Security and data handling
How we protect our clients’ data in every AI solution.
We work with banks and insurers. These are the practices we apply by default in the solutions we build and operate, and the limits we set for ourselves. What is not here, we do not promise.
Principles
Minimum necessary data
Before any data reaches an AI model we define with the client what goes in, what does not, and why. We pseudonymize or minimize identities whenever the case allows it.
Fail closed
If an AI stage, a credential or a key is unavailable, the solution stops and says so. It never publishes a partial result as success.
Provenance and audit
Every relevant output records provider, model, prompt version and input data. Decisions and audits are stored append-only: never edited or deleted.
Human oversight
Deterministic rules for what is repeatable; the model for what needs language; a person for what affects other people.
Practices in the solutions we operate
| Identity and access | Federated sign-in with Microsoft Entra ID (OIDC) and roles enforced in the API, not only in the interface. HttpOnly cookies and strict SameSite. Without configured identity, access fails closed. |
|---|---|
| Data sent to AI models | In Change Tracker Core the model only receives aliases; real names never enter the prompt. In Risk Preparations historical examples are minimized before reaching the model. We use the providers’ commercial APIs, which do not train on data sent through the API, and we do not train models on client data. |
| Encryption and isolation | Encryption in transit (TLS) and at rest where personal data exists (for example AES-GCM in Kai). Isolation per client and per project with row-level policies in the database; without an explicit project, queries return zero rows. |
| Retention and purge | Retention rules agreed per campaign or project (in Kai, a maximum of 90 days for authorized text) with a recorded purge. Purge certificates state exactly what they cover and what they do not. |
| Integrations | Jira, Microsoft 365 and other integrations connect with the client’s credentials, minimum scope and durable queues with retries. Nothing is created in client systems without an explicit action. |
| Testing and release | Automated suites and quality gates before every release; test environments with synthetic data; cryptographic fingerprint of critical components in conversational products. |
| People | In organizational listening (Kai): audio never stored, consent to the exact version of the text, aggregated results with a minimum of five people per cut, and design prohibitions: no biometrics, emotion analysis, lie detection or decisions about individuals. |
Certifications and compliance
Agility Changes is in the process of ISO/IEC 27001 certification, targeting 2027. We do not hold that certification today; we will publish the scope and certificate here when it exists. In the meantime, our controls are documented per solution and shared with each client’s security team during contracting.
We work under the confidentiality and data-processing agreements each client requires, and adapt data residency and retention rules to their regulation.
Security contact
For security questionnaires, vulnerability reports or documentation requests: info@agilitychanges.com.