AI for HR teams: a practical guide

This guide is for HR generalists, people ops leads, and recruiting leads at companies of roughly 20 to 500 people. Most of their lost time is not judgment but writing and chasing: job descriptions from six bullet points, late scorecards, candidates waiting a week for a reply, and the same leave question answered for the fortieth time.

A common approach is one named bot per job plus a manager bot to coordinate them. This guide uses the documented type.com model instead: one Space per department, with shared connections, instructions, and skills, and a channel per workflow. HR adds one twist: the Space is also the privacy boundary, so you will set up three. It also covers what most AI recruiting guides skip: the hiring laws that apply once software evaluates people.

The examples follow Copperleaf, a fictional 182-person software company in Denver with offices in New York City and Dublin. Its people team is Priya Shah (Head of People), Marcus Bell (Recruiting Lead), Leo Tran (Recruiting Coordinator), and Dana Okafor (People Ops Generalist). It has nine open reqs and uses Lever. Times are Mountain Time, and the time ranges are estimates; measure your own baseline first.

Discrimination law applies whether a person or software makes the call. U.S. law forbids discriminating against applicants and employees because of race, color, religion, sex (including transgender status, sexual orientation, and pregnancy), national origin, age (40 or older), disability, or genetic information (EEOC), including through neutral-looking practices that disproportionately exclude a group (Title VII, 42 U.S.C. 2000e-2(k)). Federal guidelines generally treat a group's selection rate below four-fifths of the highest group's rate as evidence of adverse impact (29 CFR 1607.4(D)). A model can learn proxies from a resume: graduation years for age, gaps for caregiving or disability, names and zip codes for race. That is why AI screening is risky even when nobody intends harm.

Illinois, since January 1, 2026, bars AI that has the effect of discriminating, bars zip codes as a proxy, and requires notice when AI is used in recruitment, hiring, promotion, discipline, or discharge (Public Act 103-0804). New York City's Local Law 144 covers tools that produce a score, classification, or recommendation and substantially assist hiring or promotion decisions, which the city's rules define as output that is the only factor, outweighs every other criterion, or overrules human judgment. Using one requires an independent bias audit within the past year, a public summary of results, and notice to New York City residents 10 business days before use, with a way to request an accommodation.

The EU AI Act classes AI intended for recruitment or selection (targeted job ads, filtering applications, evaluating candidates) and for promotion, termination, task allocation, or performance evaluation as high-risk (Annex III, point 4). The date those obligations apply has been revised more than once, so confirm the current one before you rely on it (Article 113), including trained human oversight and telling workers before use (Article 26). Inferring workers' emotions has been banned since February 2, 2025 (Article 5). The GDPR already limits significant decisions based solely on automated processing (Article 22), and California's privacy law covers resident applicants and employees (CPPA).

This guide stays on the drafting side of that line. AI writes materials people approve, summarizes context people already have, organizes an interviewer's own evidence, flags process delays, and answers handbook questions with sources. It never scores, ranks, filters, or rejects a person. Whether a use counts as an automated decision tool or a high-risk system is a legal question, so review the setup with employment counsel. This guide is not legal advice.

type.com's HR & Ops template is built for recruiting and people operations and creates a private Space by default. It proposes skills, automations, and connections, but nothing is added until you approve it (templates and models).

Why three Spaces? Channels share their Space's access and are not a permission boundary (permissions and access). Hiring managers need reqs and interview prep, not reviews or exit data, and employees need only the handbook. So the People Space is HR-only, the Hiring Space adds recruiters and managers with open reqs, and the Ask People Space can read only the handbook folder. For one person's situation, such as an accommodation or an investigation, use a private thread, visible only to its participants. Private-thread and Sidekick context is not moved into workspace memory.

Organization connections use shared company access (ATS, HRIS, handbook). Personal connections use one person's account, such as a recruiter's calendar, and only they can use them (Space connections). Google Drive access is granted file by file or folder by folder. The Read-only setting tells the AI not to make changes but does not stop the service from accepting them, so enforce limits with each tool's own roles and API scopes.

When and who: on request in the Hiring Space's req-kickoff channel, owned by the recruiter with the hiring manager in the thread. Inputs: intake notes or a call transcript, the leveling guide, the pay band, and the last approved posting in the job family.

The scorecard matters more than the posting. In a structured interview, every candidate is assessed on the same competencies with planned questions and an anchored scale. It only works if the scorecard is agreed before the first screen, so nobody invents criteria after meeting a candidate they like.

Example: Aisha Karimi, the payments engineering manager, pasted six bullet points for ENG-214, Senior Backend Engineer (Payments). The draft returned five competencies, two requirements to question ('CS degree required' and '8+ years of Go'), a $172,000 to $205,000 range from the L4 band, and three open questions. Aisha dropped the degree line, rewrote the Go line, and approved the scorecard in 20 minutes.

Review rule: the hiring manager approves the scorecard, the recruiter approves the posting, and AI never posts. Save the instructions as a skill in the Hiring Space Library so every req starts the same way; teammates can suggest edits for the owner to review (how skills work).

Prep brief: weekdays at 7:30 AM MT in the Hiring Space's interview-prep channel. Leo creates it because it reads his scheduling calendar, and scheduled automations run with their creator's connections. Inputs: the calendar, Lever, and the scorecard. The brief never shares earlier ratings, so each interviewer judges independently. Most panelists aren't in the Hiring Space, so Leo pastes each brief into their invite.

Feedback draft: after each interview, the interviewer runs Part B in their Sidekick, which is private to them, with typed notes or their own Granola, Fathom, or Fireflies notes. Some U.S. states require every party's consent to record a call, so check first. The draft sorts evidence by competency and flags comments that aren't job related, but leaves the rating and recommendation blank. The interviewer chooses them and submits in the ATS, so the AI never produces a candidate score.

Example: Tom Becker interviewed Jordan Ellis for ENG-214 on Wednesday, October 7, at 2:00 PM MT. His notes included 'great culture fit, reminds me of our team', 'young but sharp', and three solid observations about idempotency keys, retry backoff, and ledger reconciliation. The draft sorted the observations, marked Collaboration 'No evidence in notes', and flagged both comments. Tom deleted them, chose 3 of 4, and submitted at 3:10 PM.

When and who: weekdays at 9:00 AM MT in the Hiring Space's pipeline channel. Marcus creates it on the organization's Lever connection, and each recruiter works their own section. Inputs: stages, stage dates, interview dates, feedback status, and last contact from synced email.

A strong candidate who hears nothing for a week may accept another offer. The sweep looks only at process facts like dates and missing feedback, never at candidate quality, so it ranks no one. It drafts a rejection note only after a person has recorded the decision in Lever.

Example: on Wednesday, October 7, the sweep found 14 items across 143 active candidates: 6 stuck, 4 overdue feedback forms, 3 candidates waiting over two business days, and 1 offer expiring Friday. Marcus sent the 3 follow-ups by 9:40 AM, after editing one, and nudged the interviewers in Slack himself.

When and who: weekdays at 8:00 AM MT in the People Space's onboarding channel, owned by Dana. Inputs: the HRIS report, onboarding templates, and the role's approved posting and scorecard. Building the 30-60-90 plan from the scorecard means the hiring decision, the plan, and the first review use the same language.

Example: Maya Lin starts Monday, October 19, as a Customer Success Manager in New York. On October 7 the check posted her draft: a day-one schedule, 23 tasks across IT, her manager, payroll, and People, and a plan built on the role's four competencies. Her manager rewrote the 60-day goal, and Dana assigned the tasks.

Review rule: the manager owns the plan, and Dana sends the welcome email herself. Form I-9 documents are reviewed by a person and never uploaded to a thread; the checklist tracks only the deadline, three business days after the start date.

When and who: always on, in a public #ask-people Slack channel mapped to the Ask People Space with Mentions only (Slack setup). Dana owns it. Its only input is the Handbook folder. Because connections are assigned per Space, no question, however worded, can pull an answer from the HRIS, the ATS, or anyone's review.

Example: on Wednesday, October 7, at 10:14 AM, Ben Ortiz asked whether he could work from Portugal for three weeks in December. The answer quoted section 4.5: up to 20 working days a year abroad, approved by the manager and People at least 30 days ahead through the Work Abroad form. It noted three weeks is 15 working days, said it can't approve requests, and named Dana as the reviewer.

Review rule: Dana spot-checks 10 answers a week. If an answer is wrong because the handbook is unclear, fix the handbook, not the instructions. A harassment or discrimination report in the channel gets the private reporting route from the AI and a same-day follow-up from a person.

When and who: during each cycle (Copperleaf's runs October 19 to November 13). Managers run the skill in their Sidekick; HR uses a private thread. Inputs are only what the manager pastes: 1:1 notes, goals, the self-review, and peer feedback. Nothing comes from the HRIS, and pay is never included.

The checks are the valuable part. Recency bias shows up as evidence bunched in the last six weeks. Labels like 'abrasive' say nothing about behavior. Mentioning protected leave, health, family, or age as a performance factor is a legal risk, so the draft flags it for removal.

Example: from 14 pages of 1:1 notes, Aisha's draft showed that 7 of 9 examples came from September and flagged 'can be abrasive in code review' as a label without an example. She added two spring examples and rewrote the comment around a specific pull request.

Review rule: the manager writes the rating and submits it; calibration and pay stay with people. Never ask AI to rate, rank, or calibrate employees.

When and who: the first Monday of the month at 8:00 AM MT in the People Space's people-reports channel. Priya creates it, so it runs with her access. Most reporting arguments are about definitions, so the instructions fix them once.

Example: September's report showed headcount rising from 176 to 182 (9 hires, 3 exits), turnover of 1.7%, a median 41 days to fill and 27 days to hire, and 9 of 11 offers accepted. It flagged a hire accepted in Lever with no HRIS record, which Dana fixed before Priya shared the report.

Keep EEO self-identification data out of AI threads and run adverse-impact checks in your ATS's EEO reports. For more on recurring summaries, see how to automate a weekly team report.

Each automation has a fixed time, channel, and owner. If an output sits unread for a week, pause the automation rather than let the channel become noise.

Know whose access each run uses. In type.com, AI acts with the effective access of the person who started the work, and scheduled automations run with their creator's identity, connections, and AI subscription (test and troubleshoot automations). New automations start paused: Save, Test (which runs privately as you), then Enable. Three failed runs in a row pause a schedule. When someone leaves the HR team, recreate their automations under a current owner before removing their access.

Data handling: never put Social Security numbers, medical or accommodation records, background checks, immigration documents, or investigation notes into a thread, skill, or instruction, and keep credentials in connections. When someone asks you to delete their data, delete the related threads too; type.com may later clean up related memory. For more on access design, read permissions and approvals in a shared AI workspace and AI agent approval workflows.

Start with the HR team only. When a run is wrong, correct it in the thread, then fix the instructions or the source document.

Take the baseline the week before you start from last quarter's ATS reports and a week of time logs. Time to hire is a quarterly number, so don't expect it to move in 30 days.

Other guides in this series cover AI for small business, AI for finance and operations, and AI for customer success. For how context builds up in a Space, and what stays out of it, read shared AI memory for teams.