ZYNAGI Risk Intelligence
AI Risk Assessment for Business AI Systems
What could this use case do to patients, clients, or the business — and what has to be checked before it is approved? Work through the seven intake areas, then decide with conditions and an owner rather than a score alone.
Request a use-case reviewTL;DR — Key Takeaways
- A use case is ready to approve when you can name the data it touches, the model and version it calls, who reviews the output, who owns it, and what happens when it fails.
- Seven intake areas cover the decision: data sensitivity, model and provider access, output reliability, human oversight, security, operational impact, and accountability.
- Scoring is for ranking, not assurance. A low band means the evidence collected showed no unresolved exposure in the scored dimensions — not that the workflow is safe.
- Approve with conditions and a review trigger, or hold. An unconditional approval is the most common governance failure.
- A review is only current as of its date: model changes, new data, vendor terms, incidents, and expansion all reopen it.
Executive Summary
AI adoption is moving faster than most organizations can safely govern. Employees use general-purpose AI tools with confidential data. CRM systems integrate AI capabilities without vendor assessment. Website chatbots interact with customers without compliance review. Automation workflows process sensitive information without oversight.
The risk created by ungoverned AI is not theoretical — it is operational. Organizations accumulate exposure across vendor relationships, data handling, workflow dependencies, compliance obligations, and operational reliability. An AI risk assessment provides the structured visibility required to identify, score, and prioritize this exposure before it manifests as an incident.
The first job is not scoring — it is producing an answer you can defend: what data this use case touches, who can see the output, who can stop it, and what happens when it is wrong. The checklist below covers those questions in the order a reviewer asks them.
ZYNAGI reviews use cases with you across six risk categories, scores them across five dimensions, and produces a written report with prioritized actions. The score is a prioritization device, not a certification; the section on what a scored review proves explains the difference.
Before you approve: the intake checklist
Before a use case is approved, these seven areas need an answer you can point to. If an item cannot be answered, treat it as an open risk rather than a formality — an unchecked item is exactly what an auditor will ask about later.
Data sensitivity
- Name the most sensitive field the workflow touches (PHI, client financial data, privileged material, credentials).
- Confirm whether the provider retains inputs, trains on them, or stores them in a region you can name.
- Check that only the fields the task needs are sent, not the whole record.
Model and provider access
- Record which model and version is called, through which account, and who holds the keys.
- Confirm whether the provider can change model behaviour without notice, and how you learn about it.
- List any subprocessors or resale layers between you and the model provider.
Output reliability
- Define what a wrong answer looks like for this workflow, and how often it is tolerable.
- Test on a sample of real cases, including edge cases, and keep the results.
- Decide what the workflow does when confidence is low: stop, escalate, or ask a human.
Human oversight
- Name the person who reviews output before it reaches a patient, client, or regulator.
- Confirm reviewers can see the source material the output was based on.
- Ensure reviewers can override without friction, and that overrides are logged.
Security
- Confirm access control: which roles can use the tool, and how joiners and leavers are handled.
- Check where data is logged, how long it is retained, and who can read the logs.
- Verify prompt-injection and data-exfiltration paths have an owner.
Operational impact
- Identify what breaks if the provider has an outage or changes pricing.
- Confirm a manual fallback exists and staff know how to use it.
- Estimate the time cost of the oversight you are adding, not only the time it saves.
Accountability
- Assign one named owner for the use case, not a committee.
- Record the approval decision, the date, and the conditions attached to it.
- Set the trigger that forces a re-review: model change, new data, incident, expansion.
This checklist is a screening device, not a control framework. For a voluntary reference framework, see the NIST AI Risk Management Framework, which is guidance rather than a legal requirement in the United States.
AI Risk Categories
AI risk is not monolithic. It manifests across six distinct categories, each requiring specific evaluation, controls, and monitoring. A comprehensive AI risk assessment evaluates exposure in every category.
Vendor Risk
Risk created by third-party AI vendors who handle organizational data, operate under varying compliance postures, and create operational dependencies that are difficult to remediate after integration.
Common Examples
AI vendors without BAAs, vendors with undocumented sub-processors, tools with unclear data retention, vendors lacking SOC 2 compliance
Workflow Risk
Risk introduced when AI tools are embedded in operational workflows without evaluation of reliability, accuracy, failure modes, and human oversight requirements.
Common Examples
AI in clinical documentation without validation, automated patient communication without review, AI scheduling without fallback protocols
Data and Privacy Risk
Risk created when AI systems access, process, or transmit sensitive data — PHI, PII, financial data, confidential business information — without appropriate controls.
Common Examples
Employees entering PHI into general-purpose AI, chatbots collecting patient data without disclosure, CRM AI processing client financial data
Compliance Risk
Risk of regulatory violation when AI systems operate without evaluation against HIPAA, state privacy law, SEC regulations, professional responsibility rules, or industry-specific compliance obligations.
Common Examples
AI tools accessing PHI without BAA, financial AI without SEC compliance review, AI in legal workflows without privilege protection
Operational Risk
Risk of business disruption when AI systems fail, produce errors, or create dependencies that the organization cannot sustain or remediate.
Common Examples
AI system outage affecting operations, hallucinated outputs in customer communication, vendor discontinuation of AI service, staff inability to operate without AI tools
Reputational Risk
Risk of damage to organizational trust, brand, and client relationships when AI-related incidents become public or when AI use is discovered to be problematic.
Common Examples
Public disclosure of PHI exposure through AI, AI-generated content errors reaching clients, discovery of unassessed AI tools in operations
These categories are interrelated. Vendor risk creates data exposure. Data exposure creates compliance risk. Compliance violations create reputational harm. A risk assessment that evaluates only one or two categories provides incomplete visibility. The framework requires evaluation across all six to produce an accurate risk picture.
Risk Scoring Model
Qualitative risk identification without quantitative scoring produces lists, not decisions. ZYNAGI's AI risk scoring model converts risk findings into numerical scores that drive prioritization, resource allocation, and executive reporting.
Scoring Dimensions
Each AI deployment is scored across five dimensions, with scores weighted by organizational context and industry requirements:
Data Sensitivity
The sensitivity of data the AI system accesses — from public information to PHI, financial records, and privileged communications.
Vendor Posture
The compliance, security, and governance maturity of the AI vendor — including BAA status, SOC 2 compliance, and data handling practices.
Operational Dependency
The degree to which operations depend on the AI system — from optional tools to mission-critical workflows that cannot function without the AI.
Compliance Exposure
The regulatory obligations triggered by the AI system — HIPAA, SEC, state privacy law, professional responsibility rules, or industry-specific requirements.
Visibility and Controls
The extent to which the AI system is monitored, controlled, and governed — from fully shadow AI with no oversight to formally governed tools with active monitoring.
Risk Score Output
Scores aggregate into an overall risk score from 0 to 100, categorized into risk bands: Low (0–30), Moderate (31–55), Elevated (56–75), and High (76–100). High-risk deployments require immediate executive attention, tighter controls, and more frequent reassessment.
A worked decision, start to finish
Illustrative example — hypothetical, not a client engagement
The point of a review is not to produce a score. It is to turn a vague proposal into a decision with conditions attached. A completed example looks like this:
| Use case | An AI assistant drafts replies to patient scheduling messages for a 12-location dental group. Staff review each draft before it is sent. |
|---|---|
| Evidence available | Vendor will sign a business associate agreement. Administrative safeguard documented. Sample of 40 drafts reviewed by two office managers, with 3 corrections. Staff can send manually if the tool is unavailable. |
| Missing evidence | No written record of whether the vendor trains on submitted content. No test of what the assistant does when a patient asks about insurance coverage or a clinical symptom. No logging review of what staff paste into the tool. No named owner for model changes. |
| Safeguards proposed | Restrict the assistant to scheduling and logistics topics. Block clinical and coverage questions with a fixed fallback response. Require review of every draft. Keep a written log of vendor model changes. Name the Director of Patient Access as owner. |
| Suggested decision | Approved for a limited pilot at one location for 30 days, on condition that clinical and coverage topics are blocked in configuration and the four missing evidence items are closed before expansion. Not approved for multi-location rollout yet. |
Note what the decision is not: it is not “low risk”, and it is not a certification. It is a bounded approval with named conditions and an owner — which is what makes it repeatable when someone asks six months later.
What a scored review proves, and what it does not
The risk score on this page describes a review Zynagi performs with you. It is not produced automatically by a scanner, and there is no self-serve risk scoring tool on this site today.
What the review scores
- Data sensitivity and the sensitivity of the workflows the AI touches.
- Vendor posture: agreements, certifications, retention, subprocessors.
- Operational dependency and the depth of integration.
- Compliance exposure for the jurisdictions and sectors involved.
- Visibility and controls: whether anyone can see, stop, or override the system.
What the score is not
- Not a certification, attestation, or audit opinion.
- Not legal or clinical advice, and not a substitute for counsel review.
- Not a comparison to other organizations. There is no published benchmark behind it.
- Not a substitute for vendor security testing you are required to perform.
- Not permanent — a score reflects the evidence available on the review date.
The 0–100 figure exists for one reason: to rank which use cases need attention first and to make the ranking defensible when resources are limited. A deployment sitting in the Low band is not thereby safe; it means the evidence collected did not show unresolved exposure in the scored dimensions.
Scores move. A vendor term change, a new data source, a model upgrade, or an incident can change the picture without any change on your side. That is why a review is paired with a review trigger rather than treated as a one-off document. The AI risk management guide describes the loop after the score, including a copyable risk register.
Assessment Process
ZYNAGI's AI risk assessment follows a structured three-phase process designed to produce actionable executive intelligence, not generic recommendations.
1. Intake and Discovery
- →Structured AI risk questionnaire
- →AI inventory collection across departments
- →Vendor and data flow mapping
- →Regulatory obligation identification
2. Manual Risk Review
- →AI usage across websites, chatbots, CRM, automation
- →Employee AI behavior and data exposure analysis
- →Vendor compliance and BAA assessment
- →Risk scoring across six risk categories
3. Executive Risk Report
- →Prioritized risk findings by category
- →Quantitative risk scores and risk band assignment
- →Governance maturity assessment
- →Priority action plan for executive leadership
Assessment Deliverables
The assessment produces a practical executive report, not generic AI advice. Deliverables include:
AI tool and vendor inventory
Employee AI usage risk review
Data exposure and sensitive information map
Website chatbot and customer interaction review
CRM, workflow, and automation risk assessment
Vendor compliance and BAA status assessment
Quantitative risk scores by category
Governance maturity score
Priority action plan for executive leadership
Risk monitoring recommendations
Connecting Risk Assessment to Governance
AI risk assessment is not a standalone activity. It is a component of a broader AI governance framework that includes inventory, policy, vendor review, monitoring, and audit trails. The assessment identifies risk; the framework manages it.
Organizations that conduct risk assessments without establishing governance infrastructure find that identified risks re-emerge as new AI tools are adopted. The assessment provides the snapshot; the governance platform provides the ongoing oversight.
For organizations seeking a structured starting point, the AI policy template provides the policy foundation, while the AI Trust Score provides a quantified measure of governance maturity.
Industry Considerations
Healthcare Organizations
Healthcare AI risk assessments must address HIPAA compliance, PHI exposure through AI tools, clinical AI accountability, vendor BAA coverage, and multi-location governance for DSOs and health systems.
Financial Advisory Firms
Financial advisory AI risk assessments must address fiduciary obligations, SEC compliance, client data protection, AI use in advisory workflows, and regulatory reporting requirements.
Law Firms
Legal AI risk assessments must address attorney-client privilege, confidentiality obligations, professional responsibility rules, and AI use in legal research and document review.
Multi-Location Businesses
Organizations operating across multiple locations face compounded risk from inconsistent AI governance, vendor proliferation, and acquisition integration gaps that require centralized assessment.
Governance Checklist
- AI inventory completed across all departments
- Vendor risk assessed for all AI vendors
- Data exposure mapped across AI systems
- Compliance obligations identified for each AI deployment
- Workflow risk evaluated for AI-dependent operations
- Risk scores assigned across all five scoring dimensions
- High-risk deployments identified and flagged
- Executive risk report prepared and distributed
- Risk monitoring schedule established
- Annual reassessment schedule set
Frequently Asked Questions
Next Step
Ready to assess your AI risk?
ZYNAGI helps organizations identify governance gaps, compliance exposure, and operational risk across AI systems.