Composite case study · Local government · Aotearoa New Zealand

Where AI could fit
in a council’s work.

A pipeline-generated map of 12 role archetypes in a composite New Zealand district council—229 generated tasks and 6,101 scored steps, assessed under a specified 2026 language-model deployment profile.

12 role archetypes
229 generated tasks
6,101 generated and scored steps
55.8% of step categories AI can carry
6.3% AI carries without routine approval

Percentages describe generated step categories, not time, workload, transaction volume, staffing or full-time equivalents.

On this page

What was modelled, the technology assumptions, headline results, role-level distributions, a planner drill-in and the limitations that matter.

  1. Scope and construction
  2. Technology profile
  3. Headline result
  4. Role-by-role results
  5. Planner drill-in
  6. Why people remain required
  7. What the analysis suggests
  8. How to read it
  9. Full role-level data

§ 01 · Scope

A model council,
not a real council audit.

The organisation is a composite representation of a mid-sized New Zealand district council, constructed to demonstrate the methodology without exposing the data of a real organisation.

The twelve role archetypes were selected to span finance, regulation, strategy, governance, customer services, people, libraries and infrastructure operations. They are not an exhaustive council organisation chart and exclude executive leadership and several specialist functions.

The role maps were produced from research-based descriptions and organisational context. The production pipeline generated 229 tasks, decomposed those tasks into 6,101 steps, assessed each step against the ten-dimension metric registry, and applied the default 2026H1 language-model profile.

The pipeline outputs shown here were not manually edited. That is useful for demonstrating the end-to-end system, but it is not proof that every generated task or step is correct. In a client engagement, practitioners would review the provisional workflow map before it became a decision basis.

Included

Pipeline integrity

Tasks, steps, metric scores, attributes, dispositions and traces were produced through the versioned production process.

Not included

Observed work measurement

The analysis does not use time-and-motion observation, workflow logs, transaction volumes or task-frequency data.

Not included

Practitioner correction

The public demonstration has not been rewritten to reflect practitioner corrections. External validation is being developed separately.

Not included

Technical performance testing

No AI system was deployed inside a council workflow to test actual accuracy, reliability, cost or employee acceptance.

§ 02 · Profile

The result depends on
the system being assessed.

This is not a timeless estimate of what “AI” can do. It is the output of one versioned technology and policy profile.

A separately defined integrated profile assumes stronger access to archival records and continuously supplied system information. It changes some dispositions without changing the underlying work map. That comparison can help identify where integration investment may increase useful autonomy.

§ 03 · Headline result

Oversight dominates
the generated map.

Under the default profile, 55.8% of generated step categories are classified as work the AI can carry—but almost nine in ten of those still require human approval before the output takes effect.

AI carries AI carries, human approves Human leads, AI assists Human required Beyond profile
6.3%

AI carries

383 generated steps

49.5%

AI carries, human approves

3,020 generated steps

17.1%

Human leads, AI assists

1,042 generated steps

14.1%

Human required

862 generated steps

13.0%

Beyond this profile

794 generated steps

01

Unreviewed AI carriage is a small minority

Only 6.3% of all generated steps are classified as AI carrying the work without routine per-instance approval.

02

Human approval does most of the work

The largest category is AI carrying the primary work with a person reviewing it before effect.

03

Physical scope remains substantial

Thirteen percent of generated steps lie beyond the current profile, concentrated in infrastructure, parks and library work.

§ 04 · Role by role

Twelve roles,
twelve different mixes.

Roles are sorted by the share of generated steps the AI can carry. The complete distribution matters more than the ranking.

AI carries AI carries, human approves Human leads, AI assists Human required Beyond profile
council-demo-01 · role-level generated-step counts and distributions
Role and department Distribution AI carries + approval AI assists Human required Beyond profile Total Carry share
Accounts Payable Officer Finance 36303 614 0359 94%
Resource Consents Planner Regulatory Services 31379 4248 2502 82%
Building Consents Officer Regulatory Services 18204 1950 16307 72%
Communications Advisor Strategy and Policy 61348 12284 0615 67%
Customer Services Officer Customer and Community 51182 9434 18379 61%
Policy Advisor Strategy and Policy 22441 178141 5787 59%
Democracy Advisor Governance 25239 12566 4459 58%
Customer Services Team Leader Customer and Community 47234 14678 11516 54%
HR Advisor People and Capability 23333 161143 1661 54%
Community Librarian Customer and Community 5595 7249 181452 33%
Water Services Operator Infrastructure and Operations 0178 61133 260632 28%
Parks and Open Spaces Worker Infrastructure and Operations 1484 1622 296432 23%
All role maps 383 3,020 1,042 862 794 6,101 55.8%

Counts are not weighted by illustrative FTE, task frequency, duration or transaction volume. “Carry share” is AI carries plus AI carries with human approval, divided by all generated steps for the role.

§ 05 · Role drill-in

One planner role,
seventeen generated tasks.

The Resource Consents Planner illustrates why the task-level mix is more informative than one role-level percentage.

The tasks below are presented in an indicative consent-processing sequence. They are pipeline-generated representations, not a claim that every council or every application follows one linear path.

AI carries AI carries, human approves Human leads, AI assists Human required Beyond profile
01

Check a lodged application for statutory completeness

96%AI can carry
02

Set up the statutory processing record

100%AI can carry
03

Determine consent triggers and activity status

96%AI can carry
04

Provide pre-application planning advice

91%AI can carry
05

Coordinate specialist input on an application

83%AI can carry
06

Request and evaluate further information under section 92

95%AI can carry
07

Assess the environmental effects of a proposal

57%AI can carry
08

Negotiate proposal amendments with the applicant

75%AI can carry
09

Prepare a notification recommendation

82%AI can carry
10

Draft resource consent conditions

87%AI can carry
11

Prepare a resource consent recommendation report

85%AI can carry
12

Prepare a section 42A hearing report

72%AI can carry
13

Assess submissions on a notified application

79%AI can carry
14

Present planning evidence at a hearing

36%AI can carry
15

Issue the resource consent decision package

97%AI can carry
16

Monitor compliance with granted consent conditions

86%AI can carry
17

Respond to consent-processing queries

74%AI can carry
Where AI carriage concentrates

Document and rules work

Completeness checks, processing records, trigger assessments, draft conditions and decision-package assembly are classified heavily toward AI carriage with human approval.

Where human leadership concentrates

Judgment and live professional interaction

Environmental-effects assessment and hearing work retain more human leadership because the planner must weigh competing considerations, explain reasoning and respond in real time.

§ 06 · Human authority

The same disposition,
for different reasons.

“Human required” may result from grave consequences, value-laden judgment, relationship demands or the need for legitimate human decision authority.

Human required because the stakes are grave

Building Consents Officer

Assess fire safety, including escape routes, alarms and fire separations.

Consequence: high Value judgment: low

The assessment is substantially rule-bound, but the consequences of error may be severe. The default policy retains human authority even where a system can perform much of the checking.

Human required because the judgment is the work

Resource Consents Planner

Weigh how much a proposed building would affect each neighbour.

Consequence: high Value judgment: high

Reasonable professionals may weigh the same facts differently. The system may prepare the evidence, but a person forms and owns the final assessment.

§ 07 · Interpretation

What this demonstration
suggests.

The strongest signal is not wholesale automation. It is the prevalence of workflows in which AI performs substantial digital work while people retain review, judgment or live responsibility.

Document-heavy, structured and rule-bound steps appear most often in the AI-carriage categories. Human leadership becomes more prominent where work depends on contested judgment, interpersonal adaptation, live situational awareness or accountability for grave consequences.

Physical work creates a separate limitation. Large parts of parks, water operations and community-library work are beyond this language-model profile rather than merely low-scoring.

The profile comparison also identifies potential integration value. Under a separately modelled integrated profile—assuming stronger archival retrieval and continuously supplied system information— some steps gain autonomy because required context becomes available to the system. This is an architectural change, not a claim that the underlying model has become more intelligent.

For AI investment

Prioritise recurring digital workflows where the map shows substantial carriage potential and where required information, tools and review controls can realistically be supplied.

For workforce planning

Focus on how roles may be redesigned around review, exception handling, judgment and relationships rather than converting step shares directly into staffing estimates.

For governance

Distinguish technical capability from acceptable authority. High-consequence or value-laden decisions may retain human control even where AI can perform much of the analytical preparation.

§ 08 · Reading the results

What the numbers
do not establish.

01

This is not a real council audit

The organisation and its FTE assumptions are composite. Results should not be attributed to any actual council.

02

Generated steps are not workload measures

Counts are not weighted by task frequency, duration, transaction volume or FTE. They cannot be converted directly into time or staffing effects.

03

The workflow map remains provisional

The public demonstration has not been corrected through a complete practitioner-validation process. Some tasks or steps may be missing, misplaced or expressed at the wrong grain.

04

The profile is not a performance benchmark

The capability envelope represents evidence-informed judgments about a technology class. It does not prove that a specific product will meet a council’s required accuracy or reliability.

05

Policy thresholds remain organisational choices

Human-review and human-authority requirements reflect the stated default policy. A real council may reasonably adopt different settings, subject to law and professional obligations.

06

Implementation feasibility is not included

Integration, data access, security, privacy, testing, assurance, procurement, employee participation and change management require separate investigation.

§ 09 · Data note

Every role-level count
is published above.

The role table in § 04 contains the complete disposition counts used for the organisation-level result. The five arrangement columns sum to 6,101 generated steps.

“AI can carry” is calculated as:

AI carries + AI carries, human approves ÷ All generated steps = Carry share
AI carries
383
AI carries, human approves
3,020
Human leads, AI assists
1,042
Human required
862
Beyond profile
794
Total generated steps
6,101

Suggested citation: Deployed Intelligence Labs Limited, Where AI could fit in a council’s work: a task-by-task case study, council-demo-01, 2026.