AI Can Approve a Loan. It Can’t Take Responsibility for One.

By Jean-Philippe St-Jacques

For more than twenty years in financial services, I worked under a simple rule: every decision had a name on it. Mine, or the name of someone on my team. When a file went wrong, we didn’t look for someone to blame. We looked for the person who had decided, so we could understand, correct, and learn.

That culture of accountability never slowed the business down. It made it durable. In the organizations I led, it coexisted with sustained growth — portfolios expanding sharply, teams doubling — through economic cycles, rate shocks, and technology shifts. You can grow fast and stay accountable. One does not prevent the other.

The setting is changing. The question is not.

Today, models decide — or pre-decide — almost everything: credit approvals, pricing, fraud detection, claims handling, the prioritization of commercial files. Canada’s regulator has now formalized the expectations. OSFI’s Guideline E-23, Model Risk Management (2027), issued in final form in September 2025 and effective May 1, 2027, requires federally regulated financial institutions to manage model risk — artificial intelligence included — on an enterprise-wide basis, across the model’s entire life cycle.

The regulator sets the framework. But a framework does not decide in place of leaders. And that is where the real risk lies, and it is not technological: it is the dilution of responsibility. “The model declined it.” “The algorithm produced that price.” Every time an automated decision loses its name, someone, somewhere, has quietly shed responsibility — and the institution has accepted a risk it no longer understands.

No customer will accept that. No regulator will either. “The algorithm did it” will never be an acceptable answer to a question about fair treatment.

Four leadership moves

The good news: AI governance does not require reinventing management. It requires applying to models the same disciplines good leaders already apply to people.

  1. Name an owner for every automated decision. Not a committee, not a vendor: a person, named on the file, who answers for the outcome — including when the model “decided” on its own.
  2. Document the data, the logic, and the limits. A model whose inputs and blind spots cannot be explained is a liability, not an asset. Documentation is not bureaucracy: it is what makes a decision defensible two years later, in front of a customer, an auditor, or a board.
  3. Test for bias the way you test for fraud. No one would put a payment process into production without controls. A model trained on historical data can reproduce — and amplify — the biases embedded in that data. Testing must be periodic, documented, and taken as seriously as an audit.
  4. Guarantee a real power to override. Human supervision that cannot reverse the machine is a fiction. Employees who spot an absurd decision must be able to correct it — and be recognized, not punished, when they do.

The expectations will cascade

E-23 targets federal institutions, but its expectations will flow to their suppliers and partners: fintechs, brokers, insurers, technology providers. OSFI has been clear that third parties serving institutions will face heightened model-risk expectations. Leaders of smaller financial-sector firms who wait to be told will already be behind. Those who name an owner now, document their decision-support models, and test them for bias will turn a requirement into a commercial argument.

The conviction

This fall, I am going deeper into AI transformation and governance — IBM’s Lead AI Transformation program, then ISO/IEC 42001 certification. One conviction keeps being confirmed: the institutions that win with AI will not be the fastest adopters.

They will be the ones whose leaders put their name on the outcome.

The algorithm doesn’t have a name. Your decisions should.


Jean-Philippe St-Jacques, EMBA (McGill), Adm.A., C.Mgr., is a financial services executive with more than twenty years of leadership experience at Canadian financial institutions. He writes on ethical leadership and AI governance.

First published in the 5 Mindset newsletter on LinkedIn, October 21, 2026.