HR is a department where the proportion of repetitive work is unusually high and the cost of delay is unusually visible. A strong candidate who gets no reply for three days accepts a competitor's offer. That is not a hypothesis but an ordinary Tuesday at any company hiring actively.
Below: where the time goes in an HR department at a company of 50–200 people, which four tasks get automated first, and where the line runs beyond which an agent should not be allowed.
Where HR loses time
A typical breakdown with one or two open roles:
- screening applications — three to five hours a day during active hiring;
- answering employee questions — one and a half to two hours a day: when is my leave, where is my payslip, how do I request a certificate;
- onboarding — 8 to 16 hours per new hire;
- scheduling interviews — 30 to 60 minutes per candidate, almost entirely correspondence about timing;
- reporting to management — three to four hours a week.
What all five share: the work is necessary but requires no HR expertise. That is precisely why it can be automated — and why the interview itself cannot.
Agent 1: application screener
What it does: collects applications from every channel (job boards, Telegram, email, the site form), normalises candidate data, checks it against the role's requirements, ranks candidates and hands HR a shortlist with reasoning for each. Clearly unsuitable applicants get a courteous reply the same day.
The effect: a hundred applications get processed in minutes instead of two working days.
Require reasoning from the agent, not a verdict: "matches on stack, but no experience in your industry." A list without explanations cannot be checked, which means it cannot be trusted.
An important limitation. The agent filters on formal criteria — experience, skills, language, location. It does not assess motivation, ability to learn, or whether someone will fit the team. So the correct configuration is "filter out the clearly irrelevant and rank the rest," not "pick the best one."
Agent 2: interview coordinator
What it does: checks interviewers' free slots in the calendar, offers the candidate options, confirms the meeting, sends reminders, renegotiates when something moves, and prepares a short candidate summary for the interviewer beforehand.
The effect: arranging an interview stops taking HR's time at all. This is the simplest of the four tasks and usually the most rewarding — there is almost no room for a costly error.
Agent 3: onboarding assistant
What it does: greets a new hire in the messenger on day one, walks them through the first-week checklist, hands over documents and instructions as they become relevant, reminds them of adaptation tasks, and answers questions about internal rules.
The effect: the new person gets answers immediately rather than whenever their mentor has a spare minute. The side effect matters more than the direct one: the part of onboarding that used to depend on one specific person's availability stops getting lost.
Agent 4: answering employee questions
What it does: answers from internal documents — leave policy, how to request a certificate, insurance terms, procedures. Always with a reference to the source document and an honest "that is not in the knowledge base" when the answer genuinely is not there.
The effect: the bulk of identical questions disappears. Technically this is a RAG system built on your HR documents.
This is also where a common problem surfaces: half the answers exist only in the HR director's head. Loading the base forces them to be written down at last — a separate benefit, independent of the agent.
What should not be automated
The line runs where a decision about a person begins.
The final choice of candidate. The agent ranks, a human decides. Not only for ethical reasons: formal criteria systematically undervalue non-standard profiles, and those often turn out to be the best hires.
Rejection after an interview. An automated courteous reply at the first stage is fine. A templated rejection after someone spent two hours in conversation reads as disrespect and costs reputation.
Difficult conversations. Conflicts, dismissals, salary discussions — entirely human territory. The agent should not even attempt them.
Assessing "culture fit." This cannot be formalised, and attempting it results in a system that reproduces whatever bias was present in the historical data.
What to prepare
- role descriptions with real requirements rather than copied ones;
- selection criteria with priorities: what is mandatory, what is desirable;
- internal documents for the knowledge base;
- an onboarding checklist — if it exists only as oral tradition, it will have to be written down;
- access to the calendar and to the channel where you talk to candidates.
A realistic timeline for the first agent is two weeks, most of which goes not into development but into preparation. The day-by-day plan is in implementing AI in 2 weeks.
Frequently asked questions
Which of the four should we start with?
The screener, if you are hiring actively — that is where the manual hours are. The knowledge base, if hiring is occasional but headcount is large: then the main flow of questions comes from inside.
Should candidates be told a bot is replying?
Yes. Concealing it is pointless — people find out in their first conversation with a human and the impression lingers. One honest line in the first message settles it.
Will the agent filter out a strong candidate?
It will, if you let it make the decision. Which is why the correct setup is this: only those who clearly fail hard criteria are filtered out, everything else is ranked and passed to a human. Borderline cases should land at the top of the list, not in the rejection pile.
What about candidates' personal data?
A CV is personal data and is handled under the same rules as customer data: a defined retention period, restricted access, deletion on request. This is worth settling before launch rather than after the first question from a candidate.
What does it cost?
One process — a pilot $1,500 plus API costs, typically tens of dollars a month. The budget breakdown is in how much AI automation costs.
Where to start
Count how many hours a week your HR spends on screening and on identical employee questions. Multiply by the cost of an hour. If the number is meaningful, start with whichever of the two is larger.
Want to work it out together? We run a free process assessment and will say plainly if the volume is not yet enough to pay back.