Short answer: Use AI to
design your hiring process— never to
decide on candidates. The moment a model scores, ranks or filters applicants, it may become a regulated automated employment decision tool: NYC’s Local Law 144 requires an annual independent bias audit, and under the EU AI Act, recruitment AI is classified high-risk with full obligations enforceable from
2 August 2026. Used upstream — on the scorecard, the job advert and the interview questions — AI measurably improves hiring, because structured interviews predict performance roughly twice as well as unstructured ones.
TL;DR — Key Takeaways
Design, don’t decide. AI on your process is unregulated and effective. AI on your candidates is regulated and risky.Write the scorecard before the job ad. Most bad hires trace to a vague scorecard, not a bad interview.Structured interviews roughly double predictive validity— about .42 vs .19 in corrected estimates (Sackett et al.).** Ask what they**Hypotheticals produce rehearsed theory; past behaviour produces evidence.did, not what theywoulddo.Bias-check your own materials, not your candidates. The audit belongs on your job ad and criteria — that’s the use AI is genuinely good at.
✔ Best for Founders and managers hiring without a dedicated recruiter, small teams making their first few hires, and anyone who wants a defensible, repeatable process rather than gut-feel interviews.
✕ Skip if You’re evaluating ATS or AI screening vendors (different question — start with their bias audit documentation), or you need jurisdiction-specific compliance advice, which requires employment counsel.
On this page
Where’s the legal line on using AI in hiring? #
The line falls between designing your process and evaluating people. Writing a scorecard, drafting a job advert, or generating interview questions involves no candidate assessment and carries no special regulatory burden. Scoring, ranking, shortlisting or filtering applicants does — and that’s where obligations attach.
✔ Generally safe — designing the process
- Defining role outcomes and competencies
- Writing the scorecard
- Drafting the job description
- Generating structured interview questions
- Auditing your ownmaterials for bias - Preparing debrief structures
✕ Regulated — deciding on people
- Scoring or ranking CVs
- Shortlisting or filtering applicants
- Assessing recorded video interviews
- Predicting a candidate’s performance
- Producing a “fit score” per person
- Anything substantially replacing human judgement
⚠ What the rules actually say.
NYC Local Law 144 defines an automated employment decision tool (AEDT) as a computational process derived from machine learning, statistics or AI producing a simplified output used to
substantially assist or replace human judgementin hiring or promotion. It requires an
annual independent bias audit, and applies extraterritorially — if the candidate resides in NYC, you’re in scope regardless of where you’re based. Reported penalties run
$500–$1,500 per violation, per day. Under the
EU AI Act, AI used in recruitment and selection is
high-risk (Annex III), with full obligations enforceable from
2 August 2026 and penalties reaching
€35 million or 7% of global turnover; it applies to roles based in the EU whatever your location. In the US, the
EEOC has indicated Title VII disparate-impact analysis applies to AI hiring tools.
This is general information, not legal advice. Requirements vary by jurisdiction and change frequently. Verify current obligations for every location you hire in, and take employment law advice before deploying any tool that assesses candidates.
The good news: the upstream uses are where the actual leverage is anyway. Meta-analytic research consistently finds structured interviews predict job performance far better than unstructured ones — Schmidt and Hunter reported .51 vs .38, and more recent corrected estimates from Sackett and colleagues put it at roughly .42 vs .19. Structure is the single biggest improvement available to most hiring teams, and it’s entirely a design problem.
Step 1
Scorecard Outcomes first
Step 2
Job ad From the scorecard
Step 3
Screen Same questions, rubric
Step 4
Interview Past behaviour
Step 5
Bias check On your materials
The role context block #
ROLE CONTEXT (paste at the top of every hiring prompt):
BUSINESS: [WHAT WE DO, IN ONE SENTENCE]
SIZE & STAGE: [HEADCOUNT, GROWTH STAGE]
THE ROLE: [TITLE — and what problem it exists to solve]
WHY NOW: [WHAT'S BREAKING THAT THIS HIRE FIXES]
REPORTS TO: [WHO, AND HOW MUCH SUPERVISION EXISTS]
TEAM AROUND THEM: [WHO THEY'LL WORK WITH DAILY]
BUDGET: [SALARY RANGE — and whether it's negotiable]
CONSTRAINTS: [REMOTE/HYBRID/ONSITE, LOCATION, HOURS,
VISA/RIGHT-TO-WORK REALITIES]
WHAT WE CAN'T OFFER: [BE HONEST — no big brand name,
limited budget, no team yet, messy processes]
RULES:
- Never invent details about our business. Ask if you need one.
- Do not use gendered or age-coded language.
- Do not suggest criteria that measure background, personality
or demographics rather than ability to do this job.
- Flag any requirement that isn't genuinely necessary.
“What we can’t offer” improves output more than anything else in the block. Job adverts that acknowledge a real trade-off attract candidates who are fine with it — and repel the ones who’d leave in four months.
1. How do I write a hiring scorecard? #
Start with outcomes, not responsibilities. Most bad hires trace back to a vague scorecard rather than a bad interview — nobody agreed what success looked like, so everyone assessed something different. Define four to six measurable things the person must achieve in twelve months, then work backwards to the competencies that predict them.
(run this first)
[PASTE ROLE CONTEXT BLOCK]
Before we write any job advert, help me define success.
1. OUTCOMES — write 4-6 things this person must DELIVER in
their first 12 months. Each must be:
- Measurable (a number, a date, or an observable state)
- Specific to MY business, not generic responsibilities
- Something we'd genuinely evaluate them on
Split into: first 90 days, and by month 12.
2. COMPETENCIES — for each outcome, name the ONE competency
that most predicts it. Not a personality trait — an
observable, demonstrable capability.
3. EVIDENCE — for each competency, what evidence in someone's
history would demonstrate it? Be specific about what to
look for.
4. THE SCORING RUBRIC — for each competency, describe what a
1, 3 and 5 out of 5 looks like in a candidate's answer.
These descriptions are what every interviewer scores
against.
5. MUST-HAVE vs NICE-TO-HAVE — split the competencies, and
challenge me on anything I've marked must-have that
probably isn't.
6. FLAG: does any outcome suggest I'm actually hiring for TWO
different roles? Say so plainly.
7. FLAG: any requirement here that's a proxy for background
rather than ability — degree requirements, years of
experience, specific employer names, "culture fit".
Name them and suggest what to measure instead.
Why it works: point 4 is what makes the rest function. Without written descriptions of what a 1 and a 5 look like, “scoring out of 5” is just numbered intuition — and every interviewer applies a different scale. Point 6 catches the classic early-stage mistake of writing one job ad for two jobs.
Get the Hiring Prompt Kit
Every prompt in this guide, plus a scorecard template, structured interview score sheets, the bias-check pass, and the debrief framework. Free.
Get the kit →
2. How do I write a job description with AI? #
Generate it from your scorecard, not from a template. A job advert written from outcomes attracts people who can deliver them. One written from a generic template attracts people good at responding to generic templates — a different and much larger group.
[PASTE ROLE CONTEXT BLOCK]
THE SCORECARD: [PASTE OUTPUT FROM PROMPT 1]
Write the job advert. Structure:
- OPENING (40 words): the specific problem this person will
own. Not "we're an exciting fast-growing company."
- WHAT YOU'LL ACHIEVE: the outcomes, in plain language,
written as what they'll accomplish rather than tasks.
- WHAT WE'RE LOOKING FOR: the must-have competencies only.
Describe capability, never a personality type.
- HONEST BIT: what's hard about this role, and what we
can't offer. Include it — it filters better than any
screening question.
- PRACTICALS: salary range, location/remote reality, hours,
the actual hiring process and how long it takes.
Rules:
- Under 500 words total.
- No "rockstar", "ninja", "wear many hats", "work hard play
hard", "like a family", "fast-paced environment".
- No gendered or age-coded language ("young and dynamic",
"recent graduate", "digital native", "aggressive").
- Do not list a degree requirement unless the role legally
requires one — flag it if I've asked for one that isn't.
- No "X+ years experience" unless you can tell me what
specifically those years would give someone.
- Write at 8th-grade reading level.
Then list: which sentence is most likely to deter a
qualified candidate who'd actually be great, and why.
3. How do I structure the screening stage? #
Same questions, same order, written rubric, human scorer. Structure is what makes candidates comparable — and it’s also what keeps the process defensible. Note the deliberate design here: AI builds the screening instrument; a person applies it.
[PASTE ROLE CONTEXT BLOCK]
THE SCORECARD: [PASTE]
Design a 20-minute structured phone screen.
Give me:
1. SIX QUESTIONS, asked of every candidate in the same order.
- 2 on the single most important must-have competency
- 2 on genuine dealbreakers (practical constraints,
right to work, notice period, salary expectations)
- 1 on why this role specifically
- 1 open question that lets them raise what matters to them
2. For EACH question, a written 1-3-5 rubric so any
interviewer scores it the same way.
3. WHAT A WEAK ANSWER SOUNDS LIKE — specifically, the answer
that sounds confident and polished but contains no
evidence. This is the main failure mode.
4. FOLLOW-UP PROBES for each question — what to ask when an
answer is vague. Give exact wording.
5. THE DISQUALIFY BAR — what specific answer means "no",
defined BEFORE we start, so we don't rationalise later.
6. WHAT TO TELL THE CANDIDATE about next steps and timing.
Rules:
- Every question must relate to the ability to do this job.
- No questions about age, family, health, nationality,
religion, or anything not job-related.
- I will score these myself. Do not offer to score
candidates.
Don’t paste CVs into a chatbot to rank them. Beyond the AEDT and high-risk classification issues above, candidate CVs contain personal data you likely have no lawful basis to process this way. Use AI to build the rubric; apply it yourself.
4. What interview questions should I ask? #
Ask what they did, not what they would do. Hypothetical questions (“how would you handle a difficult client?”) measure how well someone theorises. Behavioural questions (“tell me about the last time a client escalated — what did you do?”) produce evidence you can check.
[PASTE ROLE CONTEXT BLOCK]
THE SCORECARD: [PASTE]
INTERVIEW LENGTH: [MINUTES] · WHO'S INTERVIEWING: [ROLES]
Design the structured interview.
For EACH must-have competency:
1. ONE primary question about SPECIFIC PAST BEHAVIOUR.
Must start with "Tell me about a time..." or "Walk me
through the last time..." — never "how would you" or
"what's your approach to".
2. THREE probing follow-ups that dig for what they
personally did versus what their team did. Include one
that tests whether the example is real (ask for a detail
they'd only know if it happened).
3. THE 1-3-5 RUBRIC — what a weak, adequate and strong
answer actually contains. Focus on evidence and
specificity, not confidence or polish.
4. THE RED FLAG — the answer pattern that should worry me
even if it sounds good. Usually: all "we", no "I"; no
specifics; no mention of anything going wrong.
Then:
- Which interviewer should ask which questions, so we're
not all assessing the same thing.
- 3 questions I should NOT ask, and why (legal or useless).
- The one question most likely to distinguish a strong
candidate from a merely competent one.
Rules:
- Same core questions for every candidate.
- No brainteasers, no "sell me this pen", no questions
designed to unsettle.
- Nothing about protected characteristics or anything
unrelated to job performance.
5. How do I run a bias check on my own materials? #
Audit the process, not the people. AI can’t remove bias from hiring — it can reproduce and scale it, which is exactly why automated screening attracts scrutiny. What it does well is spot exclusionary language and unjustified requirements in materials you wrote. That’s a genuine, low-risk, high-value use.
[PASTE ROLE CONTEXT BLOCK]
Audit these hiring materials. Be specific and quote the text.
JOB ADVERT: [PASTE]
SCORECARD: [PASTE]
INTERVIEW QUESTIONS: [PASTE]
Report on:
A. CODED LANGUAGE — words associated with gender, age,
class or cultural background. Quote each and suggest a
neutral alternative.
B. UNJUSTIFIED REQUIREMENTS — every requirement that isn't
genuinely necessary to perform this job. Pay particular
attention to: degree requirements, "X+ years", specific
named employers or tools, and "culture fit".
For each, ask: what capability is this actually a proxy
for, and could we measure that directly instead?
C. ACCESS BARRIERS — anything that would exclude a capable
candidate for reasons unrelated to ability: unpaid task
stages, rigid hours, undisclosed salary, an unnecessarily
long process, assumptions about equipment or location.
D. QUESTIONS THAT INVITE BIAS — anything in the interview
that would surface information about protected
characteristics, or that rewards similarity to me rather
than competence.
E. THE "CULTURE FIT" TEST — wherever I've used that phrase
or similar, tell me what specific BEHAVIOUR I actually
mean, so we can assess it directly.
F. WHAT'S MISSING — what would make this process fairer or
more accessible that I haven't included?
Rank findings by how much each is likely to cost me a good
candidate.
Section E is the sharpest part. “Culture fit” is often a post-hoc justification for preferring someone who resembles the interviewer — a weak predictor of performance and a discrimination risk. Naming the behaviour (“gives direct feedback,” “works well unsupervised”) makes it assessable and defensible.
What separates a good hiring prompt from a bad one? #
| Move | ❌ Weak (and often risky) | ✅ Strong |
|---|---|---|
| Target | “Rank these 40 CVs for me” | “Build the rubric I’ll use to score them myself” |
| Starting point | “Write a job description for a marketing manager” | Scorecard first, advert generated from the outcomes |
| Requirements | “5+ years experience, degree required” | “Flag any requirement that’s a proxy for background not ability” |
| Questions | “How would you handle a difficult stakeholder?” | “Walk me through the last time a stakeholder escalated” |
| Scoring | Gut feel, scored out of 5 with no definitions | Written 1-3-5 rubric agreed before anyone interviews |
| Bias | “Screen these candidates without bias” | “Audit MY advert and criteria for exclusionary language” |
The through-line: point AI at your artefacts, not at people. That’s simultaneously the more useful application and the one that keeps you out of regulatory scope.
Can you show a real prompt with its actual output? #
Prompt 1 (the scorecard) run for a first operations hire at a 9-person e-commerce business.
CONTEXT: 9-person DTC homeware brand, ~£2m revenue.
Hiring: Operations Manager, our first ops hire.
Why now: founder is doing supplier management, stock
forecasting, and customer service escalations, and stock-outs
have hit 3 of our top 10 SKUs this quarter.
Reports to: founder (me). Little supervision available.
Budget: £42-48k. Can't offer: brand recognition, a team to
manage, or tidy existing processes.
Define success before we write the advert.
1. Outcomes
First 90 days:
- Documented the supplier ordering and stock forecasting process end to end — currently exists only in the founder’s head
- Taken over all supplier communication, with the founder copied but not required
By month 12:
- Zero stock-outs on top 10 SKUs (baseline: 3 this quarter)
- Stock holding reduced or held flat while revenue grows — no fixing availability by over-ordering
- Customer service escalations resolved without founder involvement in 90%+ of cases
- A named backup able to run ordering during absence
2. Competencies (one per outcome)
- Process documentation from tacit knowledge — predicts outcomes 1 and 6 - Supplier relationship management under pressure — predicts 2 - Demand forecasting with imperfect data — predicts 3 and 4 - Independent judgment on escalations — predicts 5
4. Rubric — “Demand forecasting with imperfect data”
1: Describes using a system that already existed. Can’t explain the logic. Says forecasting was “handled by the tool.”3: Built or adapted a forecast. Can explain inputs. Struggles to describe what they did when it was wrong.5: Names specific variables and why. Describes a forecast that failed, what they learned, and how they changed the method. Talks about the trade-off between availability and stock holding unprompted.
6. Flag — two roles?
Possibly. “Supplier and stock management” and “customer service escalations” are different skill sets. At 9 people they can reasonably sit together, but be aware you’re weighting the hire toward supply chain — the escalations outcome is the one most likely to slip. Consider whether escalations should be a stated secondary responsibility rather than a core outcome.
7. Flag — proxies for background
- You haven’t specified a degree — keep it that way; nothing here requires one. Watch “little supervision available.” This is a real constraint, but if it becomes a screening criterion like “self-starter,” you’ll unconsciously favour candidates who present with confidence over those who work well autonomously. These aren’t the same trait. Measure it by asking for a specific example of a period they worked without a manager.No e-commerce experience requirement. Forecasting and supplier management transfer from hospitality, manufacturing and logistics. Requiring DTC experience shrinks your pool at this salary for little predictive gain.
Note point 7: the model caught that “self-starter” would function as a proxy for confident presentation rather than actual autonomy — a distinction that shows up in a lot of small-company hiring and rarely gets named. That’s the kind of finding that changes who you end up hiring.
Level-up: the interview debrief prompt #
This is the part competitors’ hiring prompt lists don’t have. Everyone focuses on asking better questions. Almost nobody addresses what happens in the debrief — where a structured process quietly collapses back into “I just liked her more.”
Run this after each interview, with your raw notes. It separates evidence from impression.
[PASTE ROLE CONTEXT BLOCK]
THE SCORECARD & RUBRIC: [PASTE]
MY RAW INTERVIEW NOTES:
[PASTE EVERYTHING — unedited, including impressions,
things that felt off, and anything I liked about them.]
MY GUT SCORE: [OUT OF 5, AND MY INSTINCT: HIRE / NO]
Audit my assessment. Be harsh. Six sections:
A. EVIDENCE vs IMPRESSION — split my notes into what the
candidate demonstrably DID (specific examples with
detail) versus my IMPRESSIONS of them (confident,
likeable, sharp, "got it"). Only the first column
supports a hiring decision.
B. SCORECARD COVERAGE — for each competency, did I actually
gather evidence? Mark EVIDENCED, THIN, or NOT ASSESSED.
Be strict: one vague reference is THIN.
C. UNASKED QUESTIONS — what did I fail to probe? Give exact
wording, and say whether I can still ask it at a later
stage.
D. BIAS CHECK ON ME — is any part of my reaction driven by
similarity to me, communication style, accent, confidence,
background, or where they worked before? Quote the note
that suggests it. Ask directly whether I'd score the same
answer identically from a different candidate.
E. THE GAP — does my gut score match my evidence? If I rated
them highly on thin evidence, say so plainly. If I rated
them low but the evidence is strong, say that too.
F. THE HONEST VERDICT — based on evidence only, what does
the scorecard say? If the answer is "we don't have enough
evidence to decide", say that rather than inventing one.
Rules:
- Do not tell me my instinct is probably right.
- Do not evaluate the candidate as a person — assess the
quality of MY evidence and MY reasoning.
- If my notes are too thin to assess something, that's a
finding about my interviewing, not about them.
Why this is the unlock: Section D asks the question that structured hiring exists to force — would I score this same answer identically from a different candidate? — and it asks it while you can still act. Section A is what makes it work: written out in two columns, most debrief notes turn out to be mostly impression. Crucially, the prompt assesses your reasoning, not the candidate, which keeps it on the safe side of the design/decide line.
Setup tip: have every interviewer run this independently before the group debrief. Comparing evidence columns is a far better conversation than comparing gut feelings — and it stops the most senior voice in the room setting the consensus.
What must AI never do in your hiring process? #
| Never | Why |
|---|---|
| Score, rank or filter candidates | May constitute an AEDT under NYC Local Law 144 (annual independent bias audit required) and is high-risk under the EU AI Act from 2 August 2026. |
| Process CVs and personal data casually | Candidate data carries data-protection obligations. Pasting CVs into a consumer chatbot is unlikely to have a lawful basis. |
| Assess video or recorded interviews | Automated assessment of candidates’ speech or appearance attracts specific regulation in several jurisdictions and has a documented bias record. |
| Make the final decision | Accountability for a hiring decision rests with you. “The tool suggested it” is not a defence. |
| Infer anything about protected characteristics | Inferring age, ethnicity, health, or family status from a CV or profile is both discriminatory and unlawful in most jurisdictions. |
The reliable pattern: AI builds the instrument; a trained human applies it. Every prompt in this guide is deliberately constructed to keep that boundary intact.
Which model for which task? #
Prompts are model-agnostic. Practical notes as of July 2026:
| Task | Best fit | Why |
|---|---|---|
| Scorecard & question design | Claude or ChatGPT | Both hold a rubric structure and sustain the challenge-my-assumptions stance without softening. |
| Interview debrief audit | Claude | Large context holds full notes across multiple interviewers, which matters for spotting inconsistent scoring. |
| Bias & language audit | Any | A well-constrained prompt does the work; model choice barely matters here. |
| Salary benchmarking | A model with live web search | Training data alone produces outdated ranges. Verify against current market sources before quoting a number. |
| Anything touching candidate data | Enterprise tier with a data agreement | Consumer tiers may lack the data-processing terms your obligations require. Check before you paste. |
We re-check these notes whenever a major model ships or regulation changes. Given the 2 August 2026 EU AI Act milestone, verify current obligations before relying on anything here.
Frequently asked questions #
Can I use ChatGPT to screen resumes and rank candidates?
Be extremely careful — doing so may bring you within scope of laws governing automated employment decision tools. NYC’s Local Law 144 applies to computational processes producing a simplified output used to substantially assist or replace human judgement in hiring, and requires an annual independent bias audit. Under the EU AI Act, AI used in recruitment is high-risk, with full obligations enforceable from 2 August 2026. Using a general chatbot to score or rank applicants can create legal exposure. Use AI to design your process and keep candidate evaluation with trained humans. This is general information, not legal advice.
What are the best AI prompts for hiring?
The most valuable hiring prompts operate before any candidate applies. They define measurable twelve-month outcomes for the role, convert those into a scorecard, generate behavioural interview questions tied to each competency, and audit your own job advert for exclusionary language. Prompts that evaluate candidates are both the least useful and the most legally risky.
Do structured interviews actually work better?
Yes, and the evidence is long-standing. Meta-analytic research consistently finds structured interviews predict job performance substantially better than unstructured ones. Schmidt and Hunter reported validity of .51 for structured versus .38 for unstructured, and more recent corrected estimates from Sackett and colleagues put structured interviews at approximately .42 against .19 — roughly double the predictive power.
How do I write a hiring scorecard?
Start with outcomes rather than responsibilities. Define four to six things the person must achieve in their first twelve months, each measurable and specific to your business, then identify the competency that predicts each outcome and the evidence that would demonstrate it. Every interviewer scores against those same criteria — which is what makes comparisons between candidates meaningful.
Can AI remove bias from hiring?
No. AI can reproduce and scale bias present in the data or in the criteria you give it, which is precisely why automated hiring tools attract regulatory scrutiny. What AI can usefully do is audit your own materials — flagging exclusionary language in a job advert, requirements that aren’t justified by the role, and criteria that measure background rather than capability. That’s a check on your process, not a substitute for human judgement.
What interview questions should I ask?
Ask about specific past behaviour rather than hypothetical situations. Questions beginning “tell me about a time you…” produce evidence; “what would you do if…” produces rehearsed theory. Ask the same core questions of every candidate, and decide in advance what a weak answer sounds like so a confident delivery isn’t mistaken for a strong response.
Is it legal to use AI to write job descriptions?
Writing job adverts with AI assistance is generally uncontroversial, because no candidate is being evaluated. The obligations under laws such as NYC Local Law 144 and the EU AI Act attach to tools that assess, score, rank or filter people. You remain responsible for the content of the advert — including any discriminatory or unjustified requirements it contains — so review the output rather than publishing it unread.
How do I avoid hiring for “culture fit” when I mean similarity?
Define the specific behaviours you mean before the interview, and score against them. Culture fit is frequently a post-hoc justification for preferring a candidate who resembles the interviewer — both a poor predictor of performance and a discrimination risk. Replacing the phrase with named behaviours, such as “gives direct feedback” or “works well without supervision,” makes the criterion assessable and defensible.
Download: The Hiring Prompt Kit
All five prompts plus the debrief audit, a scorecard template with 1-3-5 rubrics, structured interview score sheets, the bias-check pass, and a pre-hire compliance checklist.
Enter your email and we’ll send the kit plus a short monthly prompt update. Unsubscribe anytime.
Written by the Narracomm team
Narracomm is a communications and content strategy team that helps business owners, operators, and founders use AI to produce clear, credible, high-performing work. We build and test these prompt systems inside real client hiring processes — scorecards, structured interviews and debriefs — and revise them as models and regulation change. [Add specific credentials, hiring or HR experience, number of hires made or advised on, and a named reviewer — ideally someone with employment law or HR qualifications — to strengthen E-E-A-T.]
Sources & further reading #
NYC Department of Consumer and Worker Protection — Automated Employment Decision Tools (Local Law 144)DLA Piper — Audit of NYC’s AI hiring law signals increased employer risk (2026)EU AI Act — Annex III high-risk classification (employment & worker management)AI hiring compliance 2026: Local Law 144 and the EU AI ActWingate et al. — Interview criterion-related validity meta-analysis,International Journal of Selection and AssessmentMcDaniel et al. — The validity of employment interviews: a comprehensive review and meta-analysis- Schmidt & Hunter (1998); Sackett et al. (2022) — foundational meta-analyses on selection method validity.
Last reviewed and updated: July 25, 2026 · Regulatory references verified against current sources. Note: EU AI Act high-risk obligations become enforceable 2 August 2026 — this guide is scheduled for review immediately after that date. Next review due within 14 days.