NICKE STUDIO / Choosing and evaluating an AI agency
The environmental impact of hiring an AI agency
Assess an AI agency's environmental impact through energy, carbon, hardware and water. Ask for defined boundaries and clearly identified estimates.
Read in another language ↗
An AI agency's work runs on physical infrastructure. Model inference, databases, storage and network services consume energy; hardware also has manufacturing and end-of-life impacts. Buying a digital service does not remove those effects. At the same time, a universal claim about the footprint of one prompt cannot reliably describe every model, deployment and workflow.
The useful procurement question is what the proposed system consumes to deliver an accepted business outcome, which parts can be measured and which remain uncertain. That leads to better decisions than either assuming AI is impact-free or assuming every possible application has the same environmental cost.
Define what the assessment includes
The Software Carbon Intensity specification distinguishes operational emissions from the allocated embodied emissions of hardware, then relates them to a functional unit. For a business workflow, that unit could be an accepted support response or a completed document review.
Define the boundary before comparing proposals. Does it include model calls, retries, storage, database operations, monitoring and reserved idle capacity? Are upstream model training or hardware impacts included, allocated or unavailable? Two numbers with different boundaries are not a fair comparison.
| Dimension | What to request | What to avoid claiming |
|---|---|---|
| Energy | Consumption for a defined period and workload | A universal number for every prompt |
| Carbon | Energy assumptions, regional intensity and hardware allocation | That electricity and emissions are interchangeable |
| Water | Provider information with location and a clear definition | A fixed water quantity for every response |
| Hardware | Reserved resources and a declared lifecycle boundary | That a local model has no physical footprint |
Water withdrawal and water consumption describe different things. An assessment of carbon alone is not a complete assessment of water use, electronic waste or wider environmental effects. Label what is outside the calculation rather than hiding it behind a sustainability badge.
Use a transparent worked example
Here is a hypothetical calculation, not a measurement of Nicke Studio or any provider. Suppose a defined workload uses 10 kWh, with an assumed intensity of 200 grams CO₂e per kWh and 1,000 grams of allocated hardware emissions. It delivers 1,000 accepted tasks.
Operational emissions are 10 × 200 = 2,000 grams. Adding the allocated hardware gives 3,000 grams. Dividing by the 1,000 accepted tasks gives 3 grams CO₂e per accepted task within that boundary. Changing the assumptions changes the result; the number is not a universal benchmark for AI.
Record the workload and all failed attempts in the numerator. Otherwise, a system can look efficient by reporting only its successful executions while ignoring the retries that consumed resources.
Distinguish measurements from estimates
CodeCarbon measures or estimates aspects of local computing and combines energy estimates with regional carbon-intensity information. Its documentation explicitly distinguishes local hardware from remote generative-AI APIs. Running a tracker on an agency laptop does not measure the complete remote model service.
For a remote API, ask which provider data or estimation method is being used, what hardware assumptions it makes and which components it excludes. If information is unavailable, report that gap. An estimate can still guide improvement when its method and limitations are clear.
An honest environmental report shows its boundary, its assumptions and the decisions those numbers can support. — Nicke Studio editorial note
Reduce unnecessary work before changing suppliers
Practical improvements often start with the workflow:
- Use ordinary code for deterministic rules and exact lookups.
- Avoid sending unchanged documents through the model repeatedly.
- Limit unnecessary output and retrieve only relevant material.
- Test whether a smaller model meets the same acceptance criteria.
- Remove idle infrastructure and unnecessary duplicate processing.
- Evaluate caching or batching where privacy and freshness allow it.
Recheck quality after every change. A smaller model that triggers repeated corrections may increase total resource use. The tools guide and results guide help compare complete outcomes rather than isolated requests.
Ask about totals as well as efficiency
A lower footprint per task does not guarantee a lower total footprint if the number of tasks grows sharply. Track both the intensity of accepted work and the total workload. Claims about avoided travel or reduced printing also need a plausible comparison with what would otherwise have happened.
The SCI specification does not treat carbon offsets as reductions to the SCI score. Keep any compensation claim separate from actual changes in energy use, infrastructure or emissions within the chosen boundary.
Before hiring, ask who owns this reporting, when it will be updated and whether the agency can explain the limits without promising an impact-free system. Explore our services, review the red flags checklist and send your workflow if you want to discuss a smaller, measurable implementation.