An AI agent can take much of the drafting out of an HR week: job adverts, interview guides, plain-English policy answers, case chronologies and the narrative for the monthly people report. It prepares documents for you to review. Decisions about people stay with a named, accountable person.
HR work is a lot of careful writing under time pressure, often urgent and repetitive at once, which is hard on a small team. You paste the source material (the old advert, the policy clause, your interview notes, the month's numbers) into a chat and get a structured first draft back. Nothing on this page assumes access to your inbox, calendar or HR system.
Where's the line between drafting and deciding?
The rule that runs through this page: anything that affects someone's job, pay or livelihood is decided by a person who can explain the decision. The agent can organise evidence, draft the letter and show you where the gaps are. It doesn't score, rank, select or conclude.
- Hiring. It can lay out what each candidate's interview notes show against your criteria. A person decides who goes forward, and can say why. Rules on automated decision-making about people vary by country, so check yours before using any tool near a hiring decision.
- Discipline and grievances. Drafting the invitation letter or the chronology is fine. The finding, the meeting and the outcome belong to you and whoever advises you on employment matters.
- Redundancy. No agent should propose who's at risk, score selection criteria or draft an outcome. Consultation and selection rules differ between countries and mistakes are expensive.
- Contracts, offers, references, pay and promotion. These commit the organisation. An agent can produce a first draft for a qualified reviewer, and its part ends there. It shouldn't compare individuals or recommend who gets what.
- The conversation itself. Someone told their role is at risk wants a person in the room they can question.
Employment and data-protection law vary by country and sometimes by region. Before changing a process, contract or policy because of anything here, check the rules where you operate or ask your adviser.
Recruitment: adverts, interview guides and debriefs
The job advert. Paste last year's advert and the person specification, and ask for two versions (one formal, one plain) with the empty phrases cut, the requirements kept and nothing added about the team that you didn't supply. You'll usually also get a note where the requirements contradict each other, which is worth having before the advert goes live. Pick sentences from both versions.
The interview guide. Structured interviews give you notes you can compare, and many teams still improvise.
These are the five criteria for the role. Build a structured interview guide: two questions per criterion, what a strong answer contains, what a weak one sounds like. No brain teasers.
The evidence grid. After interviews, paste the panel's notes and the criteria and ask for a grid of candidate, criterion and evidence quoted from the notes, with blanks where there's none. Tell it not to rank anyone. A comment like "good culture fit" gets flagged as unsupported, and the debrief can then focus on the criteria nobody has evidence for yet.
Candidate messages. Invitations, rejections and "we're holding your application" messages are read closely and remembered. Ask for drafts that are warm, short and specific, with placeholders for names and dates and nothing that reads as a commitment. You review and send them.
Policies people can read
Policies are written for compliance and read by people in a hurry. Paste one and ask for a plain-English version that keeps every obligation, softens nothing that's a requirement, and marks each sentence where the original is ambiguous. You get a readable draft and a list of ambiguities for the policy owner, which is often the more useful output.
The same handful of policy questions arrives every week in different words. When one comes in:
An employee has asked: [question]. The relevant policy section is below. Explain what it says in plain English, say where it's silent, and list what I should check before replying.
You edit the explanation into your reply and note the gap for the next policy review. Over a month, those notes tell you which clause most needs rewriting.
Paperwork for casework
Investigations depend on dates, and memory is unreliable. Paste your notes, the relevant emails as text and the meeting record, and ask for a dated chronology: date, what happened, who was present, source. Ask for any inferred date to be marked as inferred. When accounts conflict, the agent should show the conflict, and you should tell it so.
Build a chronology from the material below. Columns: date, event, people present, source. Mark any date you're inferring. Where two sources disagree, show both.
For a formal meeting, paste your template and the facts of the case and ask for the invitation letter: what the meeting is about, when, who'll attend, and whether the employee can bring a companion under your process. Ask it to flag anything in the template that doesn't fit these facts. The process owner reviews it, plus an adviser where the case needs one.
A performance improvement plan may be read closely by someone outside the organisation one day. Give the agent the concerns, the evidence and the role expectations, and ask for specific expectations, how they'll be measured, review dates and the support offered. Tell it to invent no examples. The manager owns the final wording.
People numbers and survey comments
The monthly report. Numbers without a narrative get ignored or misread. Paste headcount, joiners, leavers and turnover with last month's figures alongside, and ask for 200 words on what changed, the two main drivers and one thing to watch. Add "don't explain the numbers away" to the prompt. Correct the drivers if they're wrong, and they will sometimes be wrong, because the agent only knows what you pasted.
Survey free text. The useful information sits in the comments nobody has time to read. Paste 180 anonymised comments and ask for themes, a count per theme and three representative quotes each, with a warning against overstating a small sample.
Exit interviews. One exit interview is a story; six start to show a pattern. Paste anonymised notes and ask which themes appear in more than two, with the supporting lines quoted and a note on what a sample that size can't tell you.
The manager briefing. When an announcement goes out, managers get asked the questions. Turn the announcement into a one-page brief: the key message, the three questions they'll face, suggested answers and what to do when they don't know.
Can you put employee data into an AI agent?
Only with agreed rules, and only the data the task needs. HR holds the most sensitive information in the organisation, so ordinary caution falls short.
Health, disability, ethnicity, religion, union membership and criminal records are treated more strictly under GDPR and most other data-protection laws. Keep them out of the chat. Case files are identifying even without names: a chronology with dates and a department often points to one person. Use role labels ("the employee", "Manager B") and leave the source documents where they belong.
Using a tool on employee data is itself a use of that data. Ask whoever owns data protection whether your privacy notice covers it, whether a formal risk assessment is needed first, and how long text may sit in chat history. What happens to your data lists what to ask a provider. If you have a works council, union or staff forum, you may need to consult them before introducing a tool that handles employee data; the employer check covers what to settle first.
Human in the loop has a concrete meaning here: a named person reviews every output that touches an employee, and the file shows the review happened.
A rhythm for the HR week
Scheduled check-ins mean the agent prompts you, which is how cases that have stalled get noticed. Scheduled check-ins explains how they work.
Every Monday at 8:30am, ask what's on this week: interviews, case meetings, deadlines. Turn my reply into a one-page plan with the three things that mustn't slip.
Every Thursday at 4pm, ask which open cases moved this week and which didn't. For anything stalled, tell me the next step and who owns it.
On the last working day of each month, ask for the people numbers, draft the narrative and remind me which figures I still owe.
On the first Monday of each quarter, ask which policies I've explained most often since the last check, and suggest the clause most worth rewriting.
If you only set one, make it the Thursday case check.
House rules worth saving
When a format has worked a couple of times, ask the agent to save it as a skill so every draft starts from it. Agent skills explained has the detail.
- House style. British spelling, "we" rather than "the company", no promises about timescales, always the name of the person to contact.
- Interview guide format. Criteria, two questions each, strong and weak answers, a notes column. A new role then takes minutes.
- Chronology format. The columns, plus the rule that inferred dates are marked. These documents get read by people outside the team, so consistency matters.
- The plain-English pass. Keep obligations, mark ambiguity, change nothing substantive without flagging it. This stops a rewrite turning into a policy change.
How we start you off, and your first week
We set Able° up for you and connect it to Telegram or LINE with you. Then we go through what it should know first: your house style, the words you avoid, how you address employees and what you never promise. That context sits in its memory, so you don't repeat it in every request.
Your first week:
- Start with a document that isn't about a person. Take a policy nobody understands and ask for the plain-English version. It's low risk and shows you how the agent handles text with obligations in it.
- Try one recruitment task, with review. An advert or an interview guide. Leave screening and scoring alone.
- Set the Thursday case check.
- Write your team's rule. One paragraph on what may be pasted, what may not and who reviews what. Put it where the whole team will see it.
Common questions
Can an AI agent screen CVs?
It can summarise what a CV contains and set it against criteria you wrote, which saves reading time. It shouldn't score, rank or reject anyone. A person owns the shortlist and should be able to explain every name on it.
Can it write redundancy or dismissal letters?
It can draft the administrative parts, such as dates, the meeting arrangements and the structure of the letter. The decision, the meeting and the final wording belong to a person, checked against the rules where you operate or by your adviser.
Can it help with performance reviews?
Yes, on the writing side: it can help a manager turn vague concerns into specific, evidenced examples and structure the review. It shouldn't rate people, compare them or suggest outcomes, since ratings feed into pay and employment decisions that need a consistent human process.
Is it worth it for a one-person HR function?
Often yes, because a solo HR lead does all the writing personally and the adverts, policy answers and reports repeat. It's less useful if your caseload is two complex cases a year: those need careful notes and an adviser more than faster drafting.