Google Cloud Architecture Framework
Primary architecture guidance on reliability, performance, security, cost, and operational requirements.
← Back to Compute College
Compute College
Define functional and nonfunctional requirements
Separate what an AI system must do from how well, safely, quickly, and reliably it must do it.
Plain-English definition
Functional requirements describe the result an AI system must produce; nonfunctional requirements describe the operating boundaries around that result, such as latency, availability, privacy, throughput, cost, and safety. Both are needed to choose a model and deployment because a function without constraints cannot be sized.
Memory trick: Function says what; nonfunction says under what conditions.
Why it matters
Teams often debate models before agreeing on what “good enough” means. Requirements make tradeoffs explicit: a slower model may be acceptable for batch work, while an interactive assistant may need lower tail latency even if that requires more capacity or a smaller model.
- Functional requirements define accepted outcomes and failure behavior.
- Nonfunctional requirements define service quality and operating limits.
- Priorities reveal which tradeoffs are acceptable when no design meets every goal.
Simple example
A claims assistant must extract policy fields and cite source text (functional), respond within 5 seconds at p95, keep data in an approved region, maintain 99.9% availability, and stay below a defined cost per accepted claim (nonfunctional).
- The output is testable instead of “help users faster.”
- Latency, region, and availability can rule out otherwise capable options.
- Cost is tied to accepted work rather than raw calls.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Primary reference
This source explains the method or risk boundary. It does not replace testing the actual model, data, provider, and workload conditions in production.
Source discipline: examples on this page are illustrative; measure the real workload before making a release or capacity decision.
Common mistake
Leaving nonfunctional requirements until implementation can invalidate the design. A privacy boundary or latency target may change the viable model, provider, and capacity plan from the start.
Practical takeaway
What you can do with this
Write a two-column requirement sheet for one feature. Add a measurable threshold, priority, test method, and owner to every important requirement.
- Mark must-have versus negotiable requirements.
- Tie each requirement to a test or operational metric.
- Revisit the sheet when traffic or user impact changes.
Decision check: if two designs produce the same answer quality, can the requirements show which one is acceptable to operate?
Compute College learning path
AI Engineering
Step 38 of 48: Define functional and nonfunctional requirements