Three lines of defense and the RACI discipline
Lesson 4 of 5 in Designing the AI Governance Operating Model.
The role architecture underneath everything you have built so far comes from banking. After the 2008 financial crisis, the US Federal Reserve issued SR 11-7 (2011), the supervisory guidance on model risk management — and with it, hard-won answers to exactly the questions AI governance now faces: who owns a model’s risk, who checks the owner, and who checks the checker. The answer is the three lines of defense (the Institute of Internal Auditors’ 2020 update restyles it the "Three Lines Model", softening "defense" to emphasise collaboration — the structure is identical).
First line: the people who build and run. Model developers, model owners, deploying business units. They own the risk of what they ship: they write the model cards, run the pre-deployment tests, operate the monitoring, and fix what breaks. Governance that treats the first line as passive subjects of review fails — risk ownership cannot be outsourced to a review committee.
Second line: the people who set standards and challenge. Risk, compliance, and the AI governance function itself. They write the tiering methodology, define what a valid assessment looks like, challenge the first line’s work ("your fairness testing used the wrong population"), and run the committee. Crucially, they do not do the first line’s work for it.
Third line: the people who assure. Internal audit, reporting to the audit committee of the board — independent of both other lines. They do not review individual models day-to-day; they audit whether the program works: is the inventory complete, do controls operate as designed, does the second line actually challenge or merely stamp?
Key terms: three lines of defense, model risk management, RACI, independent validation, internal audit
Interactive sorting exercise: An AI governance program is staffing up. Assign each activity to the correct line of defense.
Lines of defense answer which function does what. RACI answers which named role does what, stage by stage, for a given system: Responsible does the work, Accountable owns the outcome (exactly one per row — the person whose name is on it when it fails), Consulted contributes before the decision, Informed hears about it after. Building the RACI for the AI lifecycle — design, data, build, validate, approve, deploy, monitor, retire — is a half-day workshop that surfaces every ambiguity your org chart has been hiding.
| Lifecycle stage | Model owner (1st) | Data steward (1st) | Governance / risk (2nd) | Validation lead (2nd) | Internal audit (3rd) |
|---|---|---|---|---|---|
Design & intake | A/R — defines purpose, submits intake | C — data availability & permissible use | R — triages, assigns tier | I | I |
Data sourcing | A — owns fitness of data choices | R — lineage, quality, lawful basis | C — challenges representativeness | I | I |
Build & test | A/R — builds, documents, self-tests | C | C — standards questions during build | I | I |
Independent validation | C — answers challenge, remediates | C | A — owns the validation standard | R — executes the validation | I |
Approval & deploy | R — presents case, deploys on approval | I | A — committee decision & conditions | C — validation verdict feeds decision | I |
Monitor & retire | A/R — operates monitoring, triggers retirement | C | C — reviews breaches, re-triage calls | C — revalidates on major change | R — periodic program audit |
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.