AI Governance Services

AI governance is the set of policies, processes, and technical controls that ensure AI systems behave as intended, can be audited and challenged, and comply with applicable regulations. In practice: documenting how a model was built and what it was trained on, evaluating whether it produces biased outcomes for specific groups, building the infrastructure to explain individual decisions, designing human override paths for automated decisions that affect people, and monitoring the model in production for performance changes that could indicate data drift or unexpected behaviour. Governance is what separates an AI system that can be defended to a regulator from one that cannot.

Governance requirements are highest for AI systems that make or influence consequential decisions about people: credit scoring and lending decisions, insurance underwriting and claims assessment, candidate screening and hiring recommendations, pricing decisions that vary by customer attribute, content moderation, medical diagnosis support, and benefit eligibility determinations. Regulatory requirements vary by jurisdiction and industry — GDPR's automated decision-making rules (Article 22), the EU AI Act's risk tiers, financial services model risk management guidelines (SR 11-7 in the US, PRA SS1/23 in the UK), and sector-specific rules in healthcare and insurance. We help you understand which rules apply to your specific use case before scoping the governance work.

A model card is a standardised document that describes a machine learning model: what it does, what data it was trained on, how it was evaluated, its performance across different population segments, its intended use cases, and its known limitations. Model cards originated at Google and are now a standard component of responsible AI deployment. You need one whenever a model makes decisions that could affect people differently based on protected characteristics, or whenever you need to demonstrate to a regulator, auditor, or customer that your AI system was built and tested responsibly. We produce model cards as a deliverable of the governance engagement, not as documentation produced after the fact.

Bias evaluation covers three stages. First, dataset analysis: examining the training data for representation imbalances across protected attributes (gender, race, age, disability status, geography) that could produce systematically different outcomes for different groups. Second, model evaluation: measuring performance metrics (accuracy, false positive rate, false negative rate) separately for each protected group and identifying where the model performs materially worse for specific segments. Third, outcome analysis: for deployed models, analysing whether actual decisions differ systematically by protected attribute after controlling for legitimate predictive factors. The specific fairness metrics used depend on the use case — equalised odds, demographic parity, and calibration each capture different notions of fairness, and the right metric depends on what discrimination would mean in your context.

A focused governance engagement for one deployed model — model card, bias evaluation across key protected attributes, SHAP-based explainability report, and an audit trail design — typically runs $15,000--$40,000 and takes 4--8 weeks. A comprehensive governance programme covering multiple models, ongoing monitoring, a governance policy framework, and regulatory mapping typically runs $40,000--$100,000. We scope after a call to understand which models are in scope, what regulatory requirements apply, and what governance documentation you already have.

AI explainability means being able to produce a human-readable reason for a specific model decision. At the instance level: 'This application was declined because the debt-to-income ratio (35% vs 28% threshold) and 24-month payment history were the two factors with the largest negative influence.' At the model level: a summary of which features drive predictions across the population. Explainability is technically required under GDPR Article 22 for fully automated decisions that have legal or similarly significant effects, under the EU AI Act for high-risk AI systems, and under financial services model risk guidelines. Practically, it's required whenever a human needs to review, challenge, or override an AI decision. We implement SHAP (SHapley Additive exPlanations) for feature attribution, LIME for local approximations, and counterfactual explanations ('What would need to change for this decision to be different?') depending on the model type and the explanation audience.

Human-in-the-loop (HITL) design defines the conditions under which automated decisions are reviewed by a human before being acted on. The design covers: which decision categories require human review (borderline confidence scores, protected attribute flags, high-value cases), what information the reviewer sees (the model's recommendation, the confidence level, the feature attribution, the case data), what actions the reviewer can take (approve, override, escalate), and how the review decision is recorded (for audit trail and for model retraining). We design and build the review queue interface, the case presentation, and the override logging — not just the policy document.