Who Answers When the Algorithm Decides?

An AI system can approve a loan. It cannot take responsibility for one.

That gap is the central governance question for boards in Canada today. Credit models, hiring tools, pricing engines and fraud filters now shape outcomes long before a director sees a report. The decisions are real, the consequences land on customers and employees, and the accountability has not moved with the technology.

The question no board can delegate

Boards have always delegated execution. They have never delegated responsibility. When a model declines a small business loan, lowers an insurance premium or screens out a candidate, the organization still owns the result. Regulators, courts and customers will not ask which vendor wrote the code.

AI governance starts with a simple discipline: naming who answers for each automated decision before it goes live.

Where decisions quietly moved

Most important decisions are no longer made in a single meeting. They are made in a pipeline. Data is selected, a model is trained, a threshold is set, and a score is produced. By the time a decision reaches a human, several choices have already been made, often by people who will never sit on a committee.

This is why oversight that begins after an incident is too late. The moment that matters is the design stage, when someone decides what the system optimizes for, which data it may use, and where a person must be able to intervene.

What governance looks like in practice

Effective governance is less about policy documents and more about clear ownership. In practice, it usually means three things:

  • A named owner for each model. One accountable executive, not a shared committee, who can explain what the system does and why.
  • Clear escalation thresholds. Defined cases where a decision must be reviewed by a person, such as large amounts, unusual patterns or affected groups.
  • A record of accepted risk. A written trace of who approved the system, what limits were set and what the organization agreed to tolerate.

These are not technical exercises. They are the same habits boards already apply to financial controls, applied to a new kind of decision-maker.

Three questions to ask management this quarter

Directors do not need to understand model architecture to ask good questions. Three are enough to start:

  1. Which decisions in our business are now automated, fully or in part?
  2. Who can override them, and how often has that happened?
  3. How would we know if one of these systems began to drift from what we intended?

If management cannot answer these clearly, that is itself a finding.

Keeping human judgment at the centre

Human oversight is a design choice. It has to be built into the system, assigned to a person and tested before deployment. A review added after a complaint is a reaction, not governance.

Systems shape behaviour. The people who design, approve and oversee them shape the systems. Boards that keep that distinction in view will make better decisions about where automation belongs and where it must stop.

What boards owe the people they never see

Most of the people affected by an algorithmic decision will never know a board exists. They will not read the minutes or attend the annual meeting. They will simply receive an answer.

That is the quiet obligation of governance in the age of AI: to make sure that someone, identifiable and accountable, is answerable to those people for the decisions made in their name.