AI is good at organising an application pile and poor at deciding who deserves a job. The line matters, because a screening rule that quietly filters out a protected group is still discrimination even when nobody intended it.
Key takeaways
The first is extraction: pulling structured information out of unstructured documents so a pile of resumes becomes a table. The second is summarising: producing a short, neutral account of an application against the requirements you actually published. Both leave the assessment with a person.
The third is deciding: scoring, ranking, or filtering candidates out before a human sees them. That is a different activity with a different risk profile, and it is the one worth being deliberate about. Our position on it is simple. AI prepares the pile. A person decides who progresses.
Section 11 of Ontario’s Human Rights Code provides that a right is infringed where a requirement, qualification or factor exists that is not discrimination on a prohibited ground but that results in the exclusion, restriction or preference of a group of persons identified by a prohibited ground, unless the requirement is reasonable and bona fide in the circumstances. The Ontario Human Rights Commission calls this constructive, or adverse effect, discrimination.
That is exactly the shape of the risk with automated screening. A filter never has to mention a protected ground to have an effect along one. Gaps in employment history track parental and medical leave. Graduation years track age. A preference for local experience tracks place of origin, which is one reason Ontario has now prohibited Canadian experience requirements in job postings outright. None of those filters look discriminatory in a settings screen.
Where a neutral requirement does have an adverse effect, the employer can try to show that it is reasonable and bona fide, and must also show that the affected person or group cannot be accommodated without undue hardship. That is a real defence, and it is available to any employer who can articulate why a requirement is genuinely necessary for the job.
The problem with an inherited screening model is that you often cannot articulate it. If the answer to why a candidate scored low is that the model weighted something it learned from your previous hires, you have no defence to mount, because you cannot state the requirement, let alone justify it.
Write the criteria before you open the applications, and keep them to what the posting actually said. Point the tool at those criteria only, and have it produce a summary against each one rather than a single number. A summary invites a person to disagree with it; a score invites them to accept it.
Keep the audit trail: what the tool did, what a person changed, who decided. Route any accommodation request to a human immediately rather than through a screening step. And if you operate in Ontario and use AI at any of these stages, remember that the posting itself now has to say so.
A 30-minute call is enough to tell you whether AI pays for itself here.
