Official Reply Reference: 20260457 Official Information Act Request MSD Use of AI / Automated Decision-Making and Human Accountability

Reference: 20260457
17 June 2026
Anne
[email protected]

Dear Anne
Thank you for your request made under the Official Information Act, received on 1 June
2026.

 You requested:
This request concerns the Ministry of Social Development’s use, proposed use,
approval, funding, governance, and accountability arrangements for artificial
intelligence, automated decision-making, algorithmic tools, rules engines, data
matching, risk flags, system-generated recommendations, automated notices,
and partly automated processes in welfare administration.

For clarity, this request includes any automated or partly automated process used
to make, assist, recommend, trigger, review, issue, suspend, cancel, delay,
correct, replace, or withdraw a decision affecting a person’s entitlement,
payment, debt, deduction, review right, obligation, or access to MSD-
administered support.

The purpose of this request is to understand:
1. why MSD is using or proposing to use AI or automated decision-making;
2. what problem MSD says AI or automation is intended to solve;
3. whether the purpose is to improve entitlement, access, accuracy, and support,
or primarily to reduce cost, staff workload, debt, fraud, payment levels, or
compliance risk;
4. what decisions may be affected;
5. what safeguards exist before adverse decisions are made;
6. who is the human decision-maker accountable for any decision involving
automation;
7. whether people affected by automated or system-assisted decisions are told
that automation was involved;
8. how errors, underpayments, wrongful suspensions, wrongful cancellations, and
wrongful debts are detected and corrected.

Automated handling of this OIA request
If this OIA request is processed, triaged, summarised, classified, searched,
routed, prioritised, or answered using any automated system, artificial intelligence
tool, data-matching process, workflow engine, template response system, search
tool, or system-generated recommendation, please identify that process.
2
Please confirm that the final response has been reviewed and approved by a
human officer with lawful delegated authority.
Please preserve any audit logs, search terms, classification notes, automated
summaries, routing notes, internal prompts, system-generated recommendations,
decision records, and human review records created in handling this request.
This request does not object to efficient administrative processing. The issue is
accountability. Any response affecting public rights, records, entitlements, legal
duties, refusal grounds, transfers, extensions, or statutory interpretation must be
traceable to a human officer or office with lawful authority.

1. Purpose and justification for AI / automated decision-making

Please provide all documents explaining why MSD is using, proposing to use,
approving, funding, expanding, or seeking legislative authority for artificial
intelligence, automated decision-making, algorithmic tools, rules engines, data
matching, risk flags, system-generated recommendations, automated notices, or
partly automated welfare decision processes.

This includes all documents identifying:

1. the problem AI or automation is intended to solve;
2. whether the objective is faster payment, improved entitlement, reduced
underpayment, improved client access, reduced hardship, reduced fraud,
reduced debt, reduced staff workload, cost saving, compliance enforcement,
service modernisation, or another purpose;
3. any expected financial savings;
4. any expected staffing impacts;
5. any expected increase or decrease in payment approvals, declines,
suspensions, cancellations, debts, deductions, reviews, appeals, or back-
payments;
6. any documents comparing automated decision-making with human decision-
making;
7. any documents considering whether automation may make existing MSD
errors faster, larger, or more difficult for clients to challenge;
8. any documents considering whether automation may disadvantage elderly,
disabled, medically unwell, digitally excluded, cognitively impaired, language-
limited, homeless, or otherwise vulnerable people.

2. Systems proposed, approved, piloted, used, paused, or withdrawn

Please provide the name, purpose, and description of each AI, automated, partly
automated, algorithmic, rules-based, data-matching, risk-flagging, eligibility, debt,
compliance, client-segmentation, fraud, payment, review, or system-generated
decision process currently used, proposed, piloted, approved, paused,
withdrawn, replaced, or under development by MSD.
For each system or process, please identify:
3
1. the system name;
2. the current status: proposed, piloted, approved, used, paused, withdrawn,
replaced, or under review;
3. the date it was proposed, approved, implemented, paused, replaced, or
withdrawn;
4. the decision type affected;
5. the benefit, pension, supplement, hardship grant, debt, deduction, or payment
type affected;
6. whether the system makes the final decision, recommends a decision, triggers
staff review, generates correspondence, or flags the client;
7. whether the system can affect payment, non-payment, suspension,
cancellation, debt, deduction, repayment, review, obligation status, or entitlement;
8. the data sources used;
9. the human officer, Business Owner, system owner, statutory decision-maker,
or delegated authority accountable for the decision;
10. the review, appeal, correction, and human reconsideration pathway available
to the client.

3. Ministerial, Cabinet, Treasury, and senior executive approval

Please provide all ministerial briefings, Cabinet papers, Cabinet committee
papers, aide-memoires, Treasury advice, budget documents, reports, decision
papers, implementation plans, approval papers, risk assessments, cost-benefit
analyses, legal-risk summaries, privacy assessments, human-rights
assessments, ethics assessments, procurement documents, and policy advice
concerning AI or automated decision-making in MSD.

This includes documents concerning:
1. approval of automated decision-making powers;
2. decisions to introduce or expand automation;
3. funding or savings connected to automation;
4. reasons for legislative change;
5. whether urgency was used or considered;
6. whether select committee scrutiny was bypassed, shortened, or limited;
7. legal risks;
8. privacy risks;
9. human rights risks;
10. discrimination risks;
11. administrative-law risks;
12. risk of unlawful denial, delay, suspension, cancellation, underpayment,
overpayment, or debt;
13. safeguards for vulnerable clients;
14. human review;
15. public transparency;
16. client notification;
17. lessons from overseas welfare automation failures, including Australia’s

Robo debt scheme.
4. Positive-use design: entitlement, support, and underpayment detection
4
Please provide all documents showing whether MSD’s AI or automated decision-
making systems are designed to identify, increase, or protect:

1. lawful entitlement;
2. underpayments;
3. missed entitlements;
4. hardship eligibility;
5. disability-related needs;
6. medical vulnerability;
7. pension eligibility;
8. supplementary assistance;
9. emergency assistance;
10. client support needs;
11. situations where a person should be paid first and reviewed later;
12. situations where delay, suspension, cancellation, or debt would cause
hardship or disproportionate harm.

Please also provide all documents comparing positive client-support uses of
automation with compliance, debt, fraud, suspension, cancellation, deduction, or
non-payment uses.

5. Human accountability for automated or system-assisted decisions
Please provide all policies, standards, delegations, manuals, training material,
governance documents, and internal guidance explaining how MSD identifies the
human person or office accountable where an automated or system-assisted
process is involved.

Please include documents explaining:

1. who is legally responsible for an automated or system-assisted decision;
2. how the accountable person is recorded;
3. whether the client is told who is accountable;
4. whether a human officer must review the decision before adverse action
occurs;
5. whether a client can require human reconsideration;
6. how MSD prevents the same automated error being repeated during review;
7. what happens if MSD cannot identify the human decision-maker;
8. whether the Minister, Chief Executive, Business Owner, system owner,
delegated officer, or frontline staff member is accountable for harm caused by
automated decision-making;
9. whether accountability differs where the automated system only recommends,
flags, routes, scores, prioritises, drafts, or generates correspondence rather than
making a final decision.

6. Client notice, reasons, and explanation

Please provide all templates, scripts, letters, MyMSD messages, text messages,
emails, call-centre guidance, operational instructions, and public-facing wording
5
used to tell clients:

1. that AI, automation, data matching, risk flags, templates, or system-generated
recommendations were involved;
2. the reason for the decision;
3. the evidence relied on;
4. the data source relied on;
5. the human officer or office accountable;
6. the review or appeal right;
7. how to correct wrong information;
8. how to request human review before payment is stopped, delayed, reduced,
cancelled, converted into debt, or otherwise adversely affected.
7. Data sources and data matching

Please provide all documents identifying the data sources used, proposed, or
considered for AI or automated welfare decision-making, including:

1. Inland Revenue data;
2. ACC data;
3. Immigration New Zealand data;
4. Department of Internal Affairs / identity / RealMe data;
5. Corrections data;
6. health or disability-related data;
7. banking or payment data;
8. employer data;
9. tenancy, housing, or accommodation data;
10. historic MSD records;
11. debt records;
12. compliance records;
13. fraud or investigation records;
14. MyMSD data;
15. third-party supplied information.

For each data source, please provide documents explaining:

1. the lawful authority for use;
2. the data-sharing agreement or memorandum of understanding;
3. data-quality checks;
4. error-correction process;
5. client-notification process;
6. human-review process;
7. audit process;
8. safeguards against wrongful adverse decisions.
8. Adverse decisions and safeguards

Please provide all policies, risk assessments, staff guidance, and training
documents explaining safeguards before AI, automation, data matching, system
flags, or system-generated decisions affect clients by:
6
1. delaying payment;
2. refusing payment;
3. reducing payment;
4. suspending payment;
5. cancelling payment;
6. creating debt;
7. recovering debt;
8. imposing deductions;
9. changing review rights;
10. changing obligations;
11. issuing compliance, fraud, or investigation flags.

Please include safeguards for clients who may be:

1. elderly;
2. disabled;
3. medically unwell;
4. cognitively impaired;
5. digitally excluded;
6. language-limited;
7. homeless or without stable postal address;
8. reliant on a caregiver, advocate, attorney, representative, or support person;
9. at risk of hardship if payment is delayed, reduced, suspended, or cancelled.
9. Error detection, audits, correction, and compensation

Please provide all audits, reviews, briefings, quality-assurance reports, risk
assessments, dashboards, statistics, or internal reports from the last five years
concerning:

1. automated or system-generated errors;
2. incorrect entitlement decisions;
3. incorrect suspensions or cancellations;
4. incorrect debt decisions;
5. underpayments;
6. payment delays caused by AI, automation, forms, data matching, or system
flags;
7. decisions withdrawn after issue;
8. decisions corrected after review;
9. back-payments resulting from incorrect MSD decisions;
10. failures to provide proper reasons;
11. failures to identify the accountable human decision-maker;
12. systemic risks created by automation;
13. whether automation reduces or increases error rates;
14. whether compensation, apology, reinstatement, or back-payment is
considered where an automated or system-assisted process causes harm.
10. Metrics and monitoring

7
Please provide all documents showing how MSD measures whether AI or
automated decision-making is successful.
Please include all measures, key performance indicators, dashboards, reports, or
evaluations concerning:

1. speed of decision-making;
2. payment accuracy;
3. entitlement accuracy;
4. underpayment detection;
5. overpayment detection;
6. client hardship;
7. client complaints;
8. Reviews of Decision;
9. appeals;
10. overturned decisions;
11. back-payments;
12. suspensions;
13. cancellations;
14. debt creation;
15. staff workload;
16. financial savings;
17. fraud detection;
18. vulnerable-client outcomes;
19. whether the system improves access to support or mainly improves
compliance enforcement.

11. Procurement, vendors, and external providers
Please provide all documents identifying any external vendors, contractors,
consultants, software providers, cloud providers, analytics providers, or advisory
firms involved in MSD AI, automated decision-making, algorithmic tools, data
matching, risk scoring, workflow automation, or system-generated
correspondence.

Please include:
1. contracts;
2. procurement documents;
3. statements of work;
4. privacy/security assessments;
5. risk assessments;
6. model documentation;
7. testing documents;
8. implementation plans;
9. vendor accountability documents;
10. documents explaining ownership and control of client data.
12. Format, refusal, transfer, and refinement