AI in HR: From Recruitment to Performance

A technical role goes live, and two weeks later several hundred CVs sit in the inbox while the decision deadline is a week away. What follows is less selection than surrender to arrival order: the first few applications are read carefully and the rest are dismissed at a glance.

Error in HR tends to accumulate at three points: initial screening, a performance review resting on the manager's memory, and attrition noticed the day the resignation letter arrives. The job of intelligent tools here is not to replace the judge but to order the queue of judgement.

What smart screening does and does not do

• Structuring the CV: extracting skills, history and tools from free text so applicants can be compared on the same basis.
• Ranking against the job description: ordering the queue by proximity to the written requirements rather than by submission time.
• Consistency of criteria: applying one fixed measure to every file, something no human sustains by the two-hundredth CV.
• What it is not for: judging motivation, fit with the team and the final hiring decision; these stay with the manager.

Performance review without relying on memory

The alternative to an annual review built on recent recollection is a picture assembled from work data: on-time delivery, the volume of rework, and how load is distributed over time. The picture only works if the metric fits the nature of the role; judging a developer by lines of code reintroduces the old bias in new clothing and draws the team into gaming the number.

Predicting attrition, and its ethical limit

A model can bring early signals together: responsibility that has stopped growing, withdrawal from team work, long tenure in one role. The value lies in the chance to open a conversation early, not in labelling people. If the output finds its way into someone's file or into sidelining them pre-emptively, the organisation's trust takes more damage than the turnover would have cost. Stating openly what data is collected and how it is used is part of doing this properly.

Algorithmic bias: a hazard that arrives from the past

A screening model learns from past hiring records. If a group was systematically under-hired there, the model reproduces the same pattern, now behind the neutral appearance of a number. Three measures are the minimum: removing sensitive attributes and their proxies such as school name or neighbourhood, testing outcomes by group at regular intervals, and retaining the reason behind each ranking.

Why it matters in Iran's labour market

Competition for specialists in Iran has intensified, and replacing a key person, particularly in technical roles, is both slow and expensive. At many small and mid-size firms the HR function is one or two people running payroll, hiring and administration at once. In that structure, lightening the screening load frees more capacity than any other tool.

A three-month path

• Month one: pick one high-demand role, turn its job description into measurable criteria, and organise the records of past hires in that role.
• Month two: run smart screening in parallel with your current method rather than instead of it; the gap between the two shortlists is the best material for correcting the criteria.
• Month three: if shortlist quality and time-to-hire improved, extend the scope to comparable roles; if not, revisit the job description and the data.

The mistakes that cost the most

• Handing the final rejection to the system; a ranking is a suggestion, not a verdict.
• Training a model on records that were themselves the product of bias, with no review.
• Measuring performance with a metric staff can raise without any real improvement.
• Being secretive about these tools; an employee who does not know the rule will not trust its outcome.

Frequently asked questions

• Where should we start?
  With screening for one high-demand, frequently filled role; the volume is enough to see a difference and the decision risk stays manageable.
• What if we hire only rarely?
  Structuring CVs and standardising criteria is worth doing at any scale, but a predictive model built on a thin history is not reliable.
• Does it replace the HR specialist?
  It lightens screening and reporting and moves the specialist towards interviewing, retention and career design.

Takeaway

AI helps in HR precisely where volume wears down human judgement: initial screening, seeing performance patterns, and early warning of attrition. It is also the area least tolerant of error, because the output is tied to people's livelihoods. Start with one role, keep the final decision with a person, and state the rules openly to both sides.

Glossary

• CV screening: the initial sorting and ranking of applicants by fit against the requirements of a role.
• Algorithmic bias: a model trained on past data reproducing the discrimination contained in that past.
• Turnover rate: the share of employees who leave the organisation within a given period.
• Human in the loop: a process design in which the final decision always passes through a person's approval.