AI in Recruitment: A Practical Guide for Indian SMBs

If you run hiring at a 40-person startup in Bengaluru or a 180-person manufacturing firm in Pune, you have probably already had the conversation. Someone on the leadership team read about AI agents doing sourcing and screening end to end, and now wants to know why two recruiters still take six weeks to close a mid-level backend role. Meanwhile your inbox holds 900 applications for that role, most of them plausible, many written by the same three chatbots. AI in recruitment stopped being a future topic and became an operational one — not because the technology is magic, but because it changed the volume and shape of the work on both sides of the table.

The honest position sits between the vendor pitch and the cynic's shrug. AI is genuinely good at a narrow class of recruiting work: reading unstructured text, drafting first versions, matching documents against written criteria, and handling repetitive coordination. It is genuinely bad at judgement calls that carry legal and reputational weight — deciding who deserves a conversation, weighing an unusual career path, reading whether someone will thrive on your team. Most disappointment with AI recruitment tools in India comes from companies that pointed the technology at the second category and hoped.

This guide walks the hiring funnel stage by stage and says plainly what to automate, what to keep human, and how to set each piece up inside an ATS. It covers the flooded-inbox problem that AI-written applications created, a realistic stack for a 20–200 person company, reusable prompt patterns, guardrails around bias and data privacy, a metrics table, and a 30/60/90-day rollout plan.

Why AI hit recruiting first

Recruiting was always going to be first, and the reason tells you where it will help.

Recruiting runs on unstructured text. Resumes, JDs, cover notes, hiring manager briefs, interview scribbles. Almost none of it sits in tidy database fields, and language models are machines for reading and writing exactly that. Payroll, by contrast, needs exact answers with legal consequences — a much worse fit for a probabilistic system.

The top of the funnel is high-volume and repetitive. Scanning 600 resumes is the same small task 600 times. Classic automation profile. The bottom of the funnel — final interviews, negotiation, closing — is low-volume and high-stakes, the opposite profile.

First drafts are expensive; edits are cheap. Writing a JD from scratch takes 45 minutes. Editing a decent draft takes eight. Multiply that across JDs, outreach, interview guides and rejection notes.

And candidates moved first. Jobseekers adopted AI faster than employers because the cost is zero and the incentive direct. One person with a chatbot produces forty tailored applications in an evening. Your funnel has already changed shape whether or not you adopt anything, which is why "wait and see" is not a neutral position.

What AI is actually good at

• Summarisation and extraction. A five-page resume into a structured profile. Forty minutes of notes into a scorecard.
• Drafting. JDs, outreach, screening questions, interview guides, rejections, offer explainers.
• Matching against explicit criteria. Given a written rubric, sorting applications by evidence per criterion — and showing its working.

Everything else being sold to you is one of these three wrapped in a workflow, or a claim to be sceptical about.

What it is bad at

• Predicting performance from a resume. There is no reliable ground truth. The training signal — who got hired, who got promoted — carries every bias your industry already has.
• Assessing what it cannot see. Judgement, coachability, how someone handles disagreement. These surface in conversation, not documents.
• Being accountable. "The model decided" is not an answer you want to give a candidate, a regulator or a journalist.

The hiring funnel: what to automate, what to keep human

Each stage below gets the same treatment: what AI does well, what stays human, how to set it up. Read it as a menu. A 30-person company will implement half of this and be well served.

Stage 1: Workforce planning and role definition

AI does well: turning mess into a structured brief. Most SMB hiring starts with "I need another backend guy, senior-ish, done payments." Feed that plus a five-minute intake transcript into a model with a fixed template and you get a draft covering the problem the role solves, 90-day outcomes, must-haves, level, and the questions you forgot to ask.

Stays human: whether the role should exist, what band it sits in, and whether you are solving a hiring problem or a process problem. AI has no view on your runway.

Set it up: make an intake form mandatory before a requisition opens — business problem, 90-day outcomes, must-haves (max five), nice-to-haves, level, band, panel, target start. The hiring manager approves the draft in writing, and that approval becomes the source of truth for the JD, the rubric and the scorecard. Consistency down the whole funnel, for free.

The highest-leverage move here has nothing to do with AI: cap must-haves at five.

Stage 2: Job description writing

AI does well: a solid first draft in your house style, plus length rewrites, exclusionary-language checks, role-specific 90-day sections, and short variants for boards, LinkedIn and WhatsApp from one source.

Stays human: the specifics that make someone want the job — the product they own, team size, how hybrid actually works. And whether you publish a salary band. You should think hard about publishing one; a stated band is the strongest filter available and costs zero recruiter time.

Set it up: keep a house-style file — three of your best JDs, tone rules, standard sections, EO boilerplate and benefits. Every prompt references it. Store approved JDs as ATS templates so the next similar role starts at 80%.

A sceptical note: AI-written JDs converge. If everyone in your city generates from the same models, every JD reads the same and candidates stop reading. Differentiation comes from human-supplied specifics. Treat the model as a formatting engine, not a source of ideas about your company.

Stage 3: Sourcing

Sourcing is where "AI agents" are pitched hardest. Approach with care.

AI does well: generating Boolean permutations you would never think of; expanding a role into adjacent titles and skill synonyms, which matters in India where titles vary wildly between service firms, product companies and startups; summarising a long profile into fit and gaps; deduplicating records across sources.

Stays human: who to approach, and the relationship. Obviously templated messages get ignored, and worse, burn the person for the next role. In a market as networked as Indian tech hiring, a spray-and-pray reputation travels.

Set it up: use AI for the search string and profile summary; write or heavily edit the first message yourself. Cap outreach volume deliberately and track reply rate, not sends. Keep a suppression list so nobody gets contacted by three automations. If an agent drafts outreach, require human approval before send for the first several hundred messages — you will find things you do not want going out under your name.

On agentic sourcing specifically: a tool that autonomously decides who to contact is making a selection decision, and selection decisions carry fairness exposure. An agent optimising for "people who look like people we hired before" narrows your funnel in ways you will not notice for two quarters. Constrain agents to search and draft. Keep selection with a person.

Stage 4: Application handling

This is where recruitment automation for SMBs pays for itself fastest, with the least judgement at stake.

AI does well:

• Parsing resumes into structured fields, including the multi-column PDFs that used to break parsers.
• Deduplicating the same person across three roles or a reapplication after eight months.
• Auto-acknowledging every application within minutes. Sounds trivial; it is the most-cited candidate complaint in Indian hiring and costs nothing to fix.
• Tagging and routing — fresher, lateral, wrong city for an on-site role.
• Knockouts on stated, factual, job-relevant criteria: a legally required licence, willingness to work from your Chennai office, notice period compatibility.

Stays human: every judgement-based rejection. There is a bright line. "Candidate answered No to 'Can you work from Chennai three days a week?'" is a fine automated reject. "Model scored resume 4/10 on culture fit" is not.

Set it up: replace the free-text cover letter with three to five structured, role-specific questions. Send auto-acknowledgement with a realistic timeline and then meet it. Phrase knockouts unambiguously and show them to candidates before submission. Every week, sample twenty auto-rejections — if you find people you would have wanted to talk to, your knockouts are wrong.

Stage 5: Screening and shortlisting

The heart of it, and where resume screening AI is most oversold.

AI does well: reading every application against a written rubric and producing a per-criterion assessment with evidence. Note the phrasing. Not "score this candidate 78/100." Instead: "Criterion 3, production experience with distributed systems — here is the line that evidences it, or no evidence found." That output is reviewable. A single opaque score is not.

The value is not that the machine picks better than you. It probably does not. The value is that it reads all 900 applications with equal attention, whereas human attention degrades after the first hundred and Friday-evening applications get read differently from Tuesday-morning ones.

Stays human: the shortlist, always. AI produces a ranked, evidenced queue; a recruiter reads down it, samples from the middle, and decides. Every rejection here is a human decision.

Set it up:

1. Write the rubric before you see applications. Four to six criteria drawn from the approved brief, each with an explicit definition of what counts as evidence.
2. Weight criteria explicitly, summing to 100.
3. Require evidence citation per criterion and "no evidence found" rather than inference. Models will happily infer that someone from a fintech has payments experience — exactly the pattern-matching that encodes pedigree bias.
4. Screen blind. Strip name, photo, age, gender, address and college name before the model sees the document. College name carries a very heavy and unhelpful signal in Indian hiring; it tracks family income far more tightly than capability.
5. Review the queue, not just the top. Read the top 30, then ten from the 30th–60th percentile. That is how you catch systematic error.
6. Log the rubric version, model output, reviewer and decision. You want to reconstruct any decision six months later.

On AI interview screening. Asynchronous video scored by AI is the most contested tool in the category. Useful parts exist: transcription, structured question delivery, letting candidates respond on their own schedule, giving the panel one recording. Indefensible parts also exist: scoring on facial expression, tone or speech patterns. Those correlate with regional accent, first language, neurodivergence and disability. In India, accent scoring is a straightforward proxy for region and class. Use async video for content; let humans assess.

Stage 6: Assessments and work samples

AI does well: helping design an assessment that measures the actual job — the task, the marking rubric, worked examples of strong versus weak answers — plus first-pass marking on structured submissions.

Stays human: final grading, every borderline case, and any decision from a suspected-AI flag.

Set it up: design assessments assuming candidates have AI. That is no longer optional. Two approaches work:

• Assume AI and test what remains. Allow assistance openly, then assess what AI does poorly: which trade-off the candidate chose, the quality of their assumptions, whether they spotted that the brief was ambiguous. A strong version gives a deliberately flawed spec and sees who flags it.
• Test live. A 45-minute working session where the candidate reasons aloud. Expensive in interviewer time, so use it late on few people.

What fails: a take-home that AI completes in ninety seconds, marked as if it had not. That measures prompt access, not capability. And treat AI-detection flags as unreliable — they over-flag non-native English writers, which in India means penalising exactly the wrong people. Never reject on a detection score; use it at most as a reason to ask a follow-up live.

Stage 7: Interview scheduling

Automate hard. Almost no downside, immediate and large time savings.

AI does well: reading availability across panel calendars, proposing slots, sending invites with joining links and prep material, reminders, reschedules, timezones, chasing unconfirmed interviewers, rebooking dropouts.

Stays human: conversations where scheduling itself is sensitive — repeated reschedules, someone interviewing secretly around a current job, a senior candidate needing white-glove handling.

Set it up: self-scheduling links tied to real panel availability; interview kits attached to invites so the panel gets the rubric and their questions automatically; reminders at 24 hours and one hour; escalation if a slot is unconfirmed for 48 hours; and a named human the candidate can reach when automation fails, because it will.

Scheduling is usually where hiring workflow automation delivers the clearest recruiter-hours win. For lean TA teams, coordination is often the single largest consumer of time, and almost none of it requires judgement.

Stage 8: Interview support and structured scorecards

AI does well: generating role-specific questions mapped to rubric criteria with good follow-up probes; transcription; drafting a scorecard from transcript and rough notes, organised by criterion with quotes; flagging when a panel did not cover a required criterion.

Stays human: the interview, the rating, the decision. Non-negotiable.

Set it up:

• Build an interview kit per role: which criteria each interviewer covers, three or four questions each, and what strong versus weak answers look like. Structure is the highest-return, lowest-cost intervention in hiring quality — more than any tool.
• Interviewers rate and write evidence before seeing anyone else's scores. This kills anchoring.
• Let AI draft the write-up from notes; the interviewer edits and signs off. Saves 15–20 minutes per interview and beats the two-line notes most panels leave.
• Get explicit consent before recording or transcribing, and offer a no-recording path.
• Run debriefs with the AI summary as a reference document, never the decision. The subtle failure: panels read the summary instead of remembering the conversation, and nuance disappears.

Stage 9: Offer

AI does well: drafting offer letters from approved templates; producing a plain-English CTC breakdown, a real pain point in India where variable pay, ESOP, gratuity and reimbursements confuse candidates and cause late drop-offs; building comparison summaries for approvers; chasing signatures.

Stays human: the offer conversation, negotiation, anything involving counter-offers. This is a relationship moment, not a document moment.

Set it up: templated offers with locked legal clauses and variable fields; approval workflow with defined thresholds; an automatic one-page "what this offer means, month by month" explainer alongside the formal letter; and automated check-ins between acceptance and joining. That window is where Indian hiring leaks most, especially with 60–90 day notice periods.

Stage 10: Pre-boarding handoff

The stage most often broken in SMBs, because it sits between two owners and neither picks it up.

AI does well: generating document checklists by role and location, chasing missing documents, creating the employee record from ATS data without re-keying, assigning tasks to IT, finance and the manager, drafting the welcome note and first-week schedule.

Stays human: the manager's personal message and the buddy assignment — anything signalling that a person is expected and wanted.

Set it up: one trigger, offer accepted, fires the whole sequence. The candidate record becomes the employee record with no re-entry. Documents flow through a portal, not email attachments. Every task has an owner and a due date relative to joining. And one non-automated human message before the start date.

This handoff is exactly where a combined hiring-and-HRMS system beats a stitched-together toolchain: no export-import step, nothing to fall between systems.

The flooded inbox: when candidates use AI too

The pattern Indian recruiters describe: volume up sharply, apparent quality up — better formatting, tighter keyword alignment — and actual signal down, because variance collapsed. Everything reads competent. Nothing distinguishes.

That is a measurement problem, not a volume problem. Your instinct will be to filter harder. Resist it.

Why filtering harder backfires

Tighter keyword filters select for candidates who optimise for keyword filters, which is now trivial with a chatbot. You filter out the honest, un-optimised applicant and keep the well-tooled one — the opposite of what you want. More rounds add cost on both sides and lower acceptance. Raising the experience bar excludes career changers and returners, often your best value hires.

Better signals, not more filters

Structured application questions. Three or four short, specific, role-relevant questions capped at 100–150 words each.

Weak: "Why do you want to work here?" — answerable by any chatbot given your website.

Strong:

• "Our support team handles about 300 tickets a day with four people. What would you look at first in your first two weeks?"
• "Describe a time you disagreed with a decision your manager made. What did you do?"
• "What part of this role do you expect to find hardest, and why?"

AI can help answer these too, but the answers differentiate, and they give you a real conversation starter. Someone who submits a polished answer they cannot discuss for two minutes on a call reveals themselves immediately.

Small, real work samples early. Thirty minutes, tied to the actual job, sent wider than you would normally interview. Real effort self-selects for genuine interest and gives you comparable artefacts. Keep it short — anything over 90 minutes at an early stage is extractive.

Verifiable over claimed. Shipped work, public repositories, published writing, a reference who will take a five-minute call. Claims are cheap now. Artefacts are not.

A short live conversation earlier. Ten to fifteen minutes, five standard questions. Faster than carefully screening 200 written applications, and far more informative. Many lean teams are deliberately inverting the funnel for this reason.

Referrals, actively. A push per role, simple form, real bonus. A referral carries a human's reputation, which AI cannot manufacture. Watch it for homogeneity and pair it with open sourcing.

Reframing the arms race

"Candidates use AI so we need better AI to catch them" leads somewhere bad — both sides escalate, both spend more, nobody learns anything.

Better frame: AI made written self-presentation nearly free, so written self-presentation stopped being informative. Move signal collection to where it still costs something — a real conversation, a real task, a real reference. That is a return to fundamentals, not a purchase.

A practical automation stack for a 20–200 person company

You do not need eight tools. Most companies this size are best served by a small stack with the ATS as system of record.

Buying principles:

• Fewer systems. Every integration is a place data goes stale, duplicates or leaks. One adequate system beats four best-of-breed tools you must keep in sync, especially with no HR ops person.
• Insist on explainability. If a vendor cannot show why a candidate ranked where they did, in words you can repeat to that candidate, do not buy it.
• Check where data lives. Stored where, processed where, leaves India or not, used for model training or not, deletion how. In the contract, not the sales call.
• Demand a human-in-the-loop switch. If auto-reject cannot be disabled, that is disqualifying.
• Test on your own data. Pilot on a role you already closed. Does it rank your actual hire highly? Does it resurface people you rejected for good reason? A demo on vendor sample data tells you nothing.

Sequencing from spreadsheets and email: ATS with structured stages first; then auto-acknowledgement and templated comms; then self-scheduling; then scorecards and interview kits; then AI-assisted JDs and screening; then analytics and fairness monitoring. The first four involve almost no AI. That is intentional — AI applied to an unstructured process amplifies the disorder.

Prompt patterns you can reuse

Starting points, not magic strings. The common thread: give the model a role, explicit criteria, a required output format, and permission to say "I don't know."

JD drafting

``` You are helping write a job description for an Indian SMB. Plain English, short sentences, no jargon, no superlatives about the company.

ROLE BRIEF: - Title: [title] - Team and reporting line: [details] - Business problem this role solves: [2-3 sentences] - Outcomes expected in first 90 days: [3 bullets] - Must-have skills (max 5): [list] - Location and work model: [details] - Salary band we will publish: [range or "not published"]

HOUSE STYLE: [paste two of your existing JDs]

Produce: 1. A JD of 350-450 words: About the role / First 90 days / What we're looking for / Nice to have / How we hire / Compensation and location. 2. A 60-word version for job boards. 3. A 40-word version for a WhatsApp or community post.

Rules: - Invent no facts about the company, product, funding or team size. If you need a detail I have not given, insert [NEEDS INPUT: xyz]. - Do not use "rockstar", "ninja", "fast-paced", "wear many hats". - At the end, list separately any requirement that might unnecessarily exclude candidates (specific degrees, continuous employment, exact years of experience). ```

That last rule is the useful one — models spot over-specified requirements well when asked directly.

Screening rubric

``` Turn this role brief into a screening rubric I can apply consistently across all applications.

ROLE BRIEF: [paste approved brief]

Produce a table of 4-6 criteria. For each: - Criterion name - Why it predicts success in THIS role (one sentence) - What counts as strong evidence in an application - What counts as weak or absent evidence - Suggested weight (weights sum to 100)

Constraints: - Criteria must concern demonstrable skills, experience or outcomes. - Exclude criteria based on college attended, employer prestige, continuous employment, age, gender, location beyond the stated work model, or "culture fit". - If a must-have cannot be assessed from a written application, say so and mark it "assess at interview stage". ```

That last constraint saves trouble. Plenty of things worth hiring for are invisible on paper, and naming them upfront stops you pretending otherwise.

Screening applications against the rubric

``` You are assisting a human recruiter. You do not make decisions.

RUBRIC: [paste rubric with weights] APPLICATION: [paste anonymised resume + structured answers]

For each criterion output: - Criterion name - Rating: Strong / Adequate / Limited / No evidence found - Evidence: the exact phrase or fact supporting the rating, or "No evidence found in application." - Confidence: High / Medium / Low

Then output: - A two-sentence weighted summary - Two questions a recruiter should ask on a screening call to resolve the biggest uncertainty - Anything notable the rubric does not capture

Hard rules: - Never infer skills or experience that are not stated. Do not assume payments experience because a candidate worked at a fintech. - Never comment on or infer name, gender, age, religion, caste, region, marital status or family circumstances. - Do not produce a single numeric score. - If the application is too sparse to assess, say so instead of guessing. ```

Guard the "never infer" line most carefully. Inference is where pedigree bias enters, wearing the costume of reasonable judgement.

Interview question generation

``` Create an interview kit for this role.

ROLE BRIEF: [paste] RUBRIC: [paste] PANEL: [names, roles, criteria each will assess] LENGTH: [minutes per interviewer]

For each interviewer produce: - The criteria they own - 3 primary questions, behavioural or work-sample based (no hypotheticals, no brain teasers, no trivia) - 2 follow-up probes per question that dig for specifics - What Strong, Adequate and Weak answers look like, described as observable behaviour

Also produce: - 3 questions that must NOT be asked, with a one-line reason each - A 5-question standard opener every candidate gets identically ```

Scorecard summarisation

``` Turn these raw interview notes into a structured scorecard.

RUBRIC CRITERIA: [paste] RAW NOTES / TRANSCRIPT: [paste]

For each criterion: - What the candidate actually said or described, with a short quote - Whether the evidence suffices to rate the criterion - If not, say so and suggest what to probe next round

Then list: the 2 strongest pieces of evidence for hiring, the 2 strongest concerns, and open questions for the next stage.

Hard rules: - Do not assign a hire/no-hire recommendation. The interviewer decides. - Add nothing that is not in the notes or transcript. - Do not characterise personality, confidence, communication style, accent or demeanour unless the interviewer noted it as relevant to a listed criterion. - Mark anything uncertain with [UNCERTAIN]. ```

Notice the shared design: no numeric scores, evidence required, decisions withheld, inference forbidden, uncertainty visible. That is the whole philosophy in four rules.

Guardrails: fairness, explainability, privacy

Most likely to be skipped, most likely to cause a real problem. Keep it proportionate — you do not need a compliance department — but do not skip it.

Bias and adverse impact

AI does not invent hiring bias; it inherits and scales it. In the Indian context, name the specific risks:

• College tier. The strongest and most misleading signal in Indian resume screening. It tracks family income and school access far more closely than capability in most roles.
• Employer brand. Preferring big names penalises people from smaller firms, family businesses and early-stage startups, who often did broader work.
• Career gaps. Disproportionately penalise women returning after caregiving, people who dealt with illness, and people affected by layoffs.
• Language. Written English fluency reads as competence. For most roles it is not the same thing.
• Names and location. Both carry signals about region, religion and community that a model reading an un-anonymised resume will pick up.

Mitigations: blind the first pass by stripping identifying fields; forbid inference explicitly in prompts; monitor pass-through rates by stage across whatever categories you lawfully collect, and investigate the criterion rather than the candidates when one group lags; review twenty AI-influenced rejections monthly; and never fully automate rejection on a judgement criterion. That last one is the primary safeguard and makes most other failures recoverable. If you cannot collect demographic data, watch proxies you do have — pass-through by college tier or source channel often reveals the same problems.

Explainability

You should be able to answer three questions about any decision six months later: what criteria was this candidate assessed against, what evidence supported each rating, and who decided and when. Store the rubric version alongside the decision — rubrics change, and a decision only makes sense against the one in force at the time. If your tooling cannot answer these, it is not fit for purpose.

Human-in-the-loop, defined properly

"Human in the loop" is often theatre: a person clicking approve through 200 AI decisions in four minutes. Meaningful review means the human sees the evidence rather than the conclusion, has the time and mandate to disagree, reviews some low-ranked candidates too, and has their override rate tracked. A reviewer who never disagrees is rubber-stamping.

Candidate disclosure

Tell candidates you use AI, in plain language, in the JD or on your careers page. Something like: "We use software to help organise and summarise applications against the requirements listed above. A member of our team reviews every application and makes all decisions about who moves forward."

Candidates increasingly ask, and disclosure norms are tightening across jurisdictions. Getting ahead of it costs you a paragraph. If you record or transcribe interviews, get explicit consent beforehand, state retention, and offer an alternative.

Data privacy under India's data protection regime

India's framework establishes obligations around consent, purpose limitation, data minimisation, security safeguards, breach notification and individuals' rights over their data. Implementation has rolled out in phases, and what applies to you depends on your size, the data you handle and where you process it. Treat this as orientation, not legal advice, and verify current requirements with a qualified advisor. Rules here are actively evolving.

These practices are sensible under essentially any interpretation:

• Collect only what the role needs. Date of birth, marital status, photograph and family details are almost never necessary. Many Indian application forms ask out of habit. Remove them.
• Say why. A short, readable privacy notice at the point of application: what you collect, why, how long, who you share it with, how to reach you.
• Get real consent for anything beyond the immediate application — talent-pool retention, client sharing, background verification, interview recording. Separate, specific, revocable.
• Set a retention period and enforce it automatically. Something like 12–24 months for unsuccessful candidates. Data you kept by accident is data you can be asked about.
• Control and log access. Hiring managers see their roles, not the whole database.
• Know your vendors' practices. Storage location, model training, sub-processors, what happens on termination. In writing.
• Make deletion work end to end across every system — a strong argument for having fewer systems — and have a breach plan written down before you need it.

Hard lines

Regardless of what a tool promises, avoid: fully automated rejection on a model's judgement of quality or fit; scoring on facial expression, voice tone, accent or speech patterns; personality inference from resumes or social media; screening on any protected characteristic or obvious proxy; and any system whose ranking you cannot explain in a sentence.

Measuring whether it worked

"Applications received" is not a success measure. It went up because of AI, and higher volume at flat conversion means you are worse off. Measure the funnel, the time and the quality proxies.

At 20–200 people you lack the volume for statistical rigour, so do not pretend otherwise. Pick four metrics — time-to-hire, screen-to-interview, joining rate, recruiter hours per hire. Baseline for two months before changing anything. Change one thing at a time; switching on AI screening, self-scheduling and new questions in the same week teaches you nothing. Review quarterly, not weekly.

On quality of hire: nobody at your size can measure it properly. The proxies above are good enough to detect a real problem. Anyone selling a precise quality-of-hire score is selling a number with no ground truth behind it.

A 30/60/90-day rollout plan

Assumes 20–200 people, one or two recruiters, currently running spreadsheets, email and maybe a basic ATS.

Days 1–30: structure and baseline

The goal is not deploying AI. It is making your process legible enough that AI can help.

• Week 1. Document your current process end to end for one role. Pull baseline time-to-hire and stage conversion for your last 8–10 hires. Time-audit a recruiter for a week; the result usually surprises people, because coordination is typically the biggest block.
• Week 2. Write intake templates and make them mandatory before a requisition opens. Write screening rubrics and structured application questions for your three most common roles.
• Week 3. Set up or clean up the ATS. Define stages and owners, move every active candidate in, turn on auto-acknowledgement, and build email templates for every stage transition including rejections.
• Week 4. Build interview kits for those three roles. Spend 45 minutes training interviewers on structured interviewing and independent scoring. Publish your candidate AI-disclosure paragraph.

By day 30 you may have used no AI at all, and you will already be faster and more consistent.

Days 31–60: automate the mechanical

• Week 5. Turn on self-scheduling against real panel availability, with interview kits attached to invites, plus reminders and escalation rules.
• Week 6. Start AI-assisted JD drafting from your house-style file, and AI-assisted scorecard drafting from interview notes with interviewer sign-off always.
• Week 7. Pilot AI screening on one role in shadow mode — run it alongside human screening without acting on its output. Where does it disagree? Who does it miss? Is its evidence citation accurate or invented? Set up blind screening at the same time.
• Week 8. Review shadow results with hiring managers and adjust rubrics and prompts. If it performed well, move to assisted mode: AI produces the evidenced queue, a human decides, never auto-reject. Set ATS retention rules and turn on automatic deletion.

Days 61–90: extend, measure, govern

• Week 9. Extend AI screening to remaining high-volume roles with role-specific rubrics. Add small work samples where written applications are not discriminating.
• Week 10. Build your metrics dashboard and compare against the baseline. Run your first audit: sample twenty AI-influenced rejections and review them manually.
• Week 11. Connect the offer-to-onboarding handoff so acceptance triggers document collection, task assignment and employee-record creation with no re-keying. Automate the offer-to-joining check-ins.
• Week 12. Write a one-page AI-in-hiring policy: what is automated, what is not, who reviews what, how candidates are told, how data is retained, who owns it. Survey last quarter's candidates including rejections. Decide what to keep, kill and do next.

Day 90 looks like: every application acknowledged within minutes, screening turnaround under five days, scheduling running without recruiter involvement, every interview producing a structured scorecard, every rejection decided by a human, and candidate data on a defined retention schedule.

Common failure modes

• Automating a broken process. Unstructured hiring plus AI equals faster unstructured hiring. Structure first, always.
• Buying tools instead of fixing workflow. The four-tool stack nobody uses is a familiar sight. Each purchase feels like progress; none changes how work gets done.
• Trusting the score. 82/100 feels objective. It is a compression of judgements you cannot see. Demand evidence per criterion, never a bare number.
• Rubber-stamp review. Track override rates, give reviewers time, make them look at low-ranked candidates.
• Over-filtering in response to volume. Filtering harder selects for optimisation, not capability. Add better signals instead.
• Losing the human touch where it counts. Automate outreach, screening, rejection and offer all at once and candidates notice, acceptance falls, and you will not immediately know why. Automate coordination; keep conversation human.
• Prompt drift. Someone tweaks a screening prompt to fix one case, tells nobody, and screening now behaves differently from last month with no record. Version prompts and rubrics like contracts.
• Ignoring the offer-to-joining gap. You automate everything up to offer, go quiet for a 90-day notice period, and lose the candidate to a counter-offer in week ten.
• No owner. AI in hiring gets adopted by whoever is enthusiastic, who then changes roles. Name an owner and put it in their goals.
• Believing vendor claims. "Reduces time-to-hire by X%" is a marketing sentence, not a measurement of your process. Pilot on your own historical data first.
• Forgetting the rejected majority. They talk, they review you publicly, and some are future customers or future hires. A prompt, respectful, human-decided rejection is a brand investment.

Frequently asked questions

Will AI replace recruiters at a small company?

No, but it changes the job. Coordination, data entry, first-draft writing and bulk document reading shrink substantially. Relationship-building, assessment judgement, hiring manager management, closing and process design grow in proportion. Two people with good automation can handle what used to take four — but they need to be better at the human parts, because that is where the differentiation now sits entirely.

Can I let AI reject candidates automatically?

Only on stated, factual, job-relevant criteria the candidate answered themselves: location, work authorisation, a legally required licence, availability within your timeline. Never on a model's judgement of quality, fit or potential. Those should be human decisions, even if the human decides quickly from an evidenced summary. This one line protects you from most of what goes wrong.

How do I know if my AI screening is biased?

Monitor pass-through rates by stage; if one group advances at a materially lower rate, investigate the criterion driving it rather than the candidates. Sample twenty rejections manually each month. And check what the model cites as evidence — if it keeps referencing college names or employer brands, your rubric or prompt is leaking pedigree bias. Blind screening on the first pass prevents much of this before it starts.

What should I do about candidates using AI to write applications?

Stop trying to detect it and change what you measure. Written self-presentation is no longer informative because it is nearly free to produce. Add structured, specific application questions; add short work samples; move a brief live conversation earlier; weight verifiable artefacts over claims. Detection tools are unreliable and over-flag non-native English writers, which in India means penalising exactly the wrong people.

Do I have to tell candidates I'm using AI?

You should as a matter of practice, regardless of what is strictly required where you operate. A short plain-language line on your careers page covering what the software does and confirming that a human makes all decisions. For recording or transcription, get explicit consent beforehand, state retention, and offer an alternative. Disclosure requirements are tightening in several jurisdictions — verify what currently applies to you.

How long can I keep candidate data?

Set a defined period, document it, and enforce it automatically in your ATS. Twelve to twenty-four months for unsuccessful candidates is a common approach, with anything longer requiring separate consent for talent-pool retention. India's framework emphasises purpose limitation and data minimisation, and specifics depend on your circumstances — verify current requirements with a qualified advisor. As a rule: collect less, say why, delete on schedule, and be able to honour a deletion request everywhere.

What's the smallest useful starting point with almost no budget?

Three things, none needing an AI subscription. Structured application questions instead of a cover letter. A written screening rubric per role, agreed with the hiring manager before applications arrive. Interview kits with independent scoring before debrief. These cost only time and will improve hiring quality more than any tool. Add auto-acknowledgement and self-scheduling next, because they give back the most recruiter hours per rupee. AI screening comes after all of that.

How do I get hiring managers to actually use scorecards?

Make it the path of least resistance. Attach the interview kit to the calendar invite. Let AI draft the write-up from rough notes so filling it in takes three minutes, not twenty. Require the scorecard before the debrief starts, and run the debrief strictly from scorecards. Managers resist structure when it is extra work and adopt it fast when it is less work than what they were doing.

Where this leaves you

The useful version of AI in recruitment for an Indian SMB is unglamorous. It reads applications against a rubric you wrote and shows its evidence. It drafts the JD you edit. It books the interviews. It turns messy notes into a scorecard. It chases documents. It never decides who gets rejected.

The version that causes trouble is the one promising to do the judging — score fit, read personality, rank on a number, hand you a shortlist you did not reason about. That outsources the one part of hiring that carries real consequence, using patterns learned from a history you probably do not want to repeat.

Start with structure. Automate the mechanical. Keep the judgement. Measure whether it worked. Tell candidates what you are doing. Delete data you do not need. That is the whole approach, and it is achievable in a quarter with a team of two.

If you are stitching this together across a spreadsheet, an inbox and three tools that do not talk to each other, that is the problem worth fixing first. CozyHR brings hiring, onboarding and employee records into one system — structured requisitions and application questions, an ATS with scorecards and audit trails, AI assistance that shows its evidence instead of handing you a black-box score, scheduling that runs itself, and a one-click handoff from accepted offer to employee record with no re-keying. Built for Indian SMBs, with the retention and access controls you would otherwise have to bolt on. Have a look and see whether it fits how your team actually hires.