A software house in Pakistan advertising a junior developer role can expect several hundred applications, many of them irrelevant. Reading all of them properly is not realistic, so something has to do the sorting.
AI does that well. The risk is the step after: letting it decide rather than sort. Those are different things and the distinction is where most of the harm lives.
Where AI genuinely helps
Extracting structure from unstructured CVs
Pulling out years of experience, technologies, education and contact details from CVs in twelve different formats. Reliable, boring, and it saves real hours.
Matching against stated requirements
Ranking applicants by how well they match what the role actually asks for. Useful as a reading order, not as a cut off.
Deduplication
The same candidate applying to four roles under slightly different names and emails. Easy win.
Scheduling and coordination
The most underrated application. Interview scheduling across three calendars is pure administrative load, and automating it costs nothing in fairness.
Drafting job descriptions and outreach
A reasonable first draft that a human edits. Faster than a blank page.
Where it goes wrong
1. It learns from your past hiring
A model trained on who you hired before will reproduce who you hired before, including the parts you would not defend. If your last forty developers came from three universities, the model learns that those universities are the signal.
This is the failure that has embarrassed large companies publicly, and it happens quietly rather than obviously.
2. It penalises non standard paths
Career changers, self taught developers, people with a gap for family or health reasons, and anyone whose CV does not follow the expected shape. In Pakistan that filters out a meaningful share of strong candidates, because non linear paths are common here. See our note on switching into tech after 30.
3. It rewards keyword stuffing
Candidates learn to game the filter. The result is that you optimise for people who know how to beat screening software, which is not a job requirement.
4. Video and voice analysis
Tools claiming to assess personality, confidence or suitability from facial expression or speech patterns have weak evidence behind them and obvious bias risks across accents, languages and neurodiversity. We would not use them and would advise against them.
What to automate and what to keep human
Automate: parsing, deduplication, checking hard requirements that are genuinely binary (right to work, specific certification where legally required), scheduling, and generating a suggested reading order.
Keep human: the decision to reject, assessment of potential, judging a non standard background, and anything involving a candidate’s character.
The line to hold: AI ranks, humans reject. A person should look at every application before it is declined, even briefly. That is the difference between a tool that saves time and a system that quietly excludes people.
How to implement it without causing harm
- Write the requirements first, separating must have from nice to have. Most bias enters through vague criteria.
- Do not train on your own hiring history unless you have audited it. Match against stated requirements instead.
- Remove name, photograph, age, gender and university from the first pass. All four are proxies for things that should not affect a first read.
- Use ranking, not filtering. A score that orders the pile, not a threshold that deletes part of it.
- Audit quarterly. Compare the demographic shape of applicants, of those shortlisted, and of those hired. If the shape changes sharply at one stage, look at that stage.
- Tell candidates that automated screening is part of the process. It is increasingly expected and it costs nothing.
- Keep records of why candidates were rejected. Useful for improving the process and necessary if a decision is ever questioned.
What it is worth
For a company hiring at volume, screening time typically drops by half to two thirds, and time to first interview shortens by days. In a market where good candidates accept the first reasonable offer, that speed is the actual return.
For a company hiring five people a year, the honest answer is that a well written job description and an hour of reading will serve you better than any tool.
Related reading: the AI tools small businesses actually keep using and HR and payroll software.
Frequently asked questions
Is AI screening legal in Pakistan?
There is no specific prohibition, and general principles around fair treatment and data protection still apply. Keep records of decisions, tell candidates the process involves automated screening, and do not use it to reject without human review.
Will AI screening find better candidates?
It finds them faster within the pool that applied. It does not improve the pool, and a poorly written job advertisement remains the biggest constraint on candidate quality.
Should we use AI to conduct interviews?
For scheduling and note taking, yes. For assessing a candidate from video or voice analysis, no. The evidence is weak and the bias risk is high.
How do we check our screening is fair?
Compare the composition of applicants, shortlisted candidates and hires each quarter. A sharp change at one stage tells you where to look.
Ezitech builds AI and automation systems grounded in a business’s own data, and advises honestly on where automation should stop. See our AI solutions or describe your hiring volume.