full operational whole: product + model + data services + evaluation services + security + governance + cost + operation + lifecycle.
The central thesis of this book is that real AI value appears in the system, not in the isolated model.
🏗️ How this shows up in production¶
A classifier with 98% offline accuracy can fail in production because:
- production data differs from training data;
- the most informative feature is only available with a delay;
- the downstream consumer interprets the output incorrectly;
- there is no rollback if quality drops;
- there is no accountable person when an incident happens.
None of these problems is solved by swapping the model.
⚖️ Trade-offs¶
| Dimension | "Model" focus | "System" focus |
|---|---|---|
| Time to demo | Short | Longer |
| Time to production | Usually long | Shorter |
| Maintenance cost | High and hidden | Visible from the start |
| Operational risk | Concentrated in incidents | Distributed and mitigated |
| Success at scale | Unpredictable | Predictable |
🚨 Failure modes¶
- "Works in the notebook" as proof of production-readiness.
- Offline metrics disconnected from the business KPI.
- Model swapped without behavior regression testing.
- Upstream data failure causing silent degradation.
- A good model serving the wrong data (cross-tenant, old version, different language).
🛡️ Controls and mitigations¶
- Treat AI as critical software: versioning, tests, observability, CI/CD, rollback, runbooks.
- Document accepted failure modes and unacceptable failure modes.
- Distinguish what is the responsibility of the model, the product and the system.
🔭 How to observe and measure¶
- Technical metrics: latency, throughput, error rate, cost per call.
- Quality metrics: accuracy, F1, groundedness, recall@k.
- Business metrics: conversion, satisfaction, ticket deflection, MTTR.
- Governance metrics: policy violations, DLP events, pending approvals.
🧪 How to test¶
- Unit and integration tests in the product.
- Eval harness (offline and online) on the model.
- Replay of real traces.
- Canary deploy for changes.
🧰 Related practical example (planned)¶
EX-FUND-01— compare the output of an isolated model vs the same model embedded in a product with pre-processing and post-processing.
📌 Checklist¶
- [ ] Does the system have an owner for model, product and operation?
- [ ] Is there documentation of the input/output contract?
- [ ] Is there an upstream data validation pipeline?
- [ ] Is there a rollback process?
- [ ] Is there a runbook?
📚 References¶
- Evidently AI — Concept drift and model operations: https://www.evidentlyai.com/ml-in-production/concept-drift
- NIST AI RMF 1.0: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf