Sustainable AI in Public Administration: Reducing the Energy Footprint of Your Projects

Why sustainability matters in public AI projects

Public administrations must do more than follow the law: they also set an example. AI projects can consume significant energy (training, inference, storage and data transfer). Cutting that energy footprint brings three concrete benefits: lower operating costs, reduced reputational risk, and alignment with local and national sustainability goals.

In addition, Law 9/2017 on Public Sector Contracts allows sustainability criteria to be included in procurement, so integrating energy efficiency into technical specifications is both compliant and advisable.

Where energy is consumed in an AI deployment

• Model training (if done in-house): the most intensive phase.
• Updates and periodic retraining (continuous learning).
• Inference in production: model calls to respond to users or process batches.
• Storage and transfer of datasets, including logs and backups.
• Supporting infrastructure: orchestration, monitoring, load balancers, etc.

Understanding these layers helps prioritize effective interventions.

Practical levers to reduce the energy footprint

1. Model design and selection

• Prefer compact models optimized for the task: distillation, pruning or specialized models cost less at inference.
• Assess whether a full LLM is necessary or if smaller task-specific models (classification, extraction) suffice.
• Consider hybrid strategies: a local model for quick checks + calls to a large model only for complex cases.

2. Operation and architecture

• Cache frequent responses and use request batching to amortize the cost per inference.
• Design staged workflows: first fast filters (rules / lightweight NLP), then more expensive models only when needed.
• Schedule intensive processing windows (e.g., overnight reindexing) when energy demand or tariffs are lower.

3. Infrastructure and hardware

• Use inference on efficient hardware (TPUs / ML accelerators) when volume justifies it.
• Prefer providers that offer renewable energy or compensation certificates; request transparency on their energy mix.
• Consider on-prem deployments only if they allow reusing existing resources and improving efficiency; in many cases an optimized cloud is more efficient.

4. Public procurement and requirements for vendors

• Introduce energy-efficiency metrics and reporting obligations in tender documents (for example, consumption per 1,000 inferences, energy mix).
• Require continuous improvement mechanisms and technology review clauses to benefit from optimizations.
• Use Law 9/2017 to justify non-economic criteria like sustainability.

5. Governance and metrics

• Define clear operational KPIs: energy consumption per transaction, estimated emissions per function, percentage of requests resolved without calling the main model.
• Incorporate these KPIs into the project lifecycle: design, procurement, production rollout and audit.
• Record and retain traceability for transparency and compliance audits (EU AI Act) and security (ENS RD 311/2022) where applicable.

Quick checklist — actions you can start in 90 days

1. Measure a baseline: monitor instance consumption and number of inferences over one month.
2. Identify one high-traffic AI service to optimize (e.g., citizen chatbot or case classification system).
3. Assess the possibility of replacing a large model with an optimized one or a filter-based architecture.
4. Implement caching and batching where viable.
5. Ask your provider for information on their energy mix and efficiency measures.
6. Add energy-reporting clauses to upcoming tenders/contracts.
7. Define two KPIs (energy consumption per 1,000 inferences; percentage of requests resolved without calling the primary model) and set a six-month improvement target.

Risks and regulatory considerations

• Transparency and auditability: record technical decisions and consumption metrics as part of the AI system documentation (required by transparency and audit obligations).
• Security and data protection: any optimization that changes architecture must be evaluated against ENS RD 311/2022 and the GDPR (for example, when moving data to third parties or storing logs).
• EU AI Act: documentation and risk-assessment requirements still apply; integrating sustainability metrics makes the overall impact assessment easier.

Conclusion and recommended action

Energy sustainability is not an extra: it is an operational and compliance criterion that reduces costs and risks. Recommended action: launch a basic energy audit within 90 days (measurement, selection of a pilot service, cache and model improvements). Integrate efficiency requirements into the next procurement and monitor KPIs quarterly.

If your team needs a structured diagnosis that combines governance, security and energy efficiency for AI projects, tools and consultancies like OptimGov Ready can help prioritize concrete interventions that comply with public-sector regulations.