Part 8 of 12 · 1 min read
Regulation basics: interview questions
How does regulation affect how you build an agent?
"I classify the use case by risk, because obligations follow the tier. High-risk uses need documentation, logging, human oversight and an impact assessment; chatbots need disclosure. I keep evidence in a manifest and gate releases on it. Legal decides the classification; I make the system auditable." Follow-up: "Name the frameworks." EU AI Act (risk-based), GDPR (personal data), US state laws and sector rules such as PCI DSS and SOX; NIST AI RMF and ISO 42001 as organising frameworks.
What is policy as code?
"Compliance requirements as executable checks in CI: for each system, derive required evidence from its tier and fail the release if it is missing, with named reasons. In the example's sketch 4 of 10 systems passed and each failure said what was missing." Follow-up: "Does passing mean compliant?" No: it means the evidence I defined exists. Counsel and regulators decide compliance.
The AI Act dates keep moving. How do you cope?
"I keep dates and scope in a reviewed config, cite the Official Journal, and design for the strictest likely obligation so a date shift does not change the engineering plan. From what I read, high-risk Annex III moved to December 2027, but I would verify that." Follow-up: "Why design early if dates slip?" Logging, oversight and documentation are costly to retrofit and are good practice anyway.
How does GDPR affect an agent with memory?
"Memory is personal data: lawful basis, minimisation, retention limits, access and erasure. Erasure must reach derived copies, which is why I track lineage (module 6), and significant automated decisions need human involvement." Follow-up: "Is retrieval data covered?" Yes if it contains personal data.