Skip to main content

Risk Management

Why this function belongs in the AI Operating Model

Not all AI initiatives are equally risky: an internal document assistant and an automated customer decision are different worlds. The AI Operating Model relies on the principle of routing by risk level (BCG): simple cases go down a lightweight self-service track, medium ones through trust-but-verify, significant ones through strategic review, and impermissible ones are prohibited. Someone has to assign the initiative a risk profile and determine its route. That is the role of risk management.

The risk function doesn't duplicate InfoSec and legal; it aggregates their opinions into a single assessment and links it to the corporate risk appetite.

Where it engages

AI Operating Model stageRole of risk management
AssessmentAssigns a risk rating and approval route for the initiative
Before productionChecks that residual risks are within appetite, locks in controls
In supportMonitors materialized risks, incidents, model drift
PortfolioRolls AI risks up into the company's overall risk profile

What the function receives as input

  • The scenario, impact on customers, processes, and finances.
  • InfoSec opinions (threats) and legal opinions (regulation) as inputs.
  • Data on process criticality and the reversibility of the AI's decisions.

What the function delivers as output

  • The initiative's risk rating and approval route through the funnel.
  • A list of mandatory controls (human-in-the-loop, limits, kill switch).
  • A decision within risk appetite: proceed, restrict, or stop.

Key touchpoint artifacts

  • Initiative risk profile — rating, route, key risks and controls.
  • Routing matrix — mapping risk level to the line of approval.
  • AI risk register — a portfolio view of the risks of all initiatives and products.

Anti-patterns

  • One heavy committee for everything. Low-risk cases go through the same heavy review as critical ones — the funnel clogs up.
  • Risk assessment without controls. The risk is recorded, but no compensating measures and monitoring are assigned.
  • No monitoring in production. The risk is assessed once at intake and forgotten, ignoring model drift and new incidents.