Chapter 05 · Section I · 17 min read
Drafting and refreshing HR policies
A policy is a contract the company is writing with its future self — AI can draft it quickly, but only if you give it the right inputs, hold the defined terms steady, and verify every Labour Act reference before the document ever reaches a staff inbox.
A policy is not an email. It is a document that, once issued, the company will be held to — by staff, by the labour office, and sometimes by a court. The leave policy you publish on Bhadau 1 becomes the policy a manager points to when refusing a casual leave in Magh, the policy an aggrieved employee shows the labour officer when they file a grievance, and the policy a partner buyer asks to see during a due-diligence review. AI can draft a serviceable policy in fifteen minutes that would have taken a Friday afternoon to write longhand. But the speed is a trap if it lets you skip the three things a policy actually needs — accurate facts about the firm, accurate references to the Labour Act 2074, and a defined-terms discipline that the model, left alone, will quietly break.
What the model needs before it writes a single sentence
The pattern that works for policies is the same fact-first pattern you have used for JDs and outreach, with two additions: the previous policy (or a comparable template) and the relevant Labour Act provisions. Without these, the model will produce something that reads well and means little — generic clauses lifted from training data, written for jurisdictions that are not Nepal.
The minimum input block for any policy refresh:
I am drafting a revised leave policy for a Nepali private company with 64 employees. Use only the facts below. Do not invent any benefit, entitlement, or procedure not provided.
Previous policy: [paste current policy in full, or attach]. The changes I want: increase paternity leave from 7 to 15 days; add 3 days of bereavement leave for parents and parents-in-law; clarify that festival leave for Dashain is 5 working days, for Tihar 3, for Chhath 2 (Madhesh staff only), for Lhosar 2 (Sherpa, Tamang, Gurung staff), and for Eid 2 (Muslim staff). Other leaves unchanged. Relevant Labour Act 2074 provisions: paste the actual text of sections governing maternity leave (98 days, of which 60 paid), paternity leave (15 days), sick leave (12 days/year accruing), home/annual leave (1 day per 20 worked), and public holidays. House tone: plain Nepali-English bilingual, no legalese where avoidable, written as instructions a manager can actually apply. Defined terms to use consistently: “Employee” (not “team member” or “staff member”), “Leave Year” (Shrawan 1 to Asar end), “Working Day” (a day on which the office is open, excluding gazetted holidays), “Line Manager” (not “supervisor” or “reporting manager”).
The output, given that input, is something you can hand to a legal reviewer with a clean conscience — it sticks to your facts, uses your defined terms, and references real Labour Act provisions because you supplied the text. If you skip any of those inputs, the model fills the gap with plausible invention. Plausible invention in a policy is the same word as future grievance.
The defined-terms discipline
The single most under-noticed failure mode in AI-drafted policies is term drift. The model, left to its own devices, will write “Employee” in the first paragraph, “team member” in the third, “staff member” in clause 7, and “the worker” in the dispute-resolution annex. To a reader skimming the document, these feel like synonyms. To a labour officer reading carefully — or a lawyer for a former employee — they are four different terms, and the question becomes: do they mean the same thing? Were “team members” intended to be excluded from the protections that apply to “Employees”? The answer is usually “no, we just got sloppy”, but the company is now arguing about its own document.
The fix is one line in the prompt and one habit in review. The prompt line: “Use these defined terms consistently throughout — do not substitute synonyms. If a defined term has not been provided for a concept you need, use the concept word in title case and flag it at the end as a term I need to define.” The review habit: before publishing, do one Ctrl-F pass for each defined term, and a second pass searching for the synonyms the model is most likely to slip in — “staff”, “team member”, “personnel”, “worker”, “associate”. If any of them appear, decide deliberately whether to replace or define.
Labour Act verification — every reference, every time
Models hallucinate section numbers. This is not occasional, it is routine. Ask a chatbot to draft a Nepali leave policy citing the Labour Act 2074 and it will confidently cite sections — 28, 41, 53, 84 — at least one of which will be wrong, missing, or paraphrased into a meaning the Act does not carry. The model has read fragments of the Act in training, has read summaries of it on third-party websites with their own errors, and has filled gaps by analogy with Indian labour law, which is not the same Act. The result reads like authority and is partially fiction.
The verification rule is simple and non-negotiable: every Labour Act 2074 reference in the draft must be checked against the actual Act before publication. Open the gazette text or a verified copy — the Department of Labour publishes one, and the Federation of Nepalese Chambers of Commerce and Industry maintains a clean reference — and for each cited section, confirm three things: the section exists, the section number is correct, and the substance of what the policy says matches what the section actually provides. If any of the three fails, fix the reference or remove the citation.
The same rule extends to the Employees’ Provident Fund, Social Security Fund, and Bonus Act provisions. EPF contribution rates, SSF coverage scope, bonus eligibility thresholds — the model will cite numbers, and the numbers will sometimes be last year’s, or last reform’s, or a confident hybrid of both. Verify against the current rules. The hour this takes is the cheapest insurance the company will ever buy on a policy.
Legal and compliance review stays — clean writing is not a substitute
The temptation, once a policy reads well, is to treat it as ready. It is not. A policy that reads well and is internally consistent can still be unenforceable, can still conflict with the Labour Act, can still create unintended contractual obligations the company did not mean to take on, and can still trigger SSF or EPF compliance issues the HR generalist would not spot. The model produces fluent prose; it does not produce legal opinion, and reading fluent prose makes humans (including HR humans) over-trust the document.
The rule the firm should hold to: every new or substantively revised policy goes to a labour lawyer or qualified compliance reviewer before publication. The cost of an hour’s review is a small fraction of the cost of one wrongly drafted termination clause that the company then has to defend at the labour office. AI compresses the drafting time from a day to an hour — spend a portion of the savings on review. The published policy is the company’s commitment for the next several years; spending three thousand rupees on a reviewer who catches one problem is the easiest decision in the calendar.
A second-order point: published policies create expectations even when not strictly contractual. If your leave policy says “additional Chhath leave for Madhesh staff”, the firm has just made a commitment that will be enforced socially even before it is enforced legally. Be deliberate about what you publish. The model will happily generate generous-sounding clauses because generous sounds good; the firm is the one that lives with them.
Five policies a Nepali SME should draft or refresh first
If the firm is starting fresh, or doing a policy audit, these are the five where AI-assisted drafting gives the largest return for the smallest risk — provided the verification discipline holds.
1. Leave policy. The most-read policy in any company. Cover annual leave, sick leave, casual leave, public holidays, maternity (98 days, 60 paid, per the Act), paternity (15 days), bereavement, festival leave varying by community (Dashain, Tihar, Chhath, Lhosar, Eid), unpaid leave, and the approval procedure. State the Leave Year clearly (most Nepali firms use Shrawan 1 to Asar end; some use the Gregorian calendar; whichever you pick, define it).
2. Code of conduct. Workplace behaviour expectations, conflicts of interest, gift acceptance, social media conduct, confidentiality. Pair it with a clear breach procedure — what happens at first offence, second offence, gross misconduct — so the policy is operational, not aspirational.
3. Remote-work and hybrid policy. Newer ground for most Nepali firms post-2020. Cover eligibility (which roles can, which cannot), expectations (working hours, availability, response times), equipment (who provides what), home office allowance if any, and the in-office days expected. Be explicit; ambiguity here creates friction monthly.
4. Grievance procedure. How an employee raises a concern, who hears it, what the timeline is, what protection from retaliation exists, and what the escalation path is if the internal process does not resolve it. A grievance policy with no actual steps is worse than none — it signals to staff that the firm prefers complaints to disappear.
5. Anti-harassment policy. Required by the Sexual Harassment at Workplace (Prevention) Act 2071. Cover the definition, the complaints committee composition, the investigation procedure, confidentiality, protection from retaliation, and consequences. This is the policy where a wrong cite or vague clause does the most damage in actual use; verify it twice and have a labour lawyer review it once.
Check your understanding
Quick check
—Your CEO asks you to draft a refreshed leave policy by Friday. You have last year’s policy, a list of changes the management team agreed on, and access to a chatbot. Which workflow is most likely to produce a policy that is both fast to draft and safe to publish?
Quick check
—True or false: a policy draft produced by AI that is clearly written, internally consistent, and uses correct-looking Labour Act 2074 citations can be published without legal or compliance review.
What comes next
A clean policy is the foundation. The harder question is what you do once the policy exists — how you announce it, how you handle the harder changes (restructures, redundancies, leadership transitions) that no policy makes easier, and how you keep AI in its proper role as drafter, not as voice. The next section is about internal announcements and change comms: where AI helps, where it must not be the voice, and the no-surprises rule that AI makes both easier to honour and easier to break.