The HR workflow, mapped to where AI helps

The month, before AI

A typical month at Himal Voice looks roughly like this. Week 1: the manager and recruiters sit with the operations head to scope the quarter’s needs, finalise JDs for the new openings, and post on Merojob, Kumari Job, LinkedIn, and the firm’s own careers page. The generalist runs the previous month’s payroll, processes provident fund, and answers the queue of leave and grievance queries that built up. Week 2: sourcing is in full swing — recruiters are screening hundreds of CVs for the entry-level roles, doing outbound outreach for the engineer roles, and scheduling first-round telephonic interviews. Week 3: interview week. Panel rounds for the engineer roles, structured assessments for the reps, reference checks for the offers about to go out. Week 4: offer rollouts, onboarding kits for the previous batch’s joiners, the new-hire orientation session, an updated leave-policy circular, and the partner’s monthly people-metrics review.

Hours, very roughly across the four-person team: forty on sourcing and screening, sixty on JDs, outreach, and scheduling, fifty on interviews and references, thirty on onboarding and policy comms, twenty on payroll and grievances, twenty on the partner’s reporting and review. That is one team’s month. Replace “BPO in Lalitpur” with “INGO programme office in Surkhet doing a Far-West programme-officer recruitment” or “commercial bank in Kathmandu hiring relationship managers across three provinces” and the shape changes but the stages do not.

Stage by stage, with AI in the loop

Walk through the same month now, but assume the team has a decent paid chatbot, an ATS that integrates an AI extractor, and a translation tool tuned for Nepali office register.

1. Workforce planning and JD drafting — high leverage. Scoping the quarter is a human conversation that does not benefit much from AI. But once you have the brief, the model produces a JD in minutes that the recruiter edits in ten. Multiply that across the four to seven JDs a quarter and the hours add up. The risk is low because every JD is human-reviewed before posting, and the cost of a mediocre JD is a small one — fewer or worse applicants, not a hurt candidate.

2. Outreach and sourcing — high leverage. Personalised LinkedIn messages to passive engineering candidates, follow-up emails to people who half-applied, the careers-page paragraph for a new role — all of these are pattern-based prose where the model produces drafts faster than a recruiter can type. The recruiter still picks the candidates, still reads the message before sending, still owns the relationship. But the typing collapses.

3. CV screening — medium leverage, high risk. This is where most “AI in HR” sales pitches concentrate, and where Himal Voice should be most careful. The model can usefully summarise a CV against the JD’s requirements and produce a structured one-paragraph note — “five years in tier-2 BPO, two years team-lead, weak on Nepali typing, lives in Kapan.” That summary saves the recruiter time when she is reading her two-hundred-and-fiftieth CV. What the model must not do is auto-reject. The recruiter reads every summary and every CV that the model flagged as borderline. The minute the team stops reading the borderline cases, silent discrimination has begun.

4. Interview scheduling and logistics — high leverage, low risk. Coordinating five panellists across three time zones for an engineering loop, sending the calendar invites in the right language, generating the interview brief for each panellist with the candidate’s CV summary and the rubric — all of this is exactly where AI tooling shines. Errors are easy to catch (a wrong meeting link surfaces immediately) and easy to fix.

5. Interview question generation and rubric structuring — high leverage. As discussed in section one, the model produces decent first-cut behavioural and technical questions, structured rubrics, and probing follow-ups. The recruiter and the hiring manager refine them. The benefit is not just speed — structured interviews are demonstrably fairer than free-form ones, and AI lowers the cost of building structure to almost nothing.

6. The interviews themselves — zero leverage. This is where the workflow map gets surprising. The actual interview — sitting across from a candidate, asking the question, listening to the answer, watching how she handles a tough probe — does not change at all. No AI sits in the room. (We will discuss “AI interview bots” in a later chapter; for now, the short answer for most Nepali HR contexts is: do not use them, and if you must, do not use them at the screening stage.) Hours unchanged.

7. Reference checks and background verification — low leverage. A reference call is a relationship and a judgement, not a document. The model can help draft the questions you will ask the referee, summarise notes after the call, and structure the file. The call itself stays with the recruiter.

8. Offer rollout — low leverage, contractual risk. The offer letter template can be generated by AI from a few inputs. But every clause — notice period, probation, salary structure, gratuity treatment, non-compete — must be checked against the Labour Act 2074 and your firm’s standard. Do not use a model-generated offer letter without legal or experienced-HR sign-off.

9. Onboarding and orientation — medium leverage. Welcome packs, system-access checklists, the orientation slide deck, the bilingual employee handbook excerpt — all of these can be drafted by AI in a fraction of the time. The orientation session itself, where the new hire meets the team and asks the awkward questions, stays human.

10. Policy and internal comms — high leverage. The leave-policy circular, the new-system announcement, the quarterly all-hands email — first drafts in minutes, finalised in fifteen. Especially valuable for the Nepali-language version, which often gets written second and rushed; AI lets you write both versions in parallel.

11. Performance management — low leverage, very high risk. Managers’ quarterly reviews of their team members are a place where AI is being aggressively marketed and where Himal Voice should be most cautious. The model can help the manager structure the conversation, draft talking points from the manager’s notes, and tidy up written feedback. The model must not generate the rating, score the employee from observed data, or write feedback the manager has not actually thought through. The conversation is the work; the document is the artefact.

12. Grievances, conflict, and exit — minimal leverage, maximum sensitivity. A grievance interview, a conflict mediation, an exit conversation with someone being let go — these are the moments HR exists for. AI’s role is limited to helping you prepare your notes and draft the post-meeting summary. The meeting is yours.

What the new month looks like

Add it up across Himal Voice. The recruiters spend less time on JD drafting, outreach typing, and CV-summary writing, and more time talking to candidates, calibrating with hiring managers, and reading the borderline CVs the model flagged. The generalist’s policy and onboarding drafting time collapses; her grievance and payroll time does not move. The HR manager spends less time on metrics deck production and more time on the conversations the metrics deck is meant to support. The partner’s review meeting starts from a cleaner package and runs more substantive.

The team has not shrunk. The work has redistributed. Roughly thirty to forty percent of the document-production hours have moved into more skilled work — judgement, calibration, conversation, exception handling. That is the actual productivity story, and it is a good one. It is also a story that quietly assumes the team is using AI on the right tasks and avoiding it on the wrong ones. The next section is about why getting that line right matters more for HR than for almost any other function.

Check your understanding

Quick check

—

Match each hiring stage to its honest AI leverage at Himal Voice. Which of the following is correct?

1. AJD drafting — low leverage; CV auto-rejection — high leverage; offer letter — high leverage; the interview itself — high leverage.
2. BOutreach drafting — high leverage; CV summarisation — medium leverage; the interview itself — zero leverage; performance-rating generation — high leverage.
3. COutreach drafting — high leverage; CV summarisation — medium leverage with risk; the interview itself — zero leverage; performance-rating generation — low leverage with very high risk.
4. DEvery stage of hiring is high leverage for AI, including the interview and the final hire decision.

Check answer→

Quick check

—

A new HR manager at Himal Voice is excited about AI and asks where it should NOT be used. Which list best captures the work that does not move?

1. AJD drafting, outreach emails, and interview question generation — these are too important to delegate.
2. BThe interview conversation itself, reference calls, grievance and conflict meetings, and the final hire / no-hire decision — these are relational or decisional acts AI cannot perform.
3. COnboarding paperwork and orientation slides — these contain confidential information.
4. DTranslation between English and Nepali — the model is unreliable in Nepali, so do all translation manually.

Check answer→

What comes next

A workflow map tells you where AI fits. It does not tell you why the verification line is so unforgiving in HR specifically. The next section makes that case directly — why HR cannot treat AI tooling the way Marketing or Sales does, what the Constitution and the Labour Act actually require, and which Nepali-context biases — gender, caste, region — make the stakes sharper than most imported AI playbooks acknowledge.