# Consent is a Step — AgenticRail

> Markdown mirror for AI agents, generated 2026-10-26 from the live page.
> Canonical: https://agenticrail.nz/spec/consent-is-a-step/
> Site context: https://agenticrail.nz/llms.txt

Spec

# Consent is a Step

Every standard that governs what an AI system does has a slot for the person as a subject. None has a slot for consent as a step.

Published 26 October 2026 · Last reviewed 26 October 2026 · Updated as primary documents are read. This page is published by AgenticRail, which runs a sequence enforcement gate.

First stated — SHA-256 — v1.0

f6142ba928540d95e480e7b7ee0d6f2155864a510c8a880810b317aefc37ac50

 This hash is SHA-256 of the canonical string defined below. It is reproducible independently of this page using any SHA-256 implementation.

 **Canonical string (pipe-delimited, UTF-8, no trailing newline):**

 `Consent is a Step|1.0|2026-10-26|TUARA KURI LIMITED|consent is a step and the standards have no slot for it|the law carries the consent half as a precondition and nothing carries the order between them|the unknowability is the finding|the instrument that makes the question answerable is the record|alignment means into line alignment measured is sequence|no new law is needed the step already exists in law and AI processes must follow the law they already operate under|the regimes differ in their exceptions they are identical in their silence|agenticrail.nz/spec/consent-is-a-step`

 Published: 2026-10-26 | Version: 1.0 | Entity: TUARA KURI LIMITED

 **v1.0 (2026-10-26):** the floor claims as first published on this date. Amend, don't edit: a change to these sentences is a new version alongside this block, never a repair of it. The page around them remains a living document; these lines do not.

## 1. The first half, and the half nobody wrote

The AI standards each carry one half of the person. The half they carry is thorough: who is affected, what the risk to them is, and what happened to them, recorded. The half they defer is the one that matters first: whether they were asked.

**ISO/IEC 42005**, AI system impact assessment, calls for the assessment to identify affected individuals, groups and stakeholders — the people the system touches, including people who never chose to interact with it — and it does not require that anyone seek their consent. Stakeholder engagement is recommended where appropriate, never a requirement. The standard that exists to assess impact on people asks them nothing.

**ISO/IEC 23894**, guidance on AI risk management, lists privacy among its objectives in an annex, and leaves legal obligations to the organisation's context. Consent enters as a matter the organisation is expected to have handled under other law, not as a step in the risk process it governs.

**ISO/IEC 24970**, AI system logging, defines what a record of an AI system's events must contain. It records user identifiers "where privacy permits" — the log deliberately steps around the person — and it has human oversight interventions as a trigger, but no event type for consent given, refused or withdrawn. A system that never asked produces a complete, compliant log.

**ISO/IEC 42001**, the AI management system standard, and **ISO/IEC 27701**, the privacy information management standard, carry the closest thing the family has: privacy as a control to manage, and consent as a data-protection input under privacy law. Even there, consent is an input to a control, not a step a process must take before its next one.

The committees are named because the people who write these documents should be able to find this page: **ISO/IEC JTC 1/SC 42** writes the AI standards, **SC 27** writes the privacy ones, and **TC 215** writes health informatics. Consent does not sit in any one of them, and that is the point. It is a step, and none of their documents has a slot for steps.

## 2. What the law already says

The law has the consent half, and states it as a precondition. In New Zealand, Right 7(1) of the Code of Health and Disability Services Consumers' Rights provides:

"Services may be provided to a consumer only if that consumer makes an informed choice and gives informed consent, except where any enactment, or the common law, or any other provision of this Code provides otherwise."

The Privacy Act 2020 governs how personal information is collected, and its newest principle, **IPP 3A**, in force from 1 May 2026, is a notification duty for indirect collection — a notice after the fact, not consent before it. The Health Information Privacy Code applies the same structure to health information.

So the two halves exist in different places. The standards carry the first half — identify, assess, log. The law carries the consent half — a precondition to providing the service. Nothing carries the order between them. A precondition written in law is not a step enforced anywhere. That is the gap.

### The same step, worldwide

New Zealand is the anchor on this page because it is the jurisdiction where the full record is public: the requirement, the promise, the measured shortfall and the empty record. The step itself is written into the law of every major jurisdiction. The details differ — which lawful bases exist, opt-in versus opt-out, the exceptions, the penalties. The step does not.

| Jurisdiction | Where the step is written | What it requires |
| European Union | GDPR, Article 6(1)(a) with Article 7; health data as a special category under Article 9 | Consent must be freely given, specific, informed and unambiguous, and withdrawal must be as easy as giving it. Fines reach 4% of worldwide turnover. |
| Australia | Privacy Act 1988, APP 3.3: sensitive information, which includes health information, may be collected only with the individual's consent unless an exception applies; the My Health Records Act 2012 adds its own unauthorised-access offences | The regulator's guidance sets five elements of valid consent: voluntary, informed, current, specific, and given by someone with capacity. |
| United States | Sectoral, and the strongest sector is health: HIPAA permits use of health information for treatment, payment and operations, and requires the individual's written authorization for the rest, with civil and criminal penalties; state laws add opt-in consent for sensitive data, and the FTC enforces deception | The authorization must state what will be disclosed, to whom, and for what purpose, and the individual may revoke it in writing. |
| Canada, Brazil, China, Singapore, India, Japan | PIPEDA's meaningful-consent principle, LGPD Article 7, PIPL Article 13, the PDPA's consent obligation, the DPDP Act, APPI — each makes consent a precondition, with its own lawful bases and exceptions | Same shape, six spellings. |
| Before the web | OECD Privacy Guidelines (collection limitation, 1980) and the Council of Europe's Convention 108 (1981) | The order rule was world law before the technology it now governs. |

In every one of these, consent is a precondition with force of law. In none of them does anything record, at the step, whether it happened. The regimes differ in their exceptions; they are identical in their silence.

## 3. The live case: scribes in every emergency department

Health New Zealand | Te Whatu Ora is rolling out the AI scribe Heidi nationally: 1,000 clinician licences plus 100 for mental health crisis teams, in every emergency department in the country by April 2026, with the stated goal of use in every hospital by the end of 2026.

**The regulator requires the step.** The Medical Council of New Zealand's guidance of 10 March 2026, *Guidance on using artificial intelligence (AI) in patient care*, covers AI scribing tools and requires informed consent when an AI scribe or transcription tool records a consultation. It also requires that the consent, and the patient's decision when they decline, be documented. The step, and the record of the step, are both required by the profession's own regulator.

**The agency promised the step.** When Health NZ endorsed Heidi and iMedX on 20 July 2025, the chair of its National AI and Algorithm Expert Advisory Group stated: "Patients will be asked for consent when clinicians are using AI during their consultations."

**The measured practice falls short.** A peer-reviewed survey of New Zealand AI-scribe users found only about 60% sought the patient's consent. An earlier GP survey put it under two-thirds. The requirement is stated twice over, and roughly two patients in five are never asked.

**The record is empty.** Health NZ's official disclosure on Heidi, an OIA release published 28 April 2026, covers governance, risk management, privacy safeguards and early evaluation findings — as a 13-megabyte scan with no searchable text. The independent privacy impact assessment records that audio is not retained, New Zealand patient data is not used for model training, the provider is ISO 27001 certified, and processing runs in Sydney with New Zealand storage planned. The HRC-funded Roadmap for AI Scribes in Aotearoa, run by the University of Auckland with Health New Zealand, still lists patient consent among the unresolved risk areas being studied. And a clinical lead at Whakarongorau has said a complaint to the Health and Disability Commissioner over scribe use without informed consent is "only a matter of time".

The step is required by the regulator, promised by the agency, delivered by three clinicians in five, and recorded by no one. A patient cannot be asked a question that leaves no record of the answer, and no record means no one can show the law was followed — and no one can show it was broken, either. The survey is the best evidence that exists, and the best evidence that exists is a self-reported guess. Both claims die in the same silence, and the only party that silence serves is the one that never has to answer. The unknowability is the finding. Right 7 has not been repealed. It has been left outside the sequence.

The missing record is not an evidentiary gap. It is the first breach. The Council's guidance requires the consent *and its documentation* (cl. 9–10); Right 4(2) requires services to comply with professional standards; and the GDPR states the principle in so many words — the controller "shall be responsible for, and be able to demonstrate compliance with" its obligations (Article 5(2)). A system that keeps no record has already failed the requirement that exists precisely so the rest could be shown.

And because no record exists, the defence is unavailable in every case — the consultation where consent was asked, and the one where it was not. The clinic cannot clear a single one. A complaint from any patient, on any day, is met with nothing to show. So every case is exposed: not "which cases were breaches" but "no case can be shown not to be". Whatever the true rate was, even if every clinician asked, the burden now rests on the whole population, case by case, with nothing to answer it.

The instrument that makes each case answerable is the record, held by a party that is not the clinic. Each encounter gets its own signed step: consent, recorded, refused when it is missing. Then each case stops being suspicious — individually, verifiably, forever.

And the documentation is worth only what attests it. A record of consent written by the clinic alone is the operator's word — self-attested, honest or not, and the two look identical from outside. The record that answers the question carries the patient's own mark: consent, signed by the person who gave it, recorded at the step, and held by a party that is neither patient nor clinic. The doer cannot witness itself; the patient is the witness of their own consent.

The requirement was written by the profession's own regulator, the promise by the deployer's own chair, the shortfall measured by a published survey, and the unresolved risk listed by the roadmap the system itself funds. Knowledge of the requirement, followed by continued conduct, is the shape negligence takes — and "nobody told us" is foreclosed by their own documents. The requirement they wrote is the ghost: every missing record is haunted by their own words — provide documentation. The silence they kept is the last one.

## 4. Consent as a step

The industry already has the word. It calls its central problem [*alignment*](https://agenticrail.nz/sequence-verification/#alignment). The word's literal sense is into line. A declared sequence is a line: the steps, in the order the law and the promise set out. A step that runs out of order is out of line, out of alignment in the measurable sense, whatever the model wanted internally. Consent skipped is the line broken, and the incidents the industry reports are the same shape: actions out of line with what was declared, found later because nothing recorded the line. Alignment, measured, is sequence.

Consent belongs in the sequence a process declares in advance: intake, then consent, then record, then proceed. Consent is the step that must precede the recording of the person. Its absence stops the process at that point, before anything else runs. Its refusal is itself the record: signed by the patient, and held by a party that is not the operator, so a process that skipped consent shows as a gap rather than as a completed run. Then the number is not believed — it is known. Every consent is a signed step, every refusal a signed refusal, every missing step a visible gap in a chain a third party holds. The instrument that makes the question answerable is the record.

No new law is needed, and nothing needs to wait on a committee. The step already exists in law, and AI processes must follow the law they already operate under. That is the way in: the sequence declares the step the law already requires, and the gate holds the order, today.

This is what [sequence verification](https://agenticrail.nz/sequence-verification/) does: each step is checked against an order declared in advance, before it runs, and the decision is signed and held by a party that is not the operator. An [order failure](https://agenticrail.nz/sequence-verification/#order-failure) is the name for what happens when the step runs anyway.

Consent belongs in every one of the standards above, because consent is not a clause. It is a step that every process touching a person must take before its next one. Risk management needs it: the risk is to a person who never agreed to be in the process. Impact assessment needs it: the impact lands on someone who was never asked. Logging needs it: the one entry that matters most is whether they said yes, and none of their information models has a field for it.

The committees can keep their first halves. They are good. The consent half needs a slot, and the slot is the declared sequence.

The same mechanism runs wherever the law writes an order rule: [customer due diligence](https://agenticrail.nz/spec/customer-due-diligence/), [supplier payments](https://agenticrail.nz/spec/supplier-payments/), the [health brief](https://agenticrail.nz/spec/nz-health/). Consent is the floor; the gate is the whole building.

The agent is the instrument; the vendor is the actor; the order is the duty. An agent cannot give consent it was never asked for, and a vendor cannot hand the law to a machine. The UK's four regulators state the principle plainly in their foresight paper on agentic AI: organisational responsibility for legal compliance is not diluted by autonomy. What remains is the only defence that works: the order — declared in advance, enforced at the step, and recorded by a party that is not the operator.

Limits

- The standard texts themselves are paywalled. This page states their publicly described provisions, with the source of each description linked below. It is updated as primary documents are read.
- No claim is made that any standard requires consent as a step. The claim is that none provides one, while the law requires consent as a precondition.
- No claim is made about any health provider's individual practices beyond what is cited.
- A sequence enforcement gate governs the steps a deployment sends through it. It does not stop activity on systems outside that deployment.
- This page is not legal advice.

## Questions

### Do the AI standards require consent?

None of them requires consent as a step. ISO/IEC 42005 calls for affected individuals to be identified, including people who never chose to interact with the system, and does not require that anyone ask them. ISO/IEC 23894 lists privacy as one objective and leaves legal obligations to the organisation's context. ISO/IEC 24970 has no event type for consent given or refused.

### Is consent required by law in New Zealand health?

Yes, as a precondition, and the profession's regulator requires the step. Right 7(1) of the Code of Health and Disability Services Consumers' Rights provides that services may be provided to a consumer only if that consumer makes an informed choice and gives informed consent, except where an enactment, the common law or another provision of the Code provides otherwise. The Medical Council's guidance of 10 March 2026 requires informed consent when an AI scribe records a consultation, and requires the consent and any decline to be documented. The Privacy Act 2020 and the Health Information Privacy Code govern the collection and handling of health information.

### Is consent required worldwide?

The step is written into every major jurisdiction — GDPR, the Australian Privacy Principles, HIPAA, PIPEDA, LGPD, PIPL, the PDPA, the DPDP Act — each with its own lawful bases and exceptions, and none with a record of the step. The international instruments predate the web. The regimes differ in their exceptions; they are identical in their silence.

### What would consent as a step look like?

Consent is declared in the sequence a process must follow: intake, then consent, then record, then proceed. The next step is refused until consent is recorded. A refusal is signed and held by a party that is not the operator, so a process that skipped consent shows as a gap rather than as a completed run. No new law is needed: the step already exists in law, and sequence verification is how a deployment holds the order the law already requires.

## 5. Updates

**26 October 2026:** first published.

**26 October 2026, later:** section 3 adds the Medical Council's guidance of 10 March 2026 requiring informed consent for AI scribes and its documentation, Health NZ's stated position of 20 July 2025, the measured consent rates from the peer-reviewed survey, and Health NZ's OIA release of 28 April 2026.

**26 October 2026, later still:** sections 3 and 4 add the point the record makes in both directions: with no record, no one can show the law was followed, and no one can show it was broken — the unknowability is the finding, and the instrument that makes the question answerable is the record.

**26 October 2026, last:** section 4 adds the literal sense of alignment — into line — and the declared sequence as the measurable version of it: consent skipped is the line broken.

**26 October 2026, later again:** section 2 gains the worldwide table — the same step in the GDPR, the Australian Privacy Principles, HIPAA, PIPEDA, LGPD, PIPL, the PDPA and the DPDP Act, and in the instruments that predate the web — and the page gains the questions section to match its structured data.

**26 October 2026, last of the day:** section 4 closes with the funnel: consent is the floor; the gate is the whole building, with customer due diligence, supplier payments and the health brief named alongside.

**26 October 2026, close of day:** section 4 closes with the attribution principle — the agent is the instrument, the vendor is the actor, the order is the duty — with the UK DRCF foresight paper cited.

**26 October 2026, night:** section 3 gains the four-move legal spine: the documentation duty is the first breach (the Council's cl. 9–10, Right 4(2), GDPR Article 5(2)); the missing record makes the defence unavailable in every case; therefore every case is exposed, whatever the true rate was; and the record held by a party that is not the clinic makes each case answerable.

**26 October 2026, night, last:** the spine gains its fifth move: the documentation must carry the patient's own signature — a record written by the clinic alone is self-attested, and the patient is the witness of their own consent.

**26 October 2026, night, final:** the spine gains its sixth move: the requirement was written by their own regulator, the promise by their own chair, the shortfall by their own measures — and carrying on after that is the shape negligence takes.

**26 October 2026, night, last of all:** the sixth move closes with the ghost: the requirement they wrote haunts every missing record — provide documentation.

## 6. Primary sources

Sources

ISO/IEC 42005:2025, Information technology — Artificial intelligence — AI system impact assessment, catalogue page and abstract — [iso.org](https://www.iso.org/standard/42005)

ISO/IEC 23894:2023, Information technology — Artificial intelligence — Guidance on risk management, catalogue page — [iso.org](https://www.iso.org/standard/23894)

ISO/IEC 24970, Artificial intelligence — AI system logging (draft), publicly summarised clause list including the user-identifier provision — [vde.com](https://www.vde.com/topics-en/artificial-intelligence/blog/eu-ai-act--ai-system-logging)

ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system, catalogue page — [iso.org](https://www.iso.org/standard/42001)

ISO/IEC 27701:2019, Security techniques — Extension to ISO/IEC 27001 and ISO/IEC 27002 for privacy information management, catalogue page — [iso.org](https://www.iso.org/standard/27701)

Code of Health and Disability Services Consumers' Rights, Right 7 — [hdc.org.nz](https://www.hdc.org.nz/your-rights/the-code-and-your-rights/)

Regulation (EU) 2016/679 (GDPR), Articles 5(2), 6, 7 and 9 — [eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2016/679/oj)

Privacy Act 1988 (Cth), Australian Privacy Principle 3.3, and the Office of the Australian Information Commissioner's consent guidance — [legislation.gov.au](https://www.legislation.gov.au/C2004A03712/latest/text) · [oaic.gov.au](https://www.oaic.gov.au/privacy/your-privacy-rights/consent)

Health Insurance Portability and Accountability Act, Privacy Rule — written authorization, 45 CFR 164.508 — [ecfr.gov](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.508)

Personal Information Protection and Electronic Documents Act (Canada), consent principle — [priv.gc.ca](https://www.priv.gc.ca/en/privacy-topics/privacy-laws-in-canada/the-personal-information-protection-and-electronic-documents-act-pipeda/)

OECD Privacy Guidelines (collection limitation, 1980; revised 2013) — [oecd.org](https://www.oecd.org/en/topics/privacy.html)

Digital Regulation Cooperation Forum (CMA, FCA, ICO, Ofcom), The Future of Agentic AI: Foresight Paper, 31 March 2026 — [drcf.org.uk](https://www.drcf.org.uk/siteassets/drcf/pdf-files/drcf-the-future-of-agentic-ai-foresight-paper.pdf)

Privacy Act 2020, information privacy principles including IPP 3A (indirect collection), in force 1 May 2026 — [legislation.govt.nz](https://www.legislation.govt.nz/act/public/2020/0031/latest/whole.html)

Health Information Privacy Code 2020 — [privacy.org.nz](https://www.privacy.org.nz/privacy-act-2020/codes-of-practice/hipc2020/)

Medical Council of New Zealand, Guidance on using artificial intelligence (AI) in patient care, 10 March 2026 — [mcnz.org.nz](https://www.mcnz.org.nz/our-standards/current-standards/guidance-on-using-artificial-intelligence-in-patient-care/)

Robyn Whittaker, chair of the National AI and Algorithm Expert Advisory Group, on Health NZ's endorsement of Heidi and iMedX, 20 July 2025 — [hinz.org.nz](https://www.hinz.org.nz/news/news.asp?id=706270)

Using AI scribes in New Zealand primary care consultations: an exploratory survey, peer-reviewed — [publish.csiro.au](https://www.publish.csiro.au/HC/HC25079)

Health New Zealand | Te Whatu Ora, Information Relating to the Use and Evaluation of Heidi AI, OIA HNZ00106690, published 28 April 2026 — [healthnz.govt.nz](https://www.healthnz.govt.nz/publications/information-relating-to-the-use-and-evaluation-of-heidi-ai-oia)

Independent Heidi Health Privacy Impact Assessment, commissioned by Collaborative Aotearoa — [nbph.org.nz](https://www.nbph.org.nz/s/Independent_Completion_-_Heidi_Health_AI_Privacy_Impact_Assessment_v002.pdf)

NZ Doctor, HDC complaint over AI scribes "only a matter of time", June 2026 — [nzdoctor.co.nz](https://www.nzdoctor.co.nz/article/news/hdc-complaint-over-ai-scribes-only-matter-time)

Health NZ's Heidi scribe rollout: 1,000 clinician licences plus 100 for mental health, every emergency department by April 2026, all hospitals by end of 2026 — [Healthcare IT News](https://www.healthcareitnews.com/news/anz/public-emergency-clinicians-begging-try-ai-scribes) · [NZ Herald](https://www.nzherald.co.nz/business/health-nz-to-roll-out-heidi-ai-scribe-in-16-emergency-departments-but-whats-the-story-with-errors-and-privacy/premium/RIYTO2QFMBBX5MJKEDXZUDDGDY/)

Roadmap for AI Scribes in Aotearoa, University of Auckland with Health New Zealand, listing patient consent among unresolved risk areas — [hinz.org.nz](https://www.hinz.org.nz/news/729179/)