Decision-Grade AI Taxonomy
A three-level classification system for AI authority in public services. The level determines how much disclosure and oversight a system requires.
Why Classification Matters
Every AI system in public services exercises a specific level of authority. Some systems highlight information for a human decision-maker. Others execute routine decisions independently. A few issue binding determinations with legal consequences.
Each level carries different obligations. Stronger authority demands stronger disclosure, more rigorous oversight, and clearer appeal paths. Classification answers a direct question: how much power does this system hold over people?
The Three Levels
| Level | Name | Description | Examples |
|---|---|---|---|
| 1 | Advisory | AI suggests or highlights information. A human decides everything. | Risk scores, checklists, flagging |
| 2 | Assistive | AI completes routine decisions when all criteria are met. A human handles complexity and exceptions. | Auto-approvals, renewals |
| 3 | Determinative | AI makes binding decisions. A human reviews only on appeal. High risk. | Mandatory investigations, automated denials |
Level 1: Advisory
The AI provides information, suggestions, or risk scores. A human decision-maker reviews this input alongside other evidence and makes the final call.
Authority: The system informs. It holds zero decision-making power.
Disclosure requirement: A public statement explaining what the system flags, what data it uses, and how staff incorporate its output into decisions.
Oversight requirement: Regular accuracy reviews, false-positive monitoring, and bias audits on a published schedule.
Level 2: Assistive
The AI executes straightforward decisions when all criteria are clearly met. Complex cases, edge cases, and exceptions go to a human reviewer.
Authority: The system decides routine cases within defined boundaries. A human decides everything else.
Disclosure requirement: A public statement covering what the system decides, the criteria for automatic action, how cases get routed to human review, and how to request manual review of any automated outcome.
Oversight requirement: Sampling of automated decisions for accuracy, tracking of override rates, demographic analysis of outcomes, and published performance metrics.
Level 3: Determinative
The AI issues binding decisions with legal or material consequences. Human review happens only when someone appeals.
Authority: The system holds direct power over outcomes. This is the highest-risk category.
Disclosure requirement: Full public documentation of decision logic, data sources, accuracy rates, and error remediation procedures. Every affected person receives plain-language notice of the AI's role and their right to appeal.
Oversight requirement: Continuous monitoring, independent audits, published accuracy thresholds with automatic remedies when exceeded, and accessible appeal processes with guaranteed human review.
Applying the Taxonomy
To classify a system, answer these questions:
- Who makes the final decision? If always a human: Level 1. If the system decides routine cases: Level 2. If the system decides and humans review only on appeal: Level 3.
- What happens when the system acts? If the output is a suggestion: Level 1. If the output triggers automatic action within defined criteria: Level 2. If the output carries binding legal effect: Level 3.
- How does someone get human review? If human review is the default: Level 1. If human review is available on request: Level 2. If human review requires a formal appeal: Level 3.
Disclosure Scales with Authority
The taxonomy establishes a direct relationship: the stronger the authority, the stronger the disclosure.
| Requirement | Advisory | Assistive | Determinative |
|---|---|---|---|
| Public disclosure statement | Required | Required | Required |
| Individual notice to affected people | Recommended | Required | Required |
| Plain-language explanation of AI role | Required | Required | Required |
| Published accuracy metrics | Recommended | Required | Required |
| Independent audit | Recommended | Recommended | Required |
| Automatic remedies for threshold breaches | — | Recommended | Required |
| Guaranteed human appeal | — | Required | Required |
Classify every AI system in your agency. Publish the classification. Match oversight to authority. When a system crosses from assistive to determinative, strengthen disclosure, add audit mechanisms, and ensure accessible appeal paths.
