# How Should Permit Offices Tier AI Risks in 2026?

urbanplanadvisor.com · September 28, 2026

> Direct Answer: Permit AI Risk Tiers Should Be Capability-Based, Not Model-Branded A permit AI risk tier is a control framework that assigns planning or...

## Direct Answer: Permit AI Risk Tiers Should Be Capability-Based, Not Model-Branded

A permit AI risk tier is a control framework that assigns planning or building applications to different levels of review according to the consequences of error, the reversibility of decisions, data sensitivity, and the degree of human or institutional authority involved. It should not simply be based on whether an application uses a model called a small, general, frontier, or “top-tier” system. By September 2026, a defensible permit-office framework would normally use four operational tiers: Tier 0 for deterministic tools, Tier 1 for low-impact assistive systems, Tier 2 for consequential recommendation systems, and Tier 3 for autonomous or agentic systems able to initiate, approve, or execute actions. Model names and vendor claims may inform technical assessment, but they do not determine legal risk by themselves.

**Also worth reading:** [How Do Cities Use AI for Permit Review Without Creating New Legal Risks?](https://urbanplanadvisor.com/knowledge/how_do_cities_use_ai_for_permit_review_without_creating_new_legal_risks.php) · [How Should Local Governments Govern AI Used in Permit Decisions?](https://urbanplanadvisor.com/knowledge/how_should_local_governments_govern_ai_used_in_permit_decisions.php) · [How Should Cities Permit AI Governance in Planning and Public Decisions?](https://urbanplanadvisor.com/knowledge/how_should_cities_permit_ai_governance_in_planning_and_public_decisions.php)

The European Union AI Act provides a useful reference because it distinguishes prohibited practices, high-risk systems, and other AI obligations rather than creating an ordinary three-tier ranking of model capability. For local permitting, however, that legal classification is only the starting point. IBM’s argument that risk should determine AI architecture is directly relevant: a public-sector deployment should be designed around its ability to cause harm, interrupt an essential service, expose protected information, or produce decisions that are difficult to reverse. Permit offices should publish the tier definitions, required controls, appeal routes, and service-level commitments before allowing vendors to market an “approved” planning AI.

A strong system also recognizes that the same foundation model can create different risks in different workflows. Claude used for drafting a public-facing FAQ may remain Tier 1 if a planner verifies every answer, while the same model connected to a permit database, able to recommend approval conditions, and trained to pursue completion goals may be Tier 2 or Tier 3. Risk comes from the combined configuration of model, data, tool access, users, authority, monitoring, and exit controls, not from the model card alone. This makes permit AI risk tiers more useful than vendor-based model rankings.

## A Practical Four-Tier Permit AI Risk Framework

Tier 0 should cover tools with no machine-generated judgment, such as fixed-rule calculators that check setbacks, parcel boundaries, floor-area ratios, or completeness against authoritative cadastral data. Although the phrase “AI” may be used in marketing, a deterministic rule engine does not warrant the same controls as a generative or predictive system. Its governing variables, calculation logic, and failure conditions can be tested directly, and a planner can reproduce the result from the same inputs. Tier 0 nevertheless remains subject to ordinary software validation, cybersecurity, accessibility, record retention, and correction procedures.

Tier 1 should include low-impact assistive uses in which AI produces drafts, summaries, translations, document classifications, or meeting notes and a licensed professional remains responsible for the output. Examples include converting a long zoning application into a searchable chronology or extracting project addresses from scanned plans. Errors can be detected through source comparison and corrected before they acquire legal or material consequences. The minimum controls should include disclosure, source traceability, staff training, privacy review, and a prohibition on using unreviewed output as the sole basis for denial or enforcement.

Tier 2 should cover systems that analyze evidence or recommend decisions in consequential cases, including flood-risk screening, traffic-impact review, historic-resource assessment, affordable-housing scoring, or identification of incomplete submissions. These systems may substantially shape a planner’s recommendation even if a human formally signs it. Controls should include documented purpose and limitations, representative validation, bias testing, logging of model and prompt versions, human override, reasons for overrides, periodic recertification, and an accessible route to contest the result. A recommendation must not become a de facto approval rule merely because staff routinely accept it.

Tier 3 should be reserved for autonomous or agentic systems that can call permit systems, submit forms, alter records, schedule inspections, negotiate remediation plans, make final recommendations within a narrowly bounded workflow, or take irreversible actions. It should also apply where a system can act across several agencies or access sensitive land, infrastructure, utility, public-housing, or protected-community data. For Tier 3, independent testing, continuous monitoring, least-privilege credentials, transaction limits, separation of duties, emergency shutdown, rollback capability, contractual remedies, and explicit public authorization may be necessary. The key threshold is not whether the system speaks like a human; it is whether it can exercise meaningful discretion or authority with limited human intervention.

## Quick answers

### What is a permit AI risk tier?

It is an assigned level of review and control based on a system’s potential to affect public safety, property rights, privacy, administrative fairness, or the reversibility of permit decisions. It considers the model, data, intended use, tool access, and human authority together.

### How many permit AI risk tiers should a city use?

Four tiers are usually sufficient: deterministic automation, low-impact assistance, consequential decision support, and autonomous or agentic action. A city can add emergency controls within any tier, but complexity should not be confused with risk.

### Does using a frontier model automatically make a permit system high risk?

No. A frontier model drafting an unverified FAQ is different from the same model recommending denial of a housing application or changing permit records. The deployment’s authority, data, tools, safeguards, and reversibility determine the relevant tier.

### Can Tier 1 or Tier 2 AI make permit decisions?

Tier 1 output should be reviewed before it materially affects a decision. Tier 2 may recommend an outcome or condition, but a qualified official should make the legal decision, consider the evidence, record reasons, and provide an appeal or correction route.

### When should a city prohibit a permit AI pilot?

A pilot should not proceed if the vendor cannot identify intended uses, provide test data, disclose material limitations, support audits, or demonstrate a way to reverse incorrect actions. Pilot language should not authorize production data, final decisions, or compulsory use.

Canonical: https://urbanplanadvisor.com/knowledge/how_should_permit_offices_tier_ai_risks_in_2026.php
Markdown: https://urbanplanadvisor.com/knowledge/how_should_permit_offices_tier_ai_risks_in_2026.php/index.md