Area 8.1

Choose the Smallest Model That Does the Job

Are we using a far larger model than this task actually needs?

Baseline: 1Enhanced: 1 2 controls

Who it applies to

Enter for every AI system. Go as deep as your system letter, and implement in your mode.

Why it exists

Match the model to the task, and stop reaching for the largest available by default — because the biggest models cost the most to run, in money and in energy, and most tasks do not need them. The reflex to use the most capable model for everything is expensive on every axis: a smaller model that handles the task well enough is cheaper per use, faster, easier to run, and lower-impact. Right-sizing is the single highest-leverage sustainability decision in AI, and it is made not by the sustainability team but by whoever chooses the model.

What to actually do

Select the smallest model that meets the requirement, not the most capable available. Establish what the task actually needs and choose the model that meets it with margin, not the largest one on the shelf. A large model doing a small model’s job is waste — of money and of energy — dressed up as prudence.
Make right-sizing part of the model-selection decision. Fold this into the approved-model and sourcing decisions of the earlier domains, so the environmental and cost consequence of model size is weighed at the point of choice, not audited afterwards.

Controls

Baseline 1 controls

Control 213

Select the smallest sufficient model for each use, by default.

BaselinePreventiveall rolesper system

What it is

The most capable model is a justified exception, not the starting point.

How to implement it · Minimum rules

Define 'sufficient' first: 3–5 concrete task requirements (e.g. language, document length, acceptable error on sample cases)
Default: start from the smallest model or service plan the provider offers for that task
Test the smaller option on the same 10–20 real sample cases before considering a larger one
Larger model only by exception: one written line naming the requirement the smaller option failed
Re-check at renewal or when the provider releases new model sizes

How to start

For each AI use already in the inventory, note which model or plan it runs on and whether a smaller option exists.

Evidence

The model chosen per use, its sample-case result and any written exception for a larger model.

Mode: light / standard / full

light: The tool's owner tries the smaller option on a few real cases and notes it in the inventory spreadsheet.
standard: Each team runs the sample comparison and records it; IT reviews exceptions.
full: Default model sizes are set centrally per task type; larger models need an approved exception.

Enhanced 1 controls

Control 214

Weigh model size in the selection decision.

EnhancedPreventiveall rolesper system

What it is

Cost and environmental consequence considered where the model is chosen.

How to implement it · Minimum structure

Task requirement and the acceptance threshold any model must meet
Candidates compared: at least the smallest viable model and the one proposed
Per candidate: measured quality on the evaluation cases, cost per 1,000 requests, response time
Environmental indicator: relative compute or energy per request, from the provider or estimated from model size
Decision, with written justification wherever a larger model is chosen
Approver, date and re-selection triggers (new model sizes, large change in volume)

How to start

Add a 'model size, cost and energy' section to the existing approved-model or sourcing decision template.

Evidence

Completed selection records showing sizes compared, and the justifications where a larger model won.

Mode: light / standard / full

light: Whoever chooses the tool fills a half-page version from the provider's published prices and sizes.
standard: The system owner completes it; the selection approver checks it before sign-off.
full: A platform function maintains size benchmarks and reviews every selection record.

Evidence

The model-selection records showing size was weighed; the justifications where a larger model was chosen.