Constraint Box

Map the boundaries first — compute, data, time, ethics, budget.

Advanced9 min read

Overview

The Constraint Box says every solution lives inside limits. Before optimising for performance or novelty, draw the box: what resources, timelines, legal boundaries, and quality bars are non-negotiable? Feasible design starts inside the box, not in an ideal world.

Why it matters

AI projects fail when teams chase state-of-the-art models without asking whether they have the data, budget, latency budget, or governance structure to deploy them. The Constraint Box prevents fantasy architectures and forces honest scoping — the same discipline used in senior engineering and product leadership.

Key principles

Constraints are not enemies of creativity — they define the playable space for good decisions.
Hard constraints (law, security, SLA) differ from soft constraints (preference, convenience) — classify them explicitly.
Tight boxes often produce better solutions than unconstrained optimisation (focus beats brute force).
When constraints conflict, prioritise by risk: safety and compliance before speed, unless explicitly accepted.
The box changes over time — revisiting constraints quarterly prevents building for a world that no longer exists.

How to apply it

1
List constraints in five categories: data, compute/infra, people/skills, time, and governance/ethics.
2
For each AI option, ask: Does it fit inside the box today, or does it require removing a constraint first?
3
Document trade-offs when you touch the edge of the box (e.g. higher accuracy vs 2× inference cost).
4
Use the box in vendor and model selection — cheapest API is wrong if latency violates the product constraint.
5
When teaching, give learners a constraint box for exercises — unlimited problems teach less than bounded ones.

Real-world examples

Startup support bot

Constraints: no dedicated ML team, <200 ms response, must cite sources, GDPR applies. Box rules out large fine-tunes; favours managed APIs + RAG with logging and EU data residency.

School student AI project

Constraints: no budget for paid APIs, school network filters, academic honesty policies. Box favours local models, teacher-approved tools, and assignments that reward process not just output.

Enterprise fraud detection

Constraints: explainability for regulators, 99.9% uptime, retraining only quarterly. Box demands interpretable features, shadow deployments, and human review queues — not whichever model tops a leaderboard.

Common mistakes

Discovering legal or security constraints after build instead of during design.
Treating budget as the only constraint while ignoring data quality or maintenance load.
Removing constraints without stakeholder sign-off ("we will ask forgiveness later").
Confusing aspirational goals with hard constraints — goals motivate; constraints bound.

Key takeaway

Innovation inside a well-drawn box is engineering; innovation that ignores the box is fantasy until proven otherwise.