Yes. Sustainability should influence the choice of AI models and tools, alongside the requirements that make the work usable. Start with a defined staff task, establish which options can complete it acceptably, then compare environmental evidence that describes those options under relevant conditions. A smaller model or a better energy rating is a reason to investigate an option, rather than permission to lower the quality bar.
For an adoption leader, the useful decision is usually narrower than naming the company's “greenest AI”. You might choose a supported configuration for one returns workflow, request a missing measurement or keep a particular task on its existing route. Record what the evidence supports and which environmental questions remain open.
Define the job before comparing the models
Suppose Sam Taylor coordinates returns at an events-equipment hire company. A lighting flightcase comes back with a return authorisation, an equipment packing list and a technician's note. Sam wants an assistant to prepare a return checksheet that a depot reviewer can inspect against those records. The sheet does not approve a refund or certify the equipment as safe to hire again.
Two candidates can produce an orderly-looking sheet. Their usefulness depends on whether they preserve the actual case identifier, equipment serials, quantities and unresolved discrepancies. A change from LX140 in the source to LX104 in the draft matters more than fluent explanations around it.
Before the trial, agree the necessary conditions:
- Correct records. The checksheet preserves identifiers and quantities, links each entry to its source and leaves conflicts visible.
- Permitted information. The task uses the approved inputs, access and handling arrangements.
- Usable timing. Staff can complete checking and handover within the returns process, including any waiting.
- Supported operation. Someone owns configuration changes, failures and the ordinary fallback.
Microsoft's sustainable AI design guidance recommends considering smaller models where applicable while explicitly warning about context, complexity and accuracy tradeoffs. That qualification belongs in the purchasing decision. A model that misses a necessary part of Sam's work has not earned a routine role by using fewer resources in a different test.
The enterprise assistant comparison guide separates handling a common source packet from completing the ordinary supported workflow. Keep that distinction here. A strong result on supplied documents does not establish the performance of a route that also has to find, extract and transfer them.
Use energy benchmarks to build a shortlist
Public benchmarks can reveal options worth testing and questions a supplier should answer. Read their methods before treating a rating as a result for your own service.
The AI Energy Score methodology describes controlled inference comparisons focused on GPU energy. Its stars are relative to the task, applicable text-generation class and leaderboard issue. Performance measures such as accuracy and latency still need consideration. Those stars do not establish Sam's accepted-checksheet rate or the complete environmental footprint of a purchased tool.
A useful evidence record keeps the result attached to its conditions:
| Evidence | What it can help you investigate | What still needs checking |
|---|---|---|
A task-specific energy rating | Which candidates deserve a closer look | The tested class, hardware, configuration and relevance to your work |
A controlled inference benchmark | How a configuration changes energy and timing | Whether your deployment and workload reproduce those conditions |
A supplier's service report | Resources attributed to the purchased service | Coverage, allocation, exclusions and compatibility with the comparison |
In its January 2026 analysis of ML.ENERGY Benchmark v3.0, the research team shows why the surrounding system matters. Output length, hardware utilisation, batching and runtime choices affect measured GPU energy. Processing more requests together can improve energy efficiency while increasing response latency. The findings describe tested configurations, rather than a universal ranking of hosted products.
Treat model size as one characteristic to examine. An offer's name, licence price or advertised response speed cannot supply the missing measurement boundary.
Check quality on the work the team will receive
Shortlisting should include alternatives to the largest general-purpose model. It should also leave room for evidence that a proposed smaller alternative cannot yet handle the task.
Sasha Luccioni's May 2025 comparison of models on report questions illustrates the value of task-specific evaluation. The exploration used generated questions about three reports, an AI judge and energy approximated from hardware power ratings and response time. It offers a reason to test different candidates, with important limits: estimated energy and judged answers in that experiment do not validate an employee's actual workflow.
For Sam, include the records that make the job difficult, such as a packing list with similar serial numbers and a return note referring to only part of a case. Have the responsible reviewer define what counts as an acceptable checksheet. Do not quietly replace those requirements after discovering which candidate struggles.
Suppose candidate A preserves the requested layout but changes LX140 to LX104. Candidate B preserves the identifiers in the packets examined so far. A needs investigation before routine use; B may proceed only within whatever limited trial its owners have accepted. Neither observation establishes an environmental winner.
Keep task evidence separate from the environmental decision
Case LX140
The return authorisation identifies the case Sam is checking.
Identifiers verified
The draft says LX104 while the source says LX140. Investigate the mismatch.
Tested packets matched
Preserved identifiers support only the bounded quality observation, not routine approval.
Comparison still open
Obtain comparable evidence for the complete supported route.
The point is to keep two decisions visible. Sam can reject an unsupported claim that A's fields are verified while leaving the energy comparison unresolved. A well-presented answer should not turn a failed task check into an acceptable tradeoff.
Count the route through an accepted result
The first model response may be only one part of completing the work. Record relevant document processing, additional requests, retries, tool use and any rescue through another model. Keep abandoned cases and manual completion visible too.
In an October 2026 LinkedIn post, web-agency technical lead Romain Avalle described inspecting an agent session's token counter and noticing repeated repository and context reads. His observation concerns his own development session, not a measured footprint for Sam's service. It exposes a useful question: does the evidence cover one call or the activity needed to finish the job?
Ask the technical and environmental owners to retain:
- The configured route. Record model and tool versions, settings, document processing and any fallback.
- The work attempted. Keep the relevant inputs, output lengths, retries and cases that did not finish through the proposed route.
- The accepted result. Apply the same checksheet requirements and preserve reviewer corrections and unresolved cases.
- The measurement boundary. Identify included equipment and services, the period, allocation choices and material exclusions.
- The operating conditions. Record relevant load, batching, waiting and the deployment information available to the assessor.
The Software Carbon Intensity specification, version 1.1.0, requires an explicit system boundary and a consistent functional unit for its carbon-intensity method. Applying that discipline here means defining the useful work before comparing rates. A rate per generated field and a rate per accepted checksheet answer different questions. A partial GPU-energy observation also remains different from a full carbon-intensity assessment.
Keep human checking time as a separate operational measure. It matters to the choice, but minutes cannot be converted into emissions without suitable evidence. The employee AI footprint guide explains how to preserve coverage and allocation limits when an environmental estimate is available.
Make a conditional choice the team can support
A smaller local model may require different infrastructure, maintenance or workflow design. Those commitments belong in the decision, alongside any supported resource advantage.
A Reddit user developing a game encountered that practical tension in a February 2026 discussion about moving from an API to a small local model. After others suggested changing the surrounding architecture, the developer said the approach was more complex than the time available, while welcoming a separate suggestion to prepare dialogue ahead of use. This unverified game-development account does not establish enterprise energy performance. It makes the support question concrete: who will maintain the arrangement that makes the smaller model useful?
For the returns workflow, take the decision in this order:
- Exclude or investigate candidates that fail essential task, information or support requirements.
- Compare relevant environmental evidence for the remaining supported routes, retaining uncertainty and differences in coverage.
- Record the chosen scope, unresolved questions and the owner who will obtain missing evidence.
- Set a review trigger for material changes to the model, configuration, workload or supplier method.
Where both routes meet the necessary requirements, a credible environmental advantage can help determine the choice. Where evidence is incomplete, it may support a bounded operational trial while the environmental comparison stays open. Do not publish a carbon-saving claim merely because a cheaper or smaller option was selected.
For Sam, the next useful step could be a request for comparable service evidence for B, with A's identifier problem retained for investigation. The company can continue learning without pretending the choice has answered every environmental question.
Questions about sustainability and AI model choice
Is a smaller AI model always the greener choice?
No. Model size can inform a shortlist, but the task, configuration, hardware and complete route affect resource use. A smaller model that needs extra processing or repeated rescue calls must be assessed with that activity included. First establish that it produces acceptable work, then compare suitable environmental evidence for the supported deployment.
Can an energy star rating choose the model for us?
An energy rating is useful shortlist evidence within its stated method. The AI Energy Score FAQ explains that its stars are relative to specific benchmark conditions and leaderboard issues. Check the applicable task and class, then test your quality, timing and information requirements. A high rating does not by itself approve a company workflow or establish its total carbon footprint.
Is the cheapest API also the lowest-carbon option?
A price is a commercial charge, not a carbon measurement. Suppliers can use different pricing, infrastructure and allocation arrangements. Request relevant resource or emissions evidence with its method and exclusions. Keep the cost comparison and environmental assessment separate until the data supports a meaningful connection.
What if a supplier cannot provide comparable environmental data?
Record the missing evidence and the decision it prevents. The organisation may still make a limited operational choice after its necessary requirements are met, but it should leave the environmental ranking unresolved. Name an owner and a review trigger, and use the footprint-boundary checks to make the supplier request specific.
Should we automatically try a small model before a larger one?
Test that route before making it the default. A smaller-first approach can add initial calls, waiting and fallback activity when it cannot complete the task. Compare the complete supported route against an appropriate alternative using the same accepted result. Define when to escalate or stop, and retain the ordinary fallback instead of asking employees to keep retrying indefinitely.
