Where we are, and what the finale is for
a1 was what AI changes about the people function. a2 was the legal map. a3 was governance - vendors, audits, notice, records. a4 was fairness in practice, and a5 drew the line around employee data and monitoring. Everything so far has been about judgment: what to look for, what to refuse, what to insist on. Tonight is about motion, because a leader who understands all five sessions and starts nothing has changed nothing, and a leader who starts in the wrong place has changed something they will spend a year undoing. So: a pilot small enough to be honest about, a quality bar specific enough to fail, metrics that do not flatter you, and a rollout story your own team believes. Then the track hands over to the HR track, where the artifacts actually get built.
One team, one artifact, six weeks, and not the funnel 6 min live
The instinct is to pilot where the pain is loudest, which in a 120-role hiring wave means screening. That is precisely the wrong start. Screening is the highest-regulated, highest-harm, hardest-to-reverse part of your process, and it is where you have the least organisational muscle for checking AI output. You earn your way there. You start where a bad draft costs an afternoon.
LiveWhy the screening decision is the wrong first pilot3 min▶
Every argument for starting at screening is an argument about volume: it is where the hours are. Every argument against is an argument about consequence, and consequence wins in a regulated function.
- It is the part with obligations attached. A tool that substantially assists a hiring decision can be an automated employment decision tool, with audit, public summary and candidate notice implications, and it sits squarely inside the AI Act's Annex III employment scope. You do not want your first experiment to be the one with a notice period.
- It is the part you cannot un-do. A weak job ad gets rewritten. A candidate rejected by a badly checked screen is not recoverable, and neither is the record of it.
- Your team has no reps yet. Checking AI output well is a learned skill. Week 1 of that skill should not be applied to somebody's career. Job ads and interview guides give people twenty repetitions of "is this actually right" at a cost of nothing.
- The small stuff is where the hours actually are anyway. Count the drafting time across a 120-role wave - ads, guides, outreach templates, onboarding plans, offer wording. It is larger than the screening time and it carries none of the exposure.
The pilot that skipped the queue. A people team started with resume screening because it was the loudest complaint. Six weeks in, they had a working setup, a genuinely faster process, and a compliance review they had not budgeted for, triggered by one candidate asking a reasonable question about how their application was assessed. The tooling survived. The programme did not - it was paused for two quarters, and when it restarted, it restarted with job ads. They got to the same place, one year later, having spent the political capital twice.
LiveWhat "one team, one artifact" buys you2 min▶
Narrow scope is not timidity. It is the only way to get an answer you can trust out of six weeks.
| Choice | Why it is not negotiable |
|---|---|
| One team | You can talk to all of them. Feedback stays real instead of becoming a survey. |
| One artifact type | A quality bar for job ads is writeable. A quality bar for "HR work" is not. |
| Six weeks | Long enough for the novelty to wear off, short enough that nobody reorganises around it. |
| A written baseline in week 2 | Without it, week 6 is a vibe. With it, week 6 is a comparison. |
| An explicit not-touching list | The list is what makes the pilot safe to approve, and it is what a5's tier map feeds. |
| A named owner | Pilots with a committee sponsor and no owner run for eleven months and conclude nothing. |
A quality bar specific enough to fail 5 min live
A quality bar that nothing ever fails is decoration. The acceptance test here is the same discipline the HR track teaches its practitioners, which is the point - the leader's bar and the practitioner's checklist have to be the same object, or you will be inspecting something nobody was building towards.
LiveThe four-line acceptance test2 min▶
Every sampled artifact gets checked against four things. Any one of them missing is a fail, not a note.
- Criteria are evidenceable. Each requirement can be demonstrated by something a candidate does, shows or answers. "Strong communication skills" fails. "Can explain a technical failure to a non-technical stakeholder, assessed in the panel exercise" passes.
- Ratings cite evidence. Not an adjective about a person, but the work sample, the answer given, the artifact seen. This is the line that makes the whole record defensible later, and it is the one most often skipped under time pressure.
- Decisions name a person. A human who could have said no, who did say why, and who wrote it down the same day. "The system ranked them fourth" is not a reason.
- The impact check ran. Not that it passed - that it ran, was recorded, and was looked at by someone who could act on it.
The bar that finally bit. One team ran a review where every artifact passed for three straight weeks, which felt excellent until someone asked what the bar would have rejected. The answer was nothing - the questions were all "did you consider" and everything had been considered. They rewrote the four lines as pass or fail with no middle, sampled again, and failed four of twelve on "ratings cite evidence". The work had not got worse. The bar had finally become capable of noticing.
LiveSample a fixed share, and define the failure path2 min▶
Reviewing everything is the instinct, and it collapses in week 3. It also produces the worst possible outcome: a review process so heavy that people quietly route around it, leaving you with no visibility at all and a false belief that you have some.
- Pick a share and hold it. One in five during the pilot is a reasonable starting point for low-blast-radius artifacts. Publish the number. Sample randomly, not by whoever is nearest.
- Raise the share for higher-consequence artifacts, do not lower it for volume. If volume makes your share unaffordable, the honest response is a smaller pilot, not a thinner check.
- Define what happens on a fail, in advance. Who is told, in what tone, within how long. A failed sample is a signal about the bar, the training or the tool - it is not a disciplinary matter, and if it feels like one, people will stop surfacing problems within a fortnight.
- Define the escalation. Two fails on the same line in one week is a systemic finding and goes to the pilot owner, not into a spreadsheet.
- Sample the un-run checks too. The most useful thing a sample can catch is not a bad artifact but a missing one - the impact check that was skipped because the slate looked fine.
What to measure, and the number that lies 5 min live
Every AI programme eventually gets asked for a number. The ones that are easy to produce are the ones that mean least, and one of them - the most quoted metric in recruiting - improves for two completely opposite reasons and never tells you which one you got.
LiveWhy the good list looks harder than the bad list3 min▶
Because it is. Every honest metric requires you to have defined something first - what an evidenceable criterion is, what counts as a recorded reason, what the slate was. The vanity metrics require nothing except a tool that logs.
- Hours moved off drafting is the honest version of productivity. It says where the time went, not that people were busy.
- Time to first response for candidates is the one improvement people outside the company actually feel. It is also the easiest early win, and it is worth putting on a board slide precisely because it is about the candidate rather than about you.
- The four-fifths ratio across the wave and not just per slate. Small slates make the ratio jump around - one seat swings it - so the aggregate is the number that means something.
- Share of criteria that are evidenceable and share of decisions with a same-day recorded reason are the two that will look bad first and improve most. They are also the two you will be grateful for if anyone ever asks.
- "AI adoption %" measures a tool's success, not yours. It can rise while quality falls, and it never falls when it should. If a board asks for it, give it alongside the sampled pass rate, never on its own.
What the b4 simulator shows about measuring. In the HR track's screening simulator, a slate that scores 61 on quality with a four-fifths ratio of 0.09 climbs all the way to 100 quality and a ratio of 1.00 through five fixes - better criteria, evidence, structure. Then there is a sixth lever, the impact check, and it moves neither number. It only reveals and blocks. That is the honest shape of this whole conversation: measuring is not fixing. The five levers fixed things and the numbers reflected it; the check that catches the problem you did not anticipate improves nothing on the dashboard and is the one you cannot drop.
Self-studyThe reporting cadence, and one rule about targets2 min read▶
Weekly to the pilot team, so it is a working instrument. Monthly to the sponsor. Once, at the gate, to whoever approves the next step. Anything more frequent upwards turns the pilot into a performance.
And the rule: do not put a target on a metric your team can hit by skipping a check. A target on time-to-hire, or artifacts produced, or adoption, is an instruction to go faster, and the fastest available route through your new process is always the one that omits the impact check. Targets belong on the metrics where the only way to move them is to do the work properly - the share of criteria that are evidenceable, the share of decisions with a same-day reason.
Capability, and the thing your HR team is actually thinking 3 min live
HR has spent two years reading that AI will replace administrative and coordination work, and HR contains a great deal of both. So when a CHRO announces an AI pilot, a proportion of the room hears a headcount plan with a training session attached. If you do not address that in the first five minutes, you will get compliance instead of engagement, and you will not be told the truth about how the tool is performing.
LiveSay the unsaid thing, then train for judgment3 min▶
- Say plainly what is not happening. If this is not a headcount exercise, say so in specific words, in the kickoff, before anything else. If it partly is, do not run the pilot as though it is not - people work that out, and the credibility you lose is not recoverable inside the same programme.
- Name what the time is for. Hours moved off drafting have to land somewhere visible: calibration, manager coaching, candidate experience, the internal mobility work a5 pointed at. "We freed up time" with no destination reads as "we found slack".
- Train judgment, not prompts. Prompt technique is the shortest-lived part of this and it changes every quarter. What lasts is what to check, what to refuse, what an evidenceable criterion looks like, and when to say "this needs a person". That is the whole HR track, and it is a professional skill rather than a tool skill.
- Nobody's accountability moved. Say it as reassurance and as a boundary in the same sentence. The hiring manager still owns the hire. Nothing about the tool changes who answers for the decision, which is protective for your team and non-negotiable for everyone else.
- Give people permission to refuse. An HR partner who says "I am not putting that in the tool" should be visibly backed, publicly, the first time it happens. That single moment sets the culture of the rollout more than any policy document will.
The question nobody asked in the room. A CHRO ran a well-designed kickoff: scope, bar, sampling, timeline. Excellent questions, good energy, and then three separate people asked their manager the same thing privately that afternoon - is this how they start reducing the team. Nobody had said otherwise, because it had not occurred to anyone that it needed saying. The second kickoff opened with one sentence about headcount and closed with the same slides. The questions in the room were completely different, and so was the quality of what came back in week 3.
Write the pilot charter ★ 12 min · everyone builds
One page. If it does not fit on one page, the pilot is too big. Write it for your own function, and be specific enough that somebody could disagree with it.
Name the team. A real one, with a real lead who has agreed. Not "recruitment" - the four people hiring for the engineering roles.
Name the one artifact. Job ads, interview guides or onboarding plans. If you are reaching for screening, re-read Part 1 and pick again.
Write the quality bar as four pass-or-fail lines: criteria evidenceable, ratings cite evidence, decisions name a person, impact check ran. Add nothing that cannot fail.
Set the sampling rule: what share, sampled by whom, reviewed against what, and exactly what happens when one fails.
Write the go/no-go criteria for week 6, with numbers, plus the stop conditions that would end it earlier.
Write the not-touching list. Explicit and specific: no screening, no ranking, no named-individual outputs, no tier 3 or worse data from a5's map, no employee monitoring feeds. This is the section your DPO and your team will both read first.
The board paragraph ★ 10 min · build your own
Three sentences. What we are doing, what we are not doing, what we measure, and what would make us stop. No adjectives - no "transformative", no "responsible", no "cutting-edge". The absence of adjectives is what makes a board believe the rest of it.
Sentence one: the doing. One team, one artifact, six weeks, with the human decision explicitly unchanged.
Sentence two: the not-doing. Name screening, ranking and named-individual outputs specifically, because that is what a board has read about.
Sentence three: the measuring and the stopping. Two or three honest metrics, and the condition that ends it.
Now strike every adjective. Read what is left. If it still says something, it is done; if it has gone empty, the paragraph was the adjectives.
Test it on someone outside HR who will tell you the truth. The question is not "does this sound good" but "what would you ask me after this".
The stop condition that won the approval. A CHRO expected the "what we are not doing" section to be the hard part of the board conversation. It went by in seconds. The question that took twenty minutes was the last line, and the reason it took twenty minutes is that no other technology proposal that year had arrived with a written way to end it. The board approved the pilot on the strength of the stop condition, and the CHRO now opens with it.
Monday, and where to go next 2 min live
Six sessions, and the whole thing compresses into three rails and one habit. The rails: a person decides and the record shows which person. Rate the job, not the polish. People data is not prompt fodder. The habit: write down the reason on the day you have it.
LiveWhat to do on Monday2 min▶
- Send the a5 tier map to your DPO with the tier 3-or-worse tool contact list attached. One email, and it is the highest-value thing you will send this month.
- Pick the team and the artifact and have the conversation with the team lead before you tell anyone else. If they are not willing, pick a different team rather than a different message.
- Find out what is already running that nobody approved - the ATS feature, the note-taker, the survey summariser. Most functions are further into this than their leader thinks.
- Book the kickoff and write the headcount sentence first. Everything else on the agenda works better after it.
- Re-verify the dates. The AI Act's Annex III timing has moved once already, and NYC's enforcement posture keeps developing. Nothing on this page is a compliance source, including this line.
Try it yourself - this week ◐ 45-60 min total
- Finish the pilot charter on one page and send it to the team lead before the sponsor. If the lead will not sign it, the charter is not ready and the sponsor's approval will not save it.
- Write the four-line quality bar and then find one real artifact from last month that would have failed it. If nothing fails, the bar is decoration.
- Pick your metric set. Name the two vanity metrics you are explicitly not reporting, and be ready to say why when someone asks for them.
- Draft the board paragraph, strike every adjective, and read what survives.
- Write the headcount sentence for the kickoff. Say it out loud once before you say it to the team.
- Book yourself into HR-track b4. Bring the list of tools already touching your hiring process.
Official sources covered
Taught from public regulation, selection research and vendor documentation. Certification content and any actual legal position stay with their owners. This course is education for HR leaders, not legal advice, obligations vary by jurisdiction, and the dates referenced here have already moved once. Verify the current position before you act. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Your talent team wants to pilot AI on resume screening, because that is where the hours are. What do you say?
Blast radius sets the order. Screening carries obligations, is irreversible for the candidate, and asks a team with no reps at checking AI output to practise on somebody's career. The drafting hours are larger anyway and carry none of the exposure.
2 · Six weeks in, time-to-hire is down 30%. What is your first question?
Time-to-hire improves for two opposite reasons and never distinguishes them on its own. Pair it with the sampled pass rate, the share of decisions with a same-day recorded reason, and the four-fifths ratio across the wave. Alone it is the most dangerous number on the page.
3 · What does the b4 simulator's sixth lever - the impact check - teach about metrics?
Five fixes take the slate from 61 quality and a 0.09 ratio to 100 and 1.00. The sixth moves neither number, because it is not a fix - it catches the problem you did not anticipate. It is also the lever most likely to be dropped under a speed target, which is why targets do not go on speed metrics.