A hiring tool promises to rank three hundred applications before lunch. The team feels relieved—until nobody can explain why a strong applicant landed near the bottom.

“The system scored them” is not a hiring reason. It is an admission that the decision moved somewhere the employer cannot see.

Artificial intelligence can help with bounded administrative work, but hiring affects access to income and opportunity. A small team needs a simple governance plan that keeps accountability with people.

Name the task before choosing the tool

Do not begin with “Where can we use AI?” Begin with the bottleneck.

There is a meaningful difference between using software to schedule interviews, summarise an approved job description, search application records and rank candidates. The closer a tool moves toward recommending or making a decision, the greater the evidence, oversight and legal scrutiny required.

Write the permitted task, prohibited uses, owner and human-check point. For example: “The tool may extract stated qualifications into a review sheet, but it may not reject or rank candidates.”

If the team cannot describe the boundary, it is not ready to deploy the tool.

Map the data and the consequence

List every item the system receives, infers, stores and sends to another provider. CVs may contain names, contact details, work histories, locations and other personal information. Video or voice analysis can introduce still more sensitive signals and questionable inferences.

Ask the vendor:

  • Which data fields are processed and for what stated purpose?
  • Is customer data used to train another model?
  • Where is it stored, for how long and who can access it?
  • Which subprocessors receive it?
  • Can data and derived scores be deleted?
  • What exactly happens when the system is uncertain?

POPIA sets conditions for lawful processing and includes protections concerning solely automated decisions with legal or similarly substantial effects. Section 71 also addresses safeguards such as an opportunity to make representations and sufficient information about the underlying logic in relevant circumstances. Obtain advice on how these duties apply to the planned use; do not treat a consent tick-box as a complete answer.

Test the claim in the real context

A vendor accuracy number may describe another country, occupation, language mix or outcome. It does not show that the tool identifies success in your vacancy.

Before live use, create a test set that reflects the role and likely applicant population. Check whether the output is consistent, whether it misses obvious evidence and whether small formatting changes alter the result. Test assistive-technology routes and reasonable accommodations. Include people who understand the job, privacy, employment equity and the system’s technical limits.

Define what failure looks like before seeing the result. A tool that saves time but excludes qualified candidates is not efficient.

NIST’s voluntary AI Risk Management Framework organises this work around govern, map, measure and manage. Its useful message is that risk management continues through the system’s lifecycle and requires documented roles, testing and monitoring.

Keep human review meaningful

A person clicking “approve” on an unexplained ranking is not meaningful oversight.

The reviewer needs enough time, authority and information to disagree. They should see the source evidence, understand what the tool did, check the score against job-related criteria and record their own reason. Create a route for uncertain or conflicting cases to receive a second review.

Do not show the algorithm’s recommendation before an independent review where anchoring would undermine the purpose. For example, a recruiter can assess a sample of applications against the scorecard, then compare the system output.

The employer remains accountable for the employment decision even when a third-party system influenced it.

Tell candidates and offer a route back

Plain-language notice builds better protection than a hidden clause inside a long privacy policy. Explain where automated assistance is used, what information it processes, what role humans play and how a candidate can request help or raise a concern.

Provide a working contact route. Someone must own the response, access the relevant record and have authority to correct an error or arrange human reconsideration.

Do not force every candidate through a format such as automated video analysis when a reasonable alternative can assess the same job requirement. Accommodation should not become a penalty or a signal attached to the decision.

Monitor outcomes and stop when necessary

Create a small decision log containing the tool version, task, reviewer, override and reason. Review false exclusions, candidate complaints, overrides and outcomes across relevant groups where lawful and methodologically sound.

Watch for drift after a model, prompt, job description or vendor process changes. Re-test material changes before relying on them.

Set stop conditions in advance. Pause the tool if the team cannot reproduce an output, a group appears to be disadvantaged, candidate data moves outside the agreed purpose, or reviewers routinely override the system for the same reason.

South Africa’s Employment Equity Act prohibits unfair discrimination and extends relevant protections to applicants. Technology does not make a discriminatory outcome neutral. It can make the path to that outcome harder to inspect.

The Bloomberg Originals video documents how AI can score applications and interviews while candidates, advocates and lawmakers question bias and transparency. That tension remains the right starting point. AI can assist a hiring team; it should never become the colleague nobody is allowed to question.

Sources and limits

The video describes practices and concerns visible in 2022; tools and laws continue to change. NIST’s framework is voluntary and not a legal safe harbour. This article offers a governance starting point, not a product endorsement, transcript or legal advice.