Inside the organisation: three lines of defense

Lesson 2 of 5 in Who’s Who in AI Governance: Actors, Roles, and Responsibilities.

Laws address organisations; organisations run on people. When a regulator writes “the provider shall establish a risk management system”, somebody specific has to wake up owning that sentence. Inside a company, AI governance is not a department — it is a relay race across several, and most failures happen at the baton passes.

The cast you will meet in any mid-sized deployer or provider: data scientists and ML engineers (build and test the models), product managers (decide what ships and when), a responsible-AI lead or chief AI officer (sets policy, arbitrates the hard calls), the data protection officer (GDPR interface — often the first person in the building who ever governed algorithms), legal and compliance (map obligations to controls), risk management (keeps the register, challenges the first line), internal audit (checks everyone else), and increasingly an AI ethics board or review committee for the cases policy does not settle.

The structure that organises them comes from banking, not from AI: the three lines of defense. Its logic is separation of duties — the people who build and profit from a system should not be the only ones judging its risks, and the people who judge risks should themselves be auditable.

The model has a famous failure mode you should watch for in AI settings: the first line outruns the second. Data-science teams ship weekly; risk committees meet quarterly. Governance that arrives after deployment is archaeology. That is why mature programs push controls into the first line — bias tests in the CI pipeline, model cards as a merge requirement — instead of relying on downstream review to catch everything.

Three lines of defense, translated to AI
LineWho sits hereWhat they ownAI-specific example

1st line — owns the risk

Product, engineering, data science — the people building and running the system

Designing, testing, and operating within policy; executing controls day to day

ML engineers run disaggregated bias tests before each release and log results in the model card

2nd line — oversees the risk

Risk management, compliance, privacy office, responsible-AI function

Setting policy, maintaining the AI inventory and risk register, challenging first-line decisions

The responsible-AI lead reviews the pre-deployment risk assessment and can block launch

3rd line — checks the checkers

Internal audit, reporting to the board’s audit committee — independent of both other lines

Independent assurance that the whole framework actually works as described

Internal audit samples deployed models yearly and tests whether the claimed oversight really happens

Key terms: three lines of defense, data protection officer, AI inventory, risk register, AI ethics board

Interactive sorting exercise: Your company deploys a résumé-screening AI. Sort each task to the line of defense that should own it.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.