--- title: "Responsible AI use policy · Carange Solutions and Munia" description: "Our commitments to a frugal, fair and sovereign municipal AI, organized along the three ESG pillars: clean hosting, a person in the loop, protected data." url: https://munia.ai/en/responsible-ai-policy/ lang: en-CA last_updated: 2026-09-29 --- Responsible AI use policy · Carange Solutions and Munia # Responsible AI use policy Our commitments to a frugal, fair and sovereign municipal AI, organized along the three ESG pillars. From the first prototype to the retirement of an agent, the same rules apply. Version of September 29, 2026 · Owner: Anthony Carange, President Read our commitments [See the Security page](/en/securite/) 01 Scope ## Who this policy applies to. From the first prototype to the retirement of an agent, the same rules apply. ### Our team All Carange Solutions employees and our subcontractors. ### The Munia platform Every assistant and agent deployed for a municipality, an MRC or a public body. ### Our training The practices we teach the organizations we support. 02 Our ESG commitments ## Three pillars, concrete commitments. What guides each of our decisions, from design to deployment. ### Environment A frugal AI that uses as little energy as possible for each answer. #### Clean-energy hosting Munia's data is hosted at OVHcloud in Beauharnois, one of the most efficient data centres in the industry, water-cooled and powered by Quebec hydroelectricity. 1.21 PUE in Beauharnois industry average: about 1.55 100% renewable energy REF, fiscal year 2025 [See OVHcloud's environmental indicators](https://corporate.ovhcloud.com/en/sustainability/environment/) #### A deliberate choice of models You do not always need the most powerful model. Through the Munia gateway, each request is routed to the model that offers the best answer quality with the lowest environmental impact. −75% compute per message since the Munia gateway 109k → 27k tokens processed on average per message Munia internal data: May 27 to 31, 2026 (before the gateway, 13,897 messages) compared with August 31 to September 27, 2026 (75,659 messages). #### Digital frugality Our agents rely on data made ready for AI. For example, rather than loading a full zoning grid into the context for every question, we transform it once into a format the model digests more easily. −98% tokens consumed on a zoning grid Industry benchmark: Google measured 0.24 Wh and 0.26 mL of water per median query in 2025, with energy per query 33 times lower than a year earlier. Measuring is how you reduce. ### Social An AI that serves residents and municipal employees, never works against them. #### Transparency A resident always knows they are interacting with an AI. Answers cite their municipal sources. #### Tested agents, in plain language All our agents go through a testing process before deployment, and answer in clear French accessible to everyone. 100% tests passed targeted before deployment 95%+ quality score targeted average #### Support by people We are an AI company, but we do not hand our support to a chatbot. Technical support is provided by five people from our training and development teams: Jean, Alexandre, Anthony, Simon and Youssef. #### Training and best practices In our training sessions, we teach how to use AI to be more productive. We also devote a section to the stakes: the impact of AI on our work, on the environment and on our cognitive functions. What we apply internally, we teach externally. ### Governance A supervised, verifiable AI under Quebec and Canadian jurisdiction. #### A person in the loop, with approval The platform supports the user in their decision-making. No action is taken without their approval, and it never makes a decision without their knowledge. #### Data protection first Data protection is at the heart of all our strategic decisions. A real-time anonymization gateway removes residents' personal information before each request: it is never received by an inference model that is not ours. An inference provider is only selected if it offers zero retention, and it only receives already anonymized data. Law 25 respected by default. #### Hosting in Quebec Your data is hosted in Quebec, with a non-American provider, under Quebec and Canadian jurisdiction. #### Security AES-256 encryption at rest and TLS 1.3 in transit, SSO authentication and access logging. 03 Our limits ## What we will never do. Clear lines, valid for every project and every client. ### Train models with your data without your authorization Your resolutions, minutes and emails remain yours. ### Let the AI decide alone No permit or service refusal without human intervention. ### Monitor residents No facial recognition, no profiling of individuals. ### Sell or share your information Your data is never sold, rented or shared with third parties. ### Use prohibited tools We comply with the AI tool prohibitions issued by the Government of Quebec. 04 Reference frameworks ## Aligned with the best standards. This policy draws on frameworks recognized in Quebec, in Canada and internationally. - Quebec Statement of Principles for AI (2025) - Law 25 - OECD AI Principles - ISO/IEC 42001 - NIST AI RMF ### Owner Anthony Carange, President, oversees the application of the policy. ### Questions or reports Write to us at [info@carangesolutions.com](mailto:info@carangesolutions.com) Want to see how we protect your data every day? [Munia security](/en/securite/)[Privacy policy](/en/politique-de-confidentialite/)