Chapter 01 · Section I · 16 min read
What AI does well for HR — and what it does not
A blunt taxonomy of the HR tasks current AI handles reliably, the ones it dresses up to look reliable, and the verification test that tells you which is which before a candidate gets hurt.
You did not come to this course to be told that AI will “transform talent.” You came because your CEO forwarded you a LinkedIn post about a screening tool that promises to cut your time-to-hire in half, and the next morning you had to write a JD for a credit officer in Biratnagar, translate the office harassment policy into Nepali, and explain to a candidate why she did not make the shortlist. This section answers the practical question those three tasks raise: where, in honest terms, does today’s AI help HR — and where does it start to put your candidates, your organisation, and your own professional standing at risk?
The honest taxonomy
HR is unusual among office functions because almost every output it produces touches a person’s livelihood. A bad marketing email is embarrassing. A bad screening decision can leave a qualified woman from Janakpur off a shortlist she should have led. So the first thing to say about AI in HR work is that the same chatbot that drafts a sparkling outreach email is, in the very next minute, capable of confidently telling you that section 11 of the Labour Act 2074 contains a clause it does not contain. Both outputs look equally fluent. The professional act is knowing which output you can trust on its face and which one demands a second source before it leaves your screen.
The clearest way to sort the work is by one question: how cheaply can you check the output, and how badly does it hurt someone if you don’t? When verification is fast and the downside is small, the model is a force-multiplier. When verification is slow or the downside lands on a candidate’s career, the model is a liability dressed as a shortcut.
What AI does well
Six categories, in roughly descending order of how confidently you can lean on the output.
1. Drafting job descriptions and outreach. Hand the model a short brief — role, level, must-haves, nice-to-haves, location, salary band if you have one — and it will produce a JD that is structurally correct and tonally competent in under a minute. Same for the LinkedIn outreach message to a passive candidate, the careers-page paragraph, the reminder email for an interview. You edit for your house voice and remove the genericisms. A JD that used to take an HR generalist forty minutes now takes ten, most of it editing.
2. Translating dense employment law into client-readable Nepali. The Labour Act 2074, the Social Security Act, the Bonus Act — these read like the legal documents they are. A line manager in a Pokhara factory does not want to read them; she wants to know, in plain Nepali, what she has to do when an employee takes maternity leave. The model handles this register shift well, especially if you give it the source clause in English and ask for plain Nepali aimed at a non-lawyer. Read the output before you send it, but the productivity gain is real.
3. Summarising candidate notes against a rubric. After three interviewers have spoken to a candidate, you have three sets of typed-up notes that do not agree on what they noticed. Hand them to the model along with your scoring rubric — “communication, technical depth, ownership, cultural add” — and ask for a one-page consolidated assessment with scores justified by quotes from the notes. You still make the decision. But the half-hour you used to spend reconciling the three notes drops to five minutes of reading.
4. First-draft policies and internal comms. A code of conduct, a remote-work policy, an updated leave guideline, the all-hands email announcing a new HR system — these are pattern-based prose where the model produces a serviceable draft fast. It will not know your organisation’s specific tone or political sensitivities; you supply those in the edit. But the blank-page hour disappears.
5. Generating interview question banks and structured rubrics. Give the model the JD and ask for twelve behavioural questions targeting two competencies each, with follow-up probes and a scoring rubric. What comes back is usable as a first cut. A senior recruiter improves it in fifteen minutes. The benefit is not just speed; structured interviews are fairer than free-form ones, and structure is exactly the kind of scaffolding the model produces cheaply.
6. Translating internal communications across English and Nepali. Most mid-sized Nepali organisations operate in a register-mixed environment — English for board decks, Nepali for the shop floor, both for the all-hands. The model handles routine translation in both directions competently. Idiomatic and formal-register translation needs a human pass; routine memos, meeting summaries, and policy reminders can go through with a quick read-over.
What AI does badly, or unsafely
The same technology fails on a different cluster of tasks — and it fails most dangerously when it looks like it is succeeding.
1. Final hire / no-hire decisions. Whatever the vendor says, no current AI system is capable of making a hiring decision in a way that is fair, explainable, and defensible against a discrimination claim. Models trained on past hiring data inherit the biases of that data; models trained on global data inherit US assumptions about names, schools, and career trajectories that misread Nepali candidates routinely. The hire decision sits with a human who can be held accountable. The model can prepare the file; it cannot sign the offer.
2. Auto-rejecting candidates without human review. A surprising number of “AI screening” tools effectively auto-reject CVs that fall below a threshold. In Nepal this is particularly dangerous: a strong candidate from Karnali or Madhesh whose CV uses different formatting conventions, different school names, or a Nepali-medium degree the model has never seen will be silently filtered out. Every rejection should be reviewable by a human, and a sample should be reviewed in fact, not just in policy.
3. Computing performance scores unsupervised. Asking a model to read a year of Slack messages, calendar events, and email and produce a performance score is not a productivity feature; it is a lawsuit. Even setting aside the data-protection issues, the model has no internal calibration for what “good performance” means in your context. It will produce a number that looks objective and is in fact a confident guess. Performance assessment stays with managers and HR using observable, agreed criteria.
4. Citing specific sections of the Labour Act 2074 without verification. The model has read about the Labour Act. It has not necessarily read the current rules and the latest amendments. It will confidently quote a section number that does not exist, or attribute a notice period to the wrong category of employee. For any guidance that turns on a specific provision — termination notice periods, gratuity calculation, leave entitlements — verify against the bare act or the Ministry of Labour’s published rules. Treat any citation the model gives you as a lead, not a source.
5. Replacing the relational and empathic work. The hardest parts of HR — the difficult conversation with an underperforming manager, the support call to an employee dealing with bereavement, the negotiation between two leaders who cannot agree on a candidate — are not document-production tasks. They are acts of presence and judgement that require a human who knows the people and carries the relationship. The model can prepare your notes. It cannot have the conversation.
The verification-cost-plus-downside test
Combine the two lists and one rule falls out. Before you use AI on any HR task, ask two questions, not one. First: if the output is wrong in a way I do not immediately spot, how long would it take me to catch the error against a trusted source? Second: if the error reaches a candidate or an employee, how much does it hurt them? Fast verification, small downside — use the model freely. Slow verification, large downside — the model is the wrong tool, no matter how convincing it sounds.
This is not a counsel of timidity. Most HR work by hours — drafting, summarising, translating, structuring — clears the test comfortably. But HR’s high-stakes hour is genuinely high-stakes in a way that, say, marketing’s is not. Knowing which hour is which is the professional contribution this course is asking you to make.
Check your understanding
Quick check
—Which of the following is the safest, most defensible use of AI on a typical day for a Nepali HR team?
Quick check
—A new HR manager at a Lalitpur BPO wants to build a defensible day-one hiring kit using AI. Which mix is most appropriate?
What comes next
Knowing which tasks AI does well in the abstract is half the picture. The other half is knowing where in a real HR month those strengths actually land — which stages of sourcing, interviewing, onboarding, and exit change shape, and which do not. The next section walks through a typical month for a small Nepali HR team and marks the map.