THE CHARLEE.AI INSURANCE AI KNOWLEDGE CENTER
AI Risk-Tiering Worksheet
Assessment completion0%

Classify the risk before selecting controls

Score inherent risk based on the exact insurance use, then apply mandatory overrides. The final tier sets the minimum governance, testing, approval, and monitoring requirements.

Risk factors

Rate inherent risk before planned controls
IB-01Consumer and business impact

Consumer impact

How directly can the use affect an applicant, policyholder, claimant, beneficiary, producer, or employee?

  1. 0No individual impact; internal support only
  2. 1Indirect or informational impact
  3. 2Material influence on service, prioritization, or decision
  4. 3Determines or executes an eligibility, pricing, coverage, claim, fraud, employment, or other consequential outcome

Suggested evidence Use-case map, affected-person analysis, decision flow

IB-02Consumer and business impact

Decision materiality

How important is the output to the final insurance or business decision?

  1. 0Output is optional and not used for decisions
  2. 1One low-weight input with independent review
  3. 2Substantial input that normally changes decisions
  4. 3Primary or determinative basis for the decision

Suggested evidence Decision policy, weighting, override analysis, workflow sample

IB-03Consumer and business impact

Severity of plausible harm

What is the most credible harm if the system is wrong, biased, unavailable, manipulated, or misused?

  1. 0Negligible inconvenience with no lasting effect
  2. 1Limited delay, rework, or small financial effect
  3. 2Material financial, coverage, privacy, legal, or reputational harm
  4. 3Severe or widespread consumer harm, major loss, loss of rights, or physical safety impact

Suggested evidence Impact assessment, scenario analysis, complaint and incident history

IB-04Consumer and business impact

Scale and concentration

How many decisions, transactions, users, or business processes can be affected before detection?

  1. 0Small pilot or infrequent use
  2. 1One team or limited population with bounded volume
  3. 2Enterprise or multi-state use with significant volume
  4. 3Mass-scale, concentrated, correlated, or rapidly propagating impact

Suggested evidence Volume forecast, jurisdictions, user population, concentration analysis

AH-01Autonomy and human control

Level of autonomy

What can the system do without a person initiating or approving each material action?

  1. 0Generate drafts or analysis only
  2. 1Recommend an action for human approval
  3. 2Execute bounded workflow steps or communications
  4. 3Act autonomously, transact, bind, approve, deny, settle, or change records

Suggested evidence Process diagram, permission model, action catalog

AH-02Autonomy and human control

Human review effectiveness

How meaningful and reliable is human review before a material outcome?

  1. 0Qualified reviewer has time, evidence, authority, and routinely challenges output
  2. 1Reviewer approves with adequate context and tested override process
  3. 2Review is rushed, sampled, inconsistent, or prone to automation bias
  4. 3No meaningful review before the outcome, or reviewer cannot understand or change it

Suggested evidence Review procedure, workload, training, override and disagreement data

AH-03Autonomy and human control

Reversibility

How easily can an affected outcome be stopped, corrected, or reversed?

  1. 0Immediate, complete, and low-cost reversal
  2. 1Reversible through a routine process with limited residual effect
  3. 2Correction requires escalation or may leave material residual harm
  4. 3Irreversible, time-critical, legally binding, or difficult to remediate

Suggested evidence Correction process, rollback test, appeal path, remediation timing

AH-04Autonomy and human control

Time to intervene

How much time exists to identify and control a harmful output or action?

  1. 0Days or longer, with review before use
  2. 1Hours, with effective alerts and intervention
  3. 2Minutes or high-volume same-day activity
  4. 3Seconds or autonomous propagation before meaningful intervention

Suggested evidence Latency map, alert thresholds, kill-switch test, escalation record

DM-01Data and model characteristics

Data sensitivity

What data does the system access, infer, generate, retain, or expose?

  1. 0Public, synthetic, or low-sensitivity operational data
  2. 1Confidential business data with limited personal information
  3. 2Personal, claims, health, financial, geolocation, biometric, or similarly sensitive data
  4. 3Highly sensitive data at scale, protected-class data or proxies, or data whose misuse could cause severe harm

Suggested evidence Data inventory, classification, lineage, privacy assessment

DM-02Data and model characteristics

External data and proxy risk

How dependent is the use on external consumer data, inferred attributes, or variables that may proxy protected characteristics?

  1. 0No external consumer data or inferred personal attributes
  2. 1Limited verified external data with low proxy potential
  3. 2Material external data or inferred attributes requiring proxy testing
  4. 3Opaque, unverified, or highly correlated data used in consequential decisions

Suggested evidence Data dictionary, source contracts, proxy analysis, actuarial support

DM-03Data and model characteristics

Model complexity and change

How difficult is the system to validate, explain, reproduce, and control as it changes?

  1. 0Deterministic rule or simple stable model
  2. 1Interpretable model with controlled periodic updates
  3. 2Complex model, ensemble, continuous data change, or limited transparency
  4. 3Adaptive, self-modifying, opaque, or externally controlled model with material behavior changes

Suggested evidence Model card, version history, validation, change-control records

DM-04Data and model characteristics

Generative and agentic behavior

Does the system generate open-ended content, retrieve untrusted information, call tools, retain memory, or pursue goals?

  1. 0No generative or agentic functionality
  2. 1Closed drafting or summarization with approved sources
  3. 2Open-ended generation, retrieval, memory, or limited tools in a controlled environment
  4. 3Agentic planning, external tools, write access, untrusted content, or autonomous multi-step action

Suggested evidence Architecture, tool registry, prompt-injection tests, trace samples

OC-01Operational and control exposure

System and tool access

What systems, records, communications, funds, or external services can the AI reach?

  1. 0Isolated environment with read-only low-risk access
  2. 1Read access to internal systems or bounded sandbox
  3. 2Write access to material records, communications, or workflows
  4. 3Privileged, production, financial, policy, claim, or externally facing access across systems

Suggested evidence Access matrix, trust boundaries, service accounts, tool permissions

OC-02Operational and control exposure

Reliability and resilience

What happens when the model, data, vendor, integration, or control fails?

  1. 0Failure is obvious and work stops safely
  2. 1Manual fallback exists and recovery is routine
  3. 2Degraded behavior can create material backlog, error, or inconsistent outcomes
  4. 3Failure can silently propagate, prevent critical service, or cause severe correlated loss

Suggested evidence Failure-mode analysis, continuity plan, recovery tests, service levels

OC-03Operational and control exposure

Transparency and contestability

Can the insurer explain the use and outcome, preserve evidence, and support correction, complaint, appeal, audit, and examination?

  1. 0Complete trace and clear explanation are readily available
  2. 1Adequate records and explanation with minor manual work
  3. 2Material explanation, trace, retention, or appeal limitations
  4. 3Outcome cannot be reconstructed or explained sufficiently for affected people or regulators

Suggested evidence Decision trace, notice sample, record-retention design, appeal test

OC-04Operational and control exposure

Security and misuse potential

How attractive and capable is the system as a target or mechanism for fraud, manipulation, exfiltration, or unauthorized action?

  1. 0Minimal attack surface and no meaningful action capability
  2. 1Standard application exposure with mature controls
  3. 2Sensitive data, external content, valuable decisions, or material tool access
  4. 3High-value or privileged target with scalable fraud, exfiltration, evasion, or destructive-action potential

Suggested evidence Threat model, security test, abuse cases, red-team findings

Mandatory overrides

Overrides can raise the tier or restrict the use
OV-01RESTRICTED

Legal or policy restriction

Is the proposed use prohibited, not legally supportable, outside risk appetite, or inconsistent with approved company policy?

If yes Do not deploy. Escalate to Legal, Compliance, Risk, and the accountable executive.

OV-02TIER 4

Automated consequential decision

Can the system determine or execute an adverse or consequential consumer outcome without meaningful human review before the outcome?

If yes Apply Tier 4 controls and obtain executive risk acceptance before any production use.

OV-03TIER 4

Binding or autonomous authority

Can the system autonomously bind coverage, change policy terms, approve or deny a claim, set or transfer material funds, execute settlement, or make another legally binding commitment?

If yes Apply Tier 4 controls, strict transaction limits, confirmations, separation of duties, and tested kill switches.

OV-04TIER 4

Safety or systemic exposure

Could failure create physical safety risk, prevent a critical insurance service, or produce severe correlated impact across a large portfolio or population?

If yes Apply Tier 4 controls, independent assurance, resilience testing, and executive approval.

OV-05MINIMUM TIER 3

Sensitive consequential use

Does the use materially influence underwriting, pricing, coverage, claims, fraud, employment, or another consequential decision using sensitive data, external consumer data, inferred attributes, or potential proxies?

If yes Require fairness, proxy, data, explanation, legal, and consumer-outcome controls before pilot or production.

OV-06MINIMUM TIER 3

Agentic external action

Can a generative or agentic system use external tools, write to production records, send external communications, or act on untrusted content?

If yes Require threat modeling, prompt-injection testing, least privilege, action controls, traceability, and rollback.

Minimum control requirements

Company policy and applicable law may require more
Control Tier 1 Tier 2 Tier 3 Tier 4
Accountable business owner and inventory record Required Required Required Required
Documented purpose, boundaries, users, data, and prohibited uses Basic Detailed Detailed Executive-approved
Legal, compliance, privacy, and security review As needed Targeted Required Required plus executive review
Independent validation and challenge Optional Risk-based Required Required with enhanced independence
Fairness, proxy, and consumer-outcome testing If relevant If relevant Required for consequential use Required before and during use
Human oversight and override testing Basic Documented Tested Demonstrably effective and monitored
Security, abuse, and prompt-injection testing Baseline Targeted Required Independent red team or equivalent
Production monitoring and incident response Basic Defined metrics Thresholds and escalation Continuous or near-real-time
Change approval and reassessment Material changes Defined changes Formal approval Formal approval plus rollback test
Approval authority Business owner Business and control owners AI governance committee Executive risk committee or delegated governing body
Reassessment cadence At material change At least annually At least semiannually At least quarterly and after triggers