We use strictly necessary cookies to make this website work. With your consent, we may also use functional, statistics and marketing cookies, some of which come from third parties. These optional cookies are off by default. You can change your choices or withdraw your consent at any time using the “Cookie settings” link at the bottom of every page.
This section brings together the most frequently asked questions related to data protection, model explainability, bias mitigation, and human oversight.
Tip: use the left and right arrow keys to switch questions.
Governance01 / 30
Does Makila’s AI make autonomous decisions?
No. Our AI is supervised: no action occurs without human validation. In certain strictly controlled cases, such as predictive analysis, the AI has some degree of autonomy to make its projections, but these decisions never lead to actual actions. In all cases, outputs are delivered as reports/tables for review and approval.
Are AI roles and responsibilities formally defined?
Yes. Responsibilities cover the provision, integration/deployment, operation, and auditing of AI. Security and data teams define the scope, controls, and traceability at each stage.
Can you provide documentation on the AI methodology in the event of an audit?
Yes. Methodology documentation outlining the data preparation pipeline, controls, validation metrics, and operational rules is available retrospectively in the event of an audit.
Is personal data used to train the models?
By default, direct identifiers (such as first name, last name, email, etc.) are excluded from training processes. We apply data minimization, anonymization/pseudonymization, and masking as required.
Are data or AI outputs ever reused to retrain the models?
No. AI outputs are not reused for training. Outputs are watermarked (visibly or via metadata) to ensure origin and integrity.
What AI techniques are used depending on the use case?
We use a mix tailored to each case:
Rule-based techniques for formats, missing values, and outliers
Data is hosted in the selected region, with options for replication and logical separation of environments (dev/test/qa/prod).
How are AI models and pipelines protected from attacks?
We use RBAC, MFA, admin bastions, logging, network segmentation (WAF, IPS, NSG/VNet), and perform adversarial tests to protect against malicious inputs and data poisoning.
Are there measures in place to stop mass data exfiltration through AI?
Yes. Mechanisms include SIEM correlation, monitoring for unusual usage and outbound traffic, and alerts for mass data exports.
Do non-production environments contain real data?
No. We use synthetic or anonymized data, with masking or pseudonymization when needed.
Is AI activity, including inputs, outputs, and performance, logged?
Yes. Detailed logs include access, API calls, infrastructure, resource usage, training/inference events, and performance metrics. Logs are retained for up to six months and are available upon request.
How do you ensure the quality and relevance of AI results?
Systematic human validation, error analysis, thresholds/business rules, false alert control, and cross-review by data and business teams.
Which metrics do you use to evaluate models?
Depending on the use case: precision, recall, F1-score, goodness-of-fit tests, stability indicators, MAE, RMSE.
Are AI models tested for adversarial attacks, such as data poisoning?
Yes. Adversarial scenarios are integrated into the testing cycle, including input validation, anomaly detectors, and robustness against targeted manipulations.
What measures are in place to avoid overfitting or underfitting in AI models?
We use regularization, cross-validation, diverse datasets, production monitoring, and periodic model reevaluation to prevent overfitting and underfitting.
Are your AI results explainable?
Yes. We provide the influencing factors, context, and rules behind the results; decisions remain human and fully traceable.
How is human oversight ensured?
AI actions are never executed without human review and adherence to business rules. Approval processes are clearly defined and formalized.
Can AI outputs be identified as such?
Yes. AI outputs include visible watermarks or metadata, allowing them to be audited and traced back to their inputs and processing pipelines.
Are AI inputs and outputs stored, and if so, for how long?
Retention is based on need. Logs and traces of performance and inference are kept for up to six months, and outputs are never reused for model training.
What happens to AI data at the end of the contract?
At contract end, disk keys are rotated and the full environment—including data, backups, and logs—is securely deleted following a formal process. Inactive environments are removed within 90 days.
Is it possible to remove an individual’s data on request?
Yes. Selective export and deletion are possible, with full traceability and logging.
How do you integrate Privacy by Design into AI?
Data minimization, anonymization/pseudonymization, masking, RBAC, controlled retention; developer guidelines and incident response plan.
What AI awareness practices are in place for teams?
Teams receive regular training on security and AI risks, stay updated on threats such as bias and adversarial attacks, participate in practical workshops, and follow internal charters and policies.
Do you publish responsible AI guidelines?
Yes. Internal Responsible AI policies cover fairness, transparency, privacy, human oversight, traceability, and non-reuse of outputs.
How is AI integrated into our IAM and existing controls?
RBAC and MFA for access to AI modules; IdP federation (SAML/OIDC); SIEM logging for AI APIs and pipelines.
What network guarantees are in place for AI services?
WAF enabled, IPS on firewalls, segmentation via NSG/VNet, DDoS protection; admin access through a bastion; 24/7 monitoring.
What is the model update cycle?
Models undergo periodic evaluations, controlled updates in QA or staging, and require approval before going into production. Outputs are never reused for training.
What does secure AI onboarding look like?
Secure AI onboarding typically includes:
AI and security workshop covering data residency, classification, and risks
IAM setup with RBAC/MFA and IdP federation
SIEM monitoring and alerting for AI-related signals
Data handling with minimization and anonymization
AI testing for quality, adversarial robustness, and metrics
Go-live with bastion access, network segmentation, WAF/IPS, and continuous monitoring
When should NLP, ML, or rule-based approaches be preferred?
Use cases for each approach:
Rules: ensure format integrity, detect simple anomalies, handle specific outliers