Skip to content

Advisory Workstreams

  1. How to use AI in our business: business process evolution + data estate
  2. How to use AI in our products: technology direction + product development
  3. How to engage with large scale infrastructure providers
  • Augmentation vs. automation
  • Prompt engineering ([Example prompt for synthetic data generation (for training a [model](Example%20prompt%20for%20synthetic%20data%20generation%20(for%20training%20a%20[model) vs. fine tuning vs. pre-training
  • Model capabilities vs. company capabilities
  • Evals: making progress towards your North Star
  • Data flywheel
  • Product management in the age of AI
  • Using AI for internal workflow audits
  • Core competence vs. borrowed competence
  • Eval Strategy (Or, Knowing how good we are): what are we measuring, how do we measure it, what does great look like?
  • Models: off the shelf vs. fine tuned vs. pre-trained. Is there something good enough? When will that thing arrive?
  • Static vs dynamic strategy: what will change as models improve? What is durable, what will get subsumed?
  • Product feedback loop
  • Prediction Strategy (Or, what part of the value chain are you using AI to lower the cost of?): how quickly will it cost reduce, what’s the cost floor, what complements become important?
  • Business process change? The big lever!

An AI audit of a company examines how artificial intelligence systems are designed, deployed, and governed — ensuring they’re ethical, compliant, secure, effective, and aligned with business goals. The specific elements depend on the scope (technical, ethical, legal, or operational), but here’s a comprehensive breakdown of the key components:

AI strategy alignment: How AI initiatives support the company’s broader mission and objectives.

Governance structure: Who is accountable for AI decisions (e.g., AI ethics board, data governance committees).

Policies & documentation: Presence of AI use policies, responsible AI principles, and lifecycle documentation.

Change management: How updates to AI systems are tracked and approved.

Data sources & provenance: Where data originates and how it’s validated.

Data quality & representativeness: Completeness, bias, and relevance of datasets.

Data privacy & security: Compliance with GDPR, CCPA, or other regulations.

Data governance processes: Consent management, retention policies, anonymization.

Model design: Choice of algorithms, rationale for architecture.

Training process: Documentation of parameters, hyperparameter tuning, training data lineage.

Testing & validation: Cross-validation, performance metrics, overfitting checks.

Version control: Tracking of model versions and associated datasets.

Continuous learning: How models are retrained and monitored over time.

Bias & discrimination testing: Systematic review of model outcomes for protected groups.

Explainability & transparency: Ability to explain AI decisions to end users and regulators.

Human-in-the-loop: Safeguards for human oversight and intervention.

Accountability: Clear assignment of responsibility for AI-driven outcomes.

Adversarial resilience: Testing against data poisoning or prompt injection (for LLMs).

Access control: Who can access model code, data, and deployment systems.

Incident response: Procedures for handling AI system failures or breaches.

System monitoring: Logging, audit trails, and anomaly detection.

Regulatory alignment: Compliance with frameworks like:

EU AI Act

U.S. AI Executive Order

ISO/IEC 42001 (AI management systems)

IP rights: Ownership of AI-generated content or models.

Liability management: Procedures for errors or harms caused by AI.

Business KPIs: ROI, productivity gains, or cost reduction attributable to AI.

Technical KPIs: Accuracy, precision, recall, latency, uptime.

User satisfaction: End-user trust, adoption, and usability feedback.

Sustainability impact: Energy consumption and carbon footprint of AI workloads.

Model cards / system cards: Standardized summaries of model purpose, data, and risks.

Risk registers: Identified AI risks and mitigation actions.

Audit logs: Records of model updates, access, and performance monitoring.

Evidence repository: Documentation for compliance verification.

Third-party assessment: Independent validation or certification.

Benchmarking: Comparing AI maturity against industry standards.

Ethical impact assessment: Structured review of social or societal consequences.

Would you like me to tailor this framework for a specific industry (e.g., finance, healthcare, retail, or government)? Each has different audit emphases (like bias/fairness in lending vs. explainability in medical AI).