Skip to main content

“A person decides”: what human review of an AI recommendation means in practice

Signing off on an AI ranking is not the same as reviewing it. A real review needs authority to disagree, the application in view, enough time to read it and a reason written down.

6 minute readAI in hiring

Human review of an AI recommendation is real when the person reviewing can disagree without penalty, has the application in front of them and not only the score, has time to read it, and records a reason in their own words. Take away any one of those and the tool is deciding while the person signs. Candidates can tell the difference, and Canadian guidance on AI draws the same line.

This article is about what happens after a screening tool has produced a ranking or a score. The statement you put in an Ontario posting is a separate question, covered in our Ontario explainer.

Where the expectation comes from

Three Canadian texts describe what human involvement should look like. Each comes at it from a different angle.

Ontario’s privacy and human rights commissioners. On January 21, 2026, the Information and Privacy Commissioner of Ontario and the Ontario Human Rights Commission released joint principles for the responsible use of AI. They are addressed to the Ontario public sector and the broader public sector, and the commissioners say the principles will ground their assessment of how organizations adopt AI. They do not bind private employers; we use them here because they are the clearest statement of what Ontario’s regulators expect. The accountability principle calls for a human-in-the-loop approach that keeps human oversight in place and allows intervention in real time, and for a designated person or persons responsible for pausing or decommissioning a system that produces unsafe outputs. The privacy principle adds that people should have at least a right of review for automated decisions that are not high risk, and the choice of opting out of high-risk ones in favour of a human decision maker.

The federal Directive on Automated Decision-Making. This Treasury Board directive applies to federal institutions subject to the Policy on Service and Digital, not to private employers. It is still the most detailed Canadian text on the subject. It sorts systems into four impact levels. At levels I and II, a system may make decisions without direct human involvement. At levels III and IV, the final decision must be made by a human, and people review the system’s decisions or recommendations for accuracy and appropriateness and can override them. The directive also asks institutions to document human overrides and to use them to find and correct problems.

Quebec’s private-sector privacy law. Section 12.1 sets duties for decisions based exclusively on automated processing of personal information: telling the person, explaining on request the information used and the reasons and principal factors behind the decision, and giving them the chance to submit observations to a staff member who is in a position to review it. Those duties attach only to decisions based exclusively on automated processing. The Act does not say how much human involvement takes a decision outside section 12.1, but our reading is that a reviewer who only signs off on what the tool produced may not be enough to take the decision outside it. In Quebec, whether your reviewer actually decides is therefore more than a formality.

Four conditions for a real review

The reviewer can say no

The person must be able to advance a candidate the tool ranked low, or hold one it ranked high, without asking permission and without extra steps that apply only when they disagree. If overriding the tool needs a manager’s sign-off while agreeing takes one click, the process leans toward agreement. Ask for the same effort in both directions: one reason for every decision, whichever way it goes.

The reviewer sees the application

A score on its own cannot be reviewed. The reviewer should be able to open the résumé or the screening answers, see which requirements the tool assessed and on what evidence, and see what it could not assess. For higher-impact systems, the federal directive requires training that keeps staff’s understanding of how the system works up to date. A reviewer who does not know what the tool looks at cannot tell when it is wrong.

The reviewer has time

Take a hypothetical agency that posts a warehouse job and receives 400 applications. One recruiter has six hours set aside for screening. Six hours is 21,600 seconds, so that is 54 seconds per application, before any phone calls. At that pace the score decides.

There are two honest ways out. Change the volume each person handles: two recruiters with six hours each get 108 seconds per application. Or change what people review: let the tool sort, then have a person read every application the tool would place below the interview line, starting with those closest to it, plus every application it could not score. What you cannot do is describe 54-second clicks as review.

The reviewer writes down why

A reason in the reviewer’s own words, tied to a requirement of the job, shows the decision was theirs and gives you something to check later. “Score 2.1” or “per AI ranking” is not a reason. “No forklift certificate; the job requires one from the first shift” is.

With each decision, keep who decided, when, what the tool recommended and whether the person agreed. That record answers a candidate’s question months later. It also tells you whether the tool is drifting, which is why the federal directive asks for overrides to be documented.

Signs the review has become a formality

  • Reviewers have agreed with the tool on every decision for months. Agreement is not wrong, but if nobody has disagreed once in 1,000 decisions, ask whether anyone could have.
  • A single click rejects everyone below a cut-off score.
  • Rejection messages go out on a timer, whether or not anyone has opened the file.
  • The reason field holds the score, or text copied from the tool’s output.
  • The reviewer can see the ranking but cannot open the application.
  • Nobody can say who is allowed to pause the tool.

When a candidate asks for a person

The IPC and OHRC principles expect public-sector institutions to have a mechanism to receive and respond to challenges about how an AI system arrived at a decision. For any employer, the same idea in practice means a named contact, a response time you can keep and, where the team is large enough, a second reviewer who did not make the first decision. The second look should start from the application, not from the first reviewer’s note.

Tell the candidate what was reviewed and what the outcome was. If the second reviewer reaches a different conclusion, record that too. It is the most useful piece of information you will get about your own process.

How we handle it

In TalentInitiator, AI output is stored as a recommendation and never as a decision. Moving a candidate forward, holding or rejecting needs a person and a reason, and a decision whose only stated basis is the AI recommendation is refused. Our AI use page describes the checks.

How to read this post

This post explains the rules in general terms, as they stood on the date above. It is not legal advice for your organization.

Sources

All posts