How this track works
Ten hands-on sessions across the whole people lifecycle, all on one running company: Kestrel, a fictional 400-person services firm scaling to 600. There is a 120-role hiring wave, a performance cycle, and an engineering attrition problem, and you will work all three. Tonight is setup. b2 defines roles and writes job descriptions that widen the pool, b3 does sourcing at volume, b4 is the live screening simulator where a perfectly reasonable-looking shortlist fails the four-fifths rule, b5-b6 take it through structured interviews to offer, b7-b9 cover onboarding, performance and internal mobility, and b10 runs a full wave end to end with the audit trail. The leader track (a1-a6) runs alongside for CHROs and HR directors - a2 there is the legal map, and it is worth reading even if you never attend.
Where AI helps, and the line it does not cross 6 min live
Most of HR's week is drafting and collating: job descriptions, outreach, interview guides, onboarding plans, feedback wording, policy answers. AI is genuinely good at all of it. Then there is a second pile - who gets hired, who gets promoted, who gets managed out - and that pile is not slow because we lack a tool. It is deliberate, it is regulated, and a person owns it.
LiveWhy HR is the highest-stakes AI domain in the business3 min▶
Marketing gets a bad campaign. Finance gets a wrong number in a deck. HR gets a person who did not get the job, a protected characteristic quietly doing the work of a criterion, and a regulator who is entitled to ask how the tool was audited.
- It is explicitly regulated. The EU AI Act classes AI used for recruitment, filtering applications, evaluating candidates, performance evaluation, promotion and termination as high-risk. The obligations for those systems now apply from 2 December 2027 after the Digital Omnibus postponement - later than first legislated, and still coming.
- Some of it is regulated already. New York City's Local Law 144 has been enforceable since 2023: if an automated employment decision tool is used on a candidate in NYC, the employer needs an independent annual bias audit, a public summary of it, and at least 10 business days' notice to the candidate. That obligation sits on the employer, not only on the vendor.
- The harm is individual and durable. A campaign gets rerun next quarter. A rejection is a person's year.
The screen that worked exactly as designed. A hiring team fed a model years of "successful hire" data and asked it to rank applicants. It learned the pattern in the data faithfully - including who had historically been hired - and reproduced it. Nobody wrote a discriminatory rule; the rule was learned from the past and applied at speed. The tool was not broken. It was doing its job on the wrong objective, which is the failure mode you will meet in b4.
Self-studyThe dates, and why you must check them yourself2 min read▶
As of mid-2026: the EU AI Act's Annex III high-risk obligations, which cover employment, were moved from 2 August 2026 to 2 December 2027. Systems already in use before that date fall outside the obligations unless they are significantly changed in design; public authorities have longer. NYC LL144 is live now, with penalties starting at $500 per violation and rising to $1,500 per day for continuing ones.
One of those dates has already moved once, which is the actual lesson: check the current position before you act, and never let a course page be your compliance source. The leader session a2 is the full legal map, a3 is the governance you build around it, and none of it is legal advice.
The HR workspace: what goes in, what never does 5 min live
An empty chat window can only give you generic HR advice, which you do not need and which is occasionally wrong for your jurisdiction. A workspace holding Kestrel's role scorecards, comp bands, interview guides and policies can draft the actual artifact. But an HR workspace has a boundary no other function has: most of your most useful-looking material is other people's personal data.
LiveThe six documents worth loading3 min▶
None of these are personal data. All of them are what makes a draft specific to your company instead of specific to the internet.
| Document | What it unlocks |
|---|---|
| Role scorecards - what good looks like per role | Rating against the job instead of against CV polish |
| Comp bands and levelling framework | Offers and ads that do not contradict your own structure |
| Interview guides and scorecards you already use | New guides in your house format, comparable across raters |
| Policies: hiring, performance, leave, conduct | Answers that match your policy, not a generic one |
| Anonymised past decision records and rationales | Consistency - and a sense of what your bar actually is |
| A glossary of your levels, teams and acronyms | Drafts that speak Kestrel instead of speaking HR-in-general |
Self-studyApproved tools, and the question to ask first2 min read▶
- Ask what is approved before you build anything. HR data usually has a tighter tool list than the rest of the business. If nobody can tell you, that is a finding worth raising - the leader session a3 is exactly this conversation.
- Ask whether it is an AEDT. If a tool substantially assists or replaces a hiring or promotion decision, obligations attach in some jurisdictions - independent bias audit, public summary, candidate notice. Drafting a job ad is not that. Ranking a slate is.
- Ask who else can see the workspace. HR content is confidential by default. A shared workspace with the wrong members is a data incident waiting for a slow Tuesday.
- Availability and features change fast. If a button named here has moved, the exercise still works - the artifact is the point.
Three rails, each one scar tissue 4 min live
Techniques change every quarter. These three do not, because each is a specific way AI-assisted HR goes wrong.
LiveRail 1 - a person decides, and the record shows which person2 min▶
Every decision about an individual - shortlist, reject, offer, rating, promotion, exit - is made by a named human who can say why, in writing, from evidence. "The system ranked them fourth" is not a reason, and in a regulated process it is close to an admission.
- Write the reason at the time. Reconstructing a rationale six months later, after a complaint, is exactly when it looks reconstructed.
- Keep the slate, not just the outcome. Who was in the pool matters as much as who came out of it - that is what an impact check needs.
- Meaningful review means you could have said no. A human who rubber-stamps a ranked list has not reviewed anything, and both the AI Act's human-oversight expectations and any serious auditor will read it that way.
LiveRail 2 - rate the job, not the polish1 min▶
The fastest way to build a biased screen is to rate what a CV looks like instead of what the job needs. Polish, prestige, fluent phrasing, an unbroken timeline - all of them travel with background rather than ability, and all of them are exactly what a language model finds easiest to reward. Structured criteria drawn from the actual job, with evidence attached to each rating, is the fix. It is also, independently, the thing that predicts performance better. In b4 you will watch both numbers move together.
The "communication skills" criterion. A team scored every candidate on communication. Nobody defined it, so raters used the CV's prose. The highest scores went to people who write English as a first language and had someone edit their CV; the lowest went to two engineers who, in the interview, explained a distributed-systems failure more clearly than anyone in the room. The criterion was not wrong. It was undefined, which meant it quietly measured background.
LiveRail 3 - people data is not prompt fodder1 min▶
No CVs, no candidate contact details, no health or accommodation information, no grievance or disciplinary material, no named pay data - not into an unapproved tool, and not into an approved one without a lawful basis and the minimum necessary data. Candidates and employees did not hand you their information so it could become context for a chat window.
Build the Kestrel workspace ★ 12 min · everyone builds
Build it for Kestrel now, rebuild it for your own function this week. Fifteen minutes once, and every session after this assumes it exists.
Create a workspace named for the function and the wave: Kestrel - hiring wave and people cycle. Check who else can see it before you put anything in it.
Load the six documents. For Kestrel tonight: three role scorecards, the levelling and comp band sheet, the standard interview guide, the hiring policy, and a glossary of levels and teams. No CVs. None.
Write the workspace instructions from the skeleton below, including the never-list.
Test it with a question that is impossible without the pack: "For the Senior Support Engineer role, what are the five must-have criteria, and which of them is most often faked on a CV?" A loaded workspace answers from your scorecard. An empty one gives you a listicle.
If the answer was generic, do not rewrite the question. Add the missing document.
Write the house rules and the candidate notice ★ 10 min · build your own
Two artifacts. The house rules are internal and you will carry them from job to job. The notice is external, it is what a candidate reads, and in some jurisdictions a version of it is required - with lead time.
Start the house rules from the three rails, then add your own hard "never" lines - the ones that come from something that has actually gone wrong where you work.
Add your artifact standards: what a criterion must look like, that every rating carries evidence, that every rejection has a recorded reason.
Now draft the candidate-facing notice: what AI is used for in your process, what it is not used for, that a person decides, and how to ask for a human review or an adjustment.
Check the timing question with whoever owns compliance. Where notice is required, it typically has to arrive before the tool is used - NYC's rule is at least 10 business days ahead - which means this is a process design question, not a footer.
The notice that improved the process. A team wrote their candidate AI notice honestly - AI drafts the job ad and the interview guide, a person reads every application, a person decides - and then noticed they could not honestly write the third line, because a tool was auto-filtering on years of experience before any human saw a CV. Writing the notice is what surfaced it. Nothing sharpens a process like having to describe it to the people it is done to.
Try it yourself - this week ◐ 30-45 min total
- Build the workspace for your own function. Six documents, none of them containing a named person's data.
- Finish the house rules. Add two "never" lines that come from your own organisation's history.
- Write your candidate AI notice, then check every sentence in it against what your process actually does today. Any sentence you cannot honestly write is this week's real finding.
- List every tool already touching your hiring process, including the ones nobody calls AI. Mark which of them substantially assist a decision. Bring that list to b4 - it is the list that matters.
- Ask who owns AI decisions in HR where you work: DPO, legal, or nobody. If it is nobody, leader session a3 is the one to forward.
Official sources covered
Taught from public regulation, selection research, and vendor documentation. Certification (SHRM, CIPD) and any actual legal position stay with their owners. This course is education for HR professionals, not legal advice, and obligations vary by jurisdiction. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · A manager asks you to paste 60 CVs in and get a ranked shortlist back by Friday. What is the first problem?
Rails 1 and 3 at once. It is an automated employment decision with no defined criteria, no evidence, no audit and no notice - and the personal data should not be there in the first place. Checking the top five afterwards is a rubber stamp, not review.
2 · Your draft screening criteria include "strong communication skills". What is wrong with it?
Rail 2. An undefined criterion is not neutral - raters fill the gap with what they can see, which is writing style. Define what communication means for this job and how it will be evidenced, or drop it.
3 · Which statement about the legal position is accurate as taught here?
The live obligation lands on the employer, not just the vendor, and it is live today in some places. The EU timeline moved once already - which is why you check the current position instead of trusting any page, including this one.