How the leader track works
Six sessions, 45 minutes each, and you will not open a tool in any of them. This track is thinking mode: what your function's position is, what you are accountable for, and what you should be asking. Every session ends with the questions to put to your team, your vendors, or your board - because the thing that actually blocks a CHRO here is not tooling, it is not having decided. a2 is the legal map. a3 is the governance you build around vendors, audits, notice and records. a4 is fairness in practice, including the arithmetic a regulator will use. a5 is people data, monitoring and employee trust. a6 is rolling it out to a real HR function without a policy nobody reads. Meanwhile the HR track (b1-b10) builds the actual artifacts behind you: the workspace, the job descriptions, the screening simulator where a reasonable-looking shortlist fails the four-fifths rule, the interview scorecards, the decision records. You do not have to attend those. You do have to know what comes out of them.
One line before anything else, and it appears on every page of this course: this is education for HR professionals, not legal advice, and obligations vary by jurisdiction. Where this course names a date or a threshold, it was correct on 31 July 2026 and one of them has already moved once. Check the current position before you act.
AI drafts. A person decides. 7 min live
Almost everything your function produces is a document: job ads, outreach, interview guides, onboarding plans, feedback wording, policy answers, calibration notes, letters nobody enjoys writing. AI is genuinely, boringly good at all of it, and the productivity story is real. Then there is a second pile - who gets hired, who gets promoted, what someone gets paid, who leaves - and that pile is not slow because your team lacks a tool. It is deliberate, it is regulated, and a named person owns it.
LiveThe three piles, and why moving them is worth real money3 min▶
Sort your function's output honestly and the drafting pile is enormous. It is also the pile that produces the least differentiated work, gets done late, and gets done worst when your team is busiest - which is exactly when it matters. That is the case for moving it, and it does not need a business-case deck.
- Attract. Role scorecards, job ads, careers-page copy, outreach at volume, the fifteen versions of a spec that a hiring manager keeps redefining. Drafting here is not a decision about anyone. It is writing.
- Assess. Interview guides, question banks tied to actual criteria, scorecard templates, structuring the notes five raters wrote in five different formats. Note that structuring the assessment is the thing selection research has liked for decades - AI makes the good version cheap for the first time.
- Grow. Onboarding plans, learning paths, the wording of difficult feedback, policy answers that match your policy rather than the internet's. Twenty-two new joiners in a month is a drafting problem, not a judgement problem.
The job ad nobody had time to fix. At Kestrel, the Senior Support Engineer ad had been copy-pasted through four hiring rounds. It asked for a degree the role does not need, ten years of experience for a job three people currently do with four, and "excellent communication skills" with no definition. Nobody wrote it badly on purpose. It was written once, at 6pm, and then it was never anyone's job again. This is exactly the kind of artifact where a good draft, reviewed by someone who knows the role, is a genuine upgrade - and it is entirely in the top row.
LiveThe fourth box, and the sentence that gives it away3 min▶
Hire, promote, pay, exit. Four decisions about an individual, and each one is a person's income, status and next few years. The rule is not "a human is somewhere in the loop". It is that a named person made the call, could have called it differently, and wrote down why on the day.
- "The system ranked them fourth" is not a reason. In a regulated process it is closer to an admission. If your team cannot produce a reason that stands on its own without the ranking, the decision was made by the tool.
- Meaningful review means you could have said no. A manager approving a ranked list at the pace of one candidate every eleven seconds has not reviewed anything, and any serious auditor will read it that way. The EU AI Act's high-risk regime expects human oversight of exactly this kind.
- Keep the slate, not just the outcome. Who was in the pool matters as much as who came out of it. You cannot check for adverse impact on a list of the people you hired.
Why HR is the highest-stakes AI domain in the business 7 min live
Every function is being told to adopt AI, and most of them can afford to experiment. Marketing gets a bad campaign and reruns it. Finance gets a wrong number and catches it in review. 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. Three reasons yours is different.
LiveOne - it is explicitly regulated, by name2 min▶
Most functions have to reason by analogy about whether AI rules apply to them. You do not. The EU AI Act names employment directly: under Annex III point 4, AI used for recruitment or selection - including placing targeted job ads, filtering applications and evaluating candidates - and AI used for performance evaluation, task allocation, monitoring, promotion and termination is classed as high-risk. The obligations that attach cover risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness.
On timing: as of mid-2026, the Digital Omnibus agreement postponed those Annex III high-risk obligations from 2 August 2026 to 2 December 2027. Systems already in use before that date are out of scope unless they are significantly changed in design, and public authorities have until 2 August 2030. That date has already moved once - check the current position rather than trusting this page.
LiveTwo - part of it is already live, and it lands on you2 min▶
While the EU timeline moves, New York City's Local Law 144 of 2021 has been in effect since 1 January 2023 and enforced by the Department of Consumer and Worker Protection since 5 July 2023. Where an automated employment decision tool is used on a candidate, it requires an independent annual bias audit of the tool, a public summary of the audit results, and at least 10 business days' notice to candidates before the tool is used on them. Penalties start at $500 per violation and rise to $1,500 per day for continuing violations.
The part that catches people: the obligation sits on the employer, not only on the vendor. "Our supplier handles compliance" is a procurement sentence, not a defence. a2 maps this properly and a3 turns it into vendor questions.
LiveThree - the harm is individual, durable, and learned from your own history3 min▶
A campaign gets rerun. A rejection is a person's year. And the failure mode here is not a rogue model doing something exotic - it is a model doing exactly what it was asked, on the wrong objective.
The learned-from-history failure. Train a screening model on years of "successful hire" data and it will learn the pattern in that data faithfully, including who your organisation has historically hired and who it has historically passed over. It will then reproduce that pattern at speed, consistently, and with the appearance of objectivity. Nobody writes a discriminatory rule. The rule is learned from the past and applied at scale, and the tool is not broken - it is working. This is why "we did not program it to look at that" is not reassuring, and why the check has to be on outcomes, not on intentions.
The criterion nobody defined. A team scored every candidate on "communication skills". Nobody wrote down what it meant, so raters used the only evidence in front of them: the prose of the CV. The highest scores went to people who write English as a first language and had someone edit their application. The lowest went to two engineers who, in interview, explained a distributed-systems failure more clearly than anyone in the room. The criterion was not malicious. It was undefined, which meant it quietly measured background. Now imagine that criterion automated across 3,000 applications in an afternoon. That is the whole risk in one story.
What a CHRO actually owns here 6 min live
Not the model. Not the vendor's roadmap. Four things, and they do not delegate cleanly, because each one is a question you will be asked personally. Treat this as the agenda for the next five sessions rather than something to solve tonight.
LiveThe boundary and the audit3 min▶
| What you own | The question you will be asked |
|---|---|
| The boundary - what AI may never decide in your function | "Who decided that this tool could rank candidates?" Written before a pilot, not after a complaint. |
| The audit - who commissions it and how independent they are | "When was this tool last audited, by whom, and what did it find?" |
| The record - what gets kept, by whom, for how long | "Show me the slate, the criteria, the evidence and the reason for this rejection." |
| The notice - what candidates are told, and when | "Did this candidate know an automated tool would be used, before it was used?" |
Note what is not on that list: choosing the vendor, configuring the tool, running the pipeline. Those delegate. The four above are the ones where "I asked my team to handle it" is the answer that ends a career rather than a conversation.
Self-studyHow the two tracks fit together2 min read▶
The HR track builds what this track governs. b1 sets up the workspace and the three rails. b2 and b3 do role definition and sourcing. b4 is the one to know about: a live screening simulator where a perfectly reasonable-looking shortlist fails the four-fifths rule, which is the fastest way to make this concrete for anyone who thinks it is theoretical. b5 to b6 take it to structured interviews and the decision record. b7 to b9 cover onboarding, performance and internal mobility, and b10 runs a whole wave end to end with the audit trail attached.
If you only forward two pages to your team, forward b1 for the rails and b4 for the arithmetic. If you only read one more page yourself, read a2.
Sort your own function's week ★ 12 min · on paper
Not Kestrel's week. Yours. This works better with a hiring manager or a talent lead in the room, and it works best if you use last week specifically rather than a typical week, because a typical week is a story and last week is evidence.
List what your function actually produced in the last seven days. Documents, decisions, conversations. Aim for twenty lines, not five. Be unglamorous about it.
Sort each line into two piles: drafting (a document that a person then reviews) and deciding (an outcome that lands on a named individual). Anything you cannot place goes in a third pile called "argue about this", which is usually the most interesting pile.
In the drafting pile, mark the three artifacts worth moving first. The test is not "hardest", it is: high volume, low differentiation, currently done late or badly, and no personal data required to produce it.
In the deciding pile, write one line beside each: who is the named person today? If the honest answer is "the process" or "the committee" or "whoever is free", you have found something more urgent than an AI strategy.
Now the important one: name one thing you will forbid outright, starting this month, in a single unambiguous sentence. Not a principle. A prohibition, specific enough that someone could break it and be told they did.
The third pile is where the finding lives. One HR director ran this exercise expecting a clean two-way split and produced a third pile with six items in it, five of which were variations on the same thing: a tool already in the applicant tracking system was auto-filtering on years of experience before any human opened an application. Nobody in the room called it AI. Nobody had approved it as a screening decision. It had shipped as a feature in a routine vendor update eighteen months earlier. The exercise did not find an AI strategy gap. It found a live one.
Draft your function's position statement ★ 10 min · one paragraph
One paragraph. What AI is used for in this function, what it is never used for, and who decides. You will be asked for this - by a board, a works council, an employee forum, a candidate's lawyer, or a journalist - and the version you write calmly tonight will be enormously better than the version you write in the ninety minutes after someone asks.
Write the "used for" sentence from your drafting pile. Be specific enough that it is checkable: name the artifacts, not the ambition.
Write the "never used for" sentence from your deciding pile, plus tonight's prohibition. This sentence is the one people actually read.
Write the "who decides" sentence. Name roles, not the organisation. "A hiring manager" is a role. "Kestrel" is not a person and cannot explain a rejection.
Now read it back and strike every sentence you cannot honestly write today. Whatever you struck is your real agenda for a3 to a6.
Take it to one person who will disagree with you - a works council rep, a sceptical engineer, your DPO - before it goes anywhere near a slide.
Before session 2 ◐ 30-40 min total
- Finish the position statement. One paragraph, no adjectives about your own values, every sentence checkable.
- Ask your talent lead for a list of every tool that touches hiring or promotion, including the ones nobody calls AI - ATS filters, ranking features, scheduling tools that deprioritise, assessment platforms. Ask when each one was last updated. You want the list, not a reassurance.
- Ask one hiring manager to explain, from memory, why the last candidate they rejected was rejected. Note whether the reason exists in writing anywhere. Do not make it a test - make it a data point about your record layer.
- Find out who owns AI decisions in HR where you work: DPO, legal, the CIO, or nobody. If the answer is nobody, that is a2 and a3's whole reason for existing.
- Read a2 before it is taught. It is the map, and it changes.
Official sources covered
Taught from public regulation, selection research and vendor documentation. Certification content (SHRM, CIPD), the normative text of the AI Act, 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 hiring manager tells you the AI "just does a first pass" and a human always reviews the shortlist. What do you need to know before that reassures you?
A review that could not have gone the other way is a rubber stamp, and both the AI Act's human-oversight expectation and any serious auditor will read it that way. A first pass that removes people is a decision about those people.
2 · Why is "the model was trained on our own successful hires" a warning rather than a reassurance?
The tool is not broken in this failure - it is working, on the wrong objective. That is why the check has to be on outcomes rather than on intentions, and why "we did not program it to look at that" reassures nobody.
3 · Which of these four does a CHRO own personally, rather than delegating to procurement or IT?
Vendor choice and configuration delegate cleanly. What AI may never decide, who commissions the audit, what is kept as a record, and what candidates are told before a tool is used on them are the four you will be asked about by name.