Where we are
Session a1 drew the boundary: AI drafts, a person decides. Session a2 mapped the law around it, from NYC's Local Law 144 to the EU AI Act's Annex III. Both of those are statements of position. Tonight is the part nobody puts on a conference slide, because it looks like admin: the machinery. Who asks the vendor the awkward question. Who commissions the audit and who is allowed to run it. Who writes the candidate notice and how many days ahead it goes out. What exists in writing afterwards, and who owns each piece of it. Kestrel's hiring wave is already running, so every one of these has a name attached by the end of the session or it does not exist.
The vendor gate: questions that actually discriminate between vendors 7 min live
Every HR tech vendor answers "is your tool fair?" with a yes. That question sorts nothing. The questions that sort are the ones a good vendor can answer in a sentence and a weak vendor has to schedule a call about. Four of them do most of the work.
LiveThe four questions, and what a real answer sounds like4 min▶
Ask these in the room, in this order, and write the answers down verbatim. The wording matters because vague answers survive paraphrasing.
- "Has this tool been independently audited for bias, when, by whom, and may I see the report rather than the marketing summary?" Three of those four clauses do the work. Plenty of vendors have an audit. Fewer have an independent one, fewer still have one from this year, and a small number will hand over the document.
- "What data was it trained and validated on?" Which population, from which years, from which labour market, and validated against what outcome. A tool trained on one country's hiring history and sold into another is not obviously wrong, but it is obviously a question.
- "What does it do when it is uncertain?" A system that abstains, flags low confidence, or refuses to score a thin application is a system someone designed carefully. One that always returns a confident number has been designed to look decisive, which is a product decision, not a scientific one.
- "What will you commit to contractually?" Specifically: notifying you of material model changes, supporting your audit obligation with data, telling you what happens to candidate data, and letting you exit with your records. Everything else is a slide.
The silent model update. A team ran the same screening tool for two years and was pleased with its consistency. Between one quarter and the next, the vendor shipped a scoring change that nobody told them about, because nothing in the contract required it. The team's own numbers moved and they spent six weeks looking for the cause in their process. The lesson is not that the vendor behaved badly. It is that "tell us when the model changes" was a sentence nobody had asked for, so it did not exist.
Self-studyThe other six questions, and how to score all ten3 min read▶
Score every answer red, amber or green, and record the score with the date and the person who answered. Amber is a legitimate outcome. "We do not know yet, here is who will find out and by when" is an amber. "Trust us" is a red wearing a smile.
| Question | What green looks like |
|---|---|
| 5 · What exactly does the tool decide or influence? | A written description of where it sits in the funnel, which we can verify |
| 6 · What happens to our candidates' data? | Where it is stored, how long, who can see it, whether it trains anything |
| 7 · Can a human override it, and is the override recorded? | Yes, and the override plus the reason lands in an exportable log |
| 8 · What accessibility and accommodation provision exists? | A documented alternative route that does not disadvantage the candidate |
| 9 · Which jurisdictions have you supported through an audit? | Named ones, with what the vendor did and what the employer had to do |
| 10 · What have you seen go wrong at other customers? | A specific story. A vendor with no failure stories has no memory |
Self-studyWhy buying does not transfer the duty2 min read▶
Under NYC Local Law 144, which has been live since 1 January 2023 and enforced by the Department of Consumer and Worker Protection since 5 July 2023, the employer using an automated employment decision tool needs an independent annual bias audit, a public summary of the results, and at least 10 business days' notice to the candidate before the tool is used. Penalties run from $500 per violation up to $1,500 per day for continuing ones. The obligations attach to the employer, not only to the vendor who built the thing.
In the EU, Annex III point 4 of the AI Act places recruitment and selection AI - targeted job ads, filtering applications, evaluating candidates - alongside performance evaluation, task allocation, monitoring, promotion and termination in the high-risk category, with obligations on risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness. Those Annex III obligations were postponed by the Digital Omnibus agreement from 2 August 2026 to 2 December 2027. That is the position as of mid-2026, and it has moved once already, so check it. None of this is legal advice, and your jurisdiction may differ from both.
The audit and the notice are processes, not paperwork 7 min live
Both of these get treated as documents to produce once and file. They are not documents. They are recurring processes with owners, dates and dependencies, and the moment you draw them as processes you discover they constrain how your funnel is allowed to run.
LiveIndependent, annual, and published3 min▶
Three words, each of which has teeth.
- Independent. Your own data science team scoring your own tool is a useful internal check and it is not an independent audit. Neither, generally, is the vendor auditing the vendor. Someone has to be able to reach an inconvenient conclusion without it costing them a customer or a bonus.
- Annual. Not once at procurement. A cadence, on a calendar, with a named commissioner who books it before the wave rather than after the complaint. If the audit is older than a year when the tool is in use, the cadence has quietly lapsed.
- Published. A public summary of the results is a very different artifact from an internal memo, and writing something you know will be read outside the building changes what you are willing to leave vague. That is a feature.
Practically, someone owns commissioning it, someone owns supplying the data to the auditor, someone owns reviewing the findings, someone owns publishing the summary, and someone owns the actions that come out of it. In most organisations at least three of those five are the same overworked person, which is fine as long as it is written down.
The audit that could not be run. A people team commissioned an independent audit, got a quote, signed it, and then discovered the auditor needed selection outcomes broken down by monitored group per stage of the funnel. Their ATS could export applicants and hires but not the stages in between, and the screening tool's scores were not exportable at all. The audit slipped a quarter, not for legal reasons but for plumbing ones. Ask the auditor what data they need before you buy the tool, and put the answer in question four.
LiveThe notice is a calendar problem3 min▶
The candidate notice looks like copywriting. It is scheduling. Notice has to arrive before the tool is used on the candidate, and NYC's floor is at least 10 business days ahead. Ten business days is two working weeks. That has consequences your talent team will feel immediately.
- It sets a floor under your funnel speed. A role posted Monday cannot have its applications machine-screened on Wednesday if the notice went out with the posting. Either the notice goes out earlier, or the screen happens later, or the screen does not happen.
- It has to reach people who never apply. If a tool is used at the sourcing stage, the notice question arrives before the application does, which is a different plumbing problem entirely.
- Somebody writes it, and it should not be the vendor. The notice describes your process, including the parts the vendor knows nothing about. HR writes it, compliance checks it, and talent operations owns delivering it on time.
- It has to be true. The fastest way to find an ungoverned tool is to try to write an honest notice and discover a sentence you cannot write.
Self-studyWhat the notice should actually say2 min read▶
Short, specific and in the candidate's language. Four things: what automated tools are used and at which stage, what they are used for and what they are not used for, that a named human makes the decision, and how to request a human review or an accommodation. Add the qualifications or characteristics the tool considers, and where required, how to ask about the data retention policy.
Avoid two temptations. The first is legal padding, which converts a notice into something nobody reads and no regulator admires. The second is soft language - "we may use technology to help us" - which is not a notice, it is a hedge. If you cannot describe the tool plainly, you do not understand the tool well enough to be using it on people.
The record: six things that must exist afterwards 6 min live
Rail 1 says a person decides and the record shows which person. This is what "the record" means in practice. It is not a folder. It is six specific artifacts, each with an owner, produced at the time rather than assembled later.
LiveWhy it has to be written at the time3 min▶
Because a rationale written after a challenge lands is a rationale written by someone who already knows there is a challenge, and everybody reading it knows that. The document is not lying. It simply cannot carry the weight you need it to carry.
- Criteria before the slate, not after it. Criteria written after you have seen the candidates are a description of your preferences with job-related vocabulary applied.
- The slate matters as much as the outcome. Who was in the pool is the denominator of every fairness number you will ever need. Systems that only keep hires throw the denominator away.
- Human oversight has to leave a trace. The AI Act's oversight expectations, and any serious auditor, will look for evidence that a person could have said no and sometimes did. An override log with zero overrides in it is a finding, not a clean record.
- Version the tool. "The screen scored her 62" is meaningless in eighteen months unless you can say which screen, configured how.
The rejection nobody could explain. Eleven months after a hiring wave closed, a candidate asked why they had not progressed. The team still had the outcome, the ATS status and a date. They did not have the criteria used at the time, the evidence behind the rating, or the name of the person who made the call, because the manager who ran it had left. Three people spent two days reconstructing a story that was probably true and looked exactly like a story reconstructed over two days. Recording it live would have cost about ninety seconds.
Self-studyRetention, access and the tool register2 min read▶
- Keep it long enough to answer a challenge, no longer. Retention that is too short leaves you defenceless and retention that is indefinite is its own data protection problem. Your DPO sets the number, and it belongs in writing next to the record itself.
- Decide who can read it. Decision records contain judgements about identifiable people. Access should be narrower than the ATS as a whole.
- Keep a tool register. Every tool touching the funnel, what it does, whether it substantially assists a decision, when it was last audited, who the vendor contact is, and which of your ten gate answers were amber. Most organisations do not have this list, which is why most organisations cannot answer the AEDT question quickly.
- Include the things nobody calls AI. CV parsers, knockout questions, keyword filters and ranked search all sit in the funnel and can filter a person out before a human sees them.
Run the vendor gate on a tool you already use ★ 12 min · everyone builds
Not a tool you are evaluating. One that is live in your funnel right now, ideally one nobody thinks of as AI. Ten questions, honest answers, one colour per answer. Most teams discover they cannot answer question two.
Pick the tool. If your instinct is to pick the newest, resist it and pick the oldest instead - the one that arrived before anyone was watching.
Write the ten questions down the left of a page. Answer each one from what you actually know today, not from what you could find out. Blank is a legitimate answer and it scores red.
Colour each answer. Green means you have the artifact. Amber means you know who will get it and by when. Red means nobody in your organisation knows.
Count the reds. Then decide, out loud, whether the tool keeps running while they are open. That decision is the actual governance moment, and it belongs to a person in the room.
Send the reds to the vendor as written questions with a date. Verbal answers to written questions have a way of not surviving contact with an auditor.
Write the escalation path for a failed impact check ★ 10 min · build your own
In b4 the practitioners watch a slate score 0.09 against a 0.80 threshold. Suppose that happens on a Tuesday in your organisation. What happens next, by name, by hour? "Review it" is not an answer. It is the absence of one.
Name who is told, and within how long. Twenty-four hours is a reasonable default and any number is better than none. Include who tells them.
Name who has the authority to pause the slate. This is the sentence most teams have never written, and it is the one that decides whether the process is real, because pausing a slate costs somebody their week.
Name who decides the remedy: re-run the screen with fixes, re-design the criteria, or reopen the pool. Different remedies, different owners, different costs. Decide the routing now, not at 4pm on the Tuesday.
Say what gets recorded and where: the number, the inputs, the decision, the remedy, and the date. This lands in the record from Part 3.
Add the sentence about what you do not do: you do not explain the number away in the room, and you do not hold it until it improves. A finding that is only escalated when it is bad news is a finding nobody will bring you twice.
The escalation path with no brake. One organisation had a beautiful monitoring dashboard, a monthly fairness review and a named owner. What it did not have was anyone with the authority to stop a live hiring slate, so when a number came back badly during a wave, the review noted it, the wave finished, and the fix went into the next cycle. Everybody behaved reasonably. Twelve people were hired through a process the organisation had already decided was defective. The missing sentence was eleven words long.
Try it yourself - this week ◐ 45-60 min total
- Build the tool register. Every tool touching your funnel, including CV parsers and knockout questions. Mark which of them substantially assist a decision. This is the list a regulator would ask for first and it usually takes an afternoon.
- Run the ten-question gate against your two highest-risk tools and send the reds to each vendor in writing, with a date for a reply.
- Find out when your last independent bias audit was, if there was one, and who commissioned it. If the answer is nobody, book the conversation rather than the audit - the conversation is the finding.
- Map the notice timing against your actual funnel calendar. Mark the day a candidate first meets an automated step and count backwards.
- Finish the escalation path and get the person named as having pause authority to confirm, in writing, that they know they have it.
- Bring one open governance question to a4, where the fairness arithmetic and the candidate-facing explanation get built.
Official sources covered
Taught from public regulation, published guidance and vendor documentation. This course is education for HR professionals, not legal advice, and obligations vary by jurisdiction. Any actual legal position stays with your counsel and your DPO. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · A vendor tells you their tool was "audited for bias by our data science team last year". What is missing?
Three words carry the weight: independent, annual, published. A team scoring its own tool cannot reach an inconvenient conclusion at no cost to itself, and "last year" needs a date before you can tell whether the cadence has lapsed. Ask for the report, not the summary.
2 · An impact check on a live slate comes back at 0.42. What must your escalation path already name?
The ratio is a screening signal that starts an investigation, not a verdict. What turns it into action is having named the roles in advance, especially the pause authority - the one sentence most organisations have never written, because using it costs somebody their week.
3 · Why must the decision-maker and the reason be recorded at the time rather than when asked?
Rail 1 in its operational form. The reconstructed version may well be true and still cannot carry the weight, because everyone reading it knows when it was written. Ninety seconds at the time replaces two days of archaeology later.