AI Risk Assessment: Framework, Template and Tiers

2026-11-21 · Tommaso Maria Ricci

AI Risk Assessment: Framework, Template and Tiers

An AI risk assessment is the step most companies skip, and the numbers show what that costs. In IBM's 2025 Cost of a Data Breach research, 13% of organizations reported breaches involving their AI models or applications, and 97% of those breached organizations lacked proper AI access controls. The same study found that 63% of breached organizations had no AI governance policy or were still writing one. The problem is rarely the model. It is that nobody asked, before launch, what could go wrong, how badly, and who would notice.

This guide gives you a working AI risk assessment framework: how to tier every AI use case as low, medium or high risk, what to check at each tier, a template you can copy into a spreadsheet tomorrow, and a 90-day plan to run it across the company. It is written for founders, COOs, CIOs and compliance leads who need a process that people will actually follow, not a 60-page policy that sits in a shared drive.

I have built and sold companies for more than twenty years, and I now spend a large part of my time helping leadership teams put AI into production. The pattern I see is consistent: the companies that move fastest with AI are not the ones with the fewest rules. They are the ones with a clear, light assessment that tells teams in one afternoon whether a use case is a green light, a yellow light or a stop.

What an AI Risk Assessment Actually Is

An AI risk assessment is a structured review of a specific AI use case that answers four questions before the system goes live:

  1. What can go wrong? Wrong outputs, data leaks, biased decisions, security abuse, legal exposure, vendor failure.
  2. How likely is it? Based on the data, the model, the users and the controls in place.
  3. How bad would it be? For customers, employees, the company's finances and its reputation.
  4. What will we do about it? Accept, reduce, transfer or avoid the risk, with a named owner.

The key word is specific. You do not assess "ChatGPT" or "AI" in general. You assess "the sales team using a generative model to draft proposals with client pricing data" or "an agent that triages inbound support tickets and issues refunds up to 200 dollars." The same model can be low risk in one use case and high risk in another.

This is also where an AI risk assessment differs from the broader topics around it. AI governance sets the rules, roles and decision rights for the whole organization. An acceptable use policy tells employees what they may and may not do. The risk assessment is the operational gate that applies those rules to one use case at a time. Without it, governance stays on paper.

It also differs from using AI for risk management, which means applying machine learning to credit, fraud or operational risk. That is a separate discipline, covered in our guide to AI in risk management. Here we are assessing the risk of the AI itself.

Why Most Companies Get AI Risk Assessment Wrong

There are three failure modes, and most organizations fall into at least one.

Failure mode one: no assessment at all. Teams sign up for tools with a credit card, connect them to email or the CRM, and start pasting in customer data. This is the world of shadow AI, and it is far more common than leaders believe. The IBM 2025 research found that one in five organizations reported a breach due to shadow AI, and that high levels of shadow AI were associated with an average of 670,000 dollars in additional breach costs. If you want to understand the scale of this inside your own company, our guide on shadow AI shows how to measure it.

Failure mode two: the same heavy review for everything. Some companies react by sending every AI idea through a legal, security and privacy committee that meets monthly. A summarization tool for internal meeting notes waits eight weeks next to a credit decision model. Teams learn that the official route is slow, so they go around it. Over-control produces shadow AI just as reliably as no control.

Failure mode three: assessing once and never again. A model that was safe at launch can drift. A vendor can change its terms, its sub-processors or the underlying model. A use case that started with internal data can quietly expand to customer data. An assessment with no review date is a snapshot that ages from the day it is signed.

The consequences show up in project outcomes, not only in incidents. Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and named poor data quality, inadequate risk controls, escalating costs and unclear business value as the causes. Risk controls added late, after a pilot has already been built, are one of the most common reasons a promising project dies before production.

The AI Risk Assessment Framework: Low, Medium, High

The core of a usable AI risk assessment framework is tiering. Low, medium and high risk tiers let you match the depth of review to the stakes. A good tiering model is simple enough that a product manager can apply it in fifteen minutes and get the same answer a risk specialist would.

Here is the tiering model I use with clients. It scores each use case on five factors, each from 1 (low) to 3 (high).

Factor1: Low2: Medium3: High
Decision impactSuggests, a human decides everythingDrafts or recommends, a human reviewsDecides or acts with no human in the loop
Data sensitivityPublic or synthetic dataInternal business dataPersonal, health, financial or client confidential data
Affected peopleInternal staff only, by choiceEmployees in their work, or business clientsConsumers, job candidates, patients, borrowers
ReversibilityErrors are easy to spot and undoErrors are fixable with effortErrors cause lasting harm or cannot be undone
ExposureInternal tool, few usersInternal at scale or limited external usePublic facing, high volume, or autonomous agents

Add the five scores together. A total of 5 to 7 is low risk. A total of 8 to 11 is medium risk. A total of 12 to 15 is high risk. One override applies: any single factor scored 3 on decision impact and 3 on affected people automatically makes the use case high risk, whatever the total. An autonomous decision about a consumer is never a medium-risk matter.

Low-risk tier: green light with guardrails

Typical examples: drafting internal emails, summarizing public research, brainstorming campaign ideas, translating non-confidential documents, coding assistance on non-production code.

What the assessment requires:

  • The tool is on the approved list, or approved through a short vendor check.
  • No personal or confidential data goes in.
  • The user remains responsible for the output.
  • The use case is registered in the AI inventory, in one line.

Time to approve: same day. The goal is to make the safe path the fast path.

Medium-risk tier: yellow light with a review

Typical examples: a support assistant that drafts replies an agent sends, sales proposals generated from CRM data, demand forecasting that feeds a planner's decisions, an internal knowledge assistant over company documents.

What the assessment requires, on top of the low tier:

  • A data review: what data goes in, where it is stored, who the vendor shares it with, how long it is kept.
  • A quality test on a representative sample before launch, with a pass threshold agreed in advance.
  • A defined human review step, and evidence that it actually happens.
  • A named business owner and a review date, usually every six months.

Time to approve: one to two weeks.

High-risk tier: red light until proven otherwise

Typical examples: screening job candidates, setting credit limits, triaging patients, pricing insurance, agents that can move money or change customer accounts, anything that falls into the high-risk categories of the EU AI Act.

What the assessment requires, on top of the medium tier:

  • A full impact assessment covering accuracy, bias across relevant groups, explainability, security testing and failure scenarios.
  • Legal review against applicable regulation in every market where the system operates.
  • Monitoring in production with alert thresholds and a kill switch.
  • Sign-off by a senior executive, not only by the project team.
  • A review at least quarterly and after any significant change to the model, data or vendor.

Time to approve: four to eight weeks, sometimes more. That is acceptable, because high-risk use cases are few. In most mid-sized companies I work with, fewer than one in ten AI use cases ends up in this tier.

The Risk Categories Your Assessment Must Cover

Tiering tells you how deep to go. The risk categories tell you what to look at. A complete assessment covers eight categories. For generative AI, the NIST Generative AI Profile (NIST AI 600-1), published in July 2024, lists twelve risks that are unique to or made worse by generative AI, including confabulation, data privacy, information security, intellectual property, harmful bias and value chain integration. The eight categories below translate that kind of catalog into business language.

1. Accuracy and reliability. Does the system give correct answers often enough for this use? Generative models confabulate: they produce confident, fluent and wrong content. The question is not whether errors happen but whether the process catches them before they cause harm.

2. Data privacy. What personal data enters the system, on what legal basis, and where does it go? Check training use, retention, sub-processors and data residency. A vendor that trains on your inputs by default is a different risk from one that contractually does not.

3. Security. Can the system be manipulated through prompt injection, poisoned documents or malicious inputs? Who can access it, and with what permissions? The IBM finding that 97% of organizations with AI breaches lacked proper access controls points directly here.

4. Bias and fairness. Could the system treat groups of people differently in ways that are unlawful or unacceptable? This matters most for hiring, lending, insurance, healthcare and any decision about individuals.

5. Transparency and explainability. Can you explain to a customer, an employee or a regulator how the system reached a result? Do people know they are interacting with an AI?

6. Legal and regulatory. Which laws apply: data protection, consumer protection, employment law, sector rules, intellectual property, the EU AI Act? Who owns the outputs, and could they infringe someone else's rights?

7. Vendor and value chain. What happens if the vendor changes the model, raises prices, suffers an outage or goes out of business? Can you switch? Do you know which third-party models sit underneath the product you bought?

8. Operational and reputational. What happens if the system fails at scale on a busy day? What would a screenshot of its worst output look like on social media?

Not every category needs depth for every use case. In a low-risk case most of them get one line. In a high-risk case each gets its own section and its own evidence.

Mapping Your Assessment to NIST, ISO 42001 and the EU AI Act

You do not need to invent your own methodology from scratch, and you should not. Three references give your assessment credibility with auditors, clients and regulators.

NIST AI Risk Management Framework. The NIST AI RMF, released in January 2023, is voluntary and widely used. It organizes AI risk work into four functions: Govern, Map, Measure and Manage. Your tiering model is part of Map, your quality and bias tests are Measure, your controls and monitoring are Manage, and the policy and ownership that sit above everything are Govern. Using its vocabulary costs nothing and makes your process easy to explain to US clients.

ISO/IEC 42001. Published in December 2023, it is the international standard for an AI management system, and it is certifiable. It requires organizations to assess AI risks and the impact of AI systems on individuals and society. If enterprise clients send you security questionnaires, expect ISO 42001 questions to appear more often. A tiered risk assessment with documented evidence is the backbone of any future certification.

EU AI Act. The EU AI Act (Regulation 2024/1689) entered into force on 1 August 2024 and takes its own risk-based approach: prohibited practices, high-risk systems, systems with transparency obligations, and minimal-risk systems. The prohibitions and the AI literacy provision of Article 4 have applied since 2 February 2025. The transparency obligations of Article 50, such as telling people they are interacting with an AI, have applied since 2 August 2026. After the 2026 amendments, the obligations for high-risk systems listed in Annex III, which covers areas such as employment, credit and access to essential services, apply from 2 December 2027, and those for AI built into regulated products under Annex I apply from 2 August 2028.

Two practical points for companies selling into or operating in Europe. First, Article 4 asks providers and deployers to take measures to support the AI literacy of their staff. It does not set a specific level or a number of training hours, but documented training is the easiest measure to demonstrate. Second, your internal tiering should flag any use case that might fall into the Act's high-risk categories early, because the documentation, human oversight and monitoring those systems require take months to build properly.

A note on alignment: your internal "high risk" and the AI Act's "high-risk" are not the same thing. A customer-facing chatbot with access to order data may be high risk for your business without being a high-risk system under the Act. Keep both labels in your template, side by side.

The AI Risk Assessment Template

This is the template I give clients. Build it as a spreadsheet or a form, one row per use case, so that the whole AI inventory and its risk profile live in one place.

Section A: Identification

  • Use case name and one-sentence description
  • Business owner (a person, not a department)
  • Technical owner
  • Tool or vendor, and the underlying model if known
  • Status: idea, pilot, production, retired
  • Date of assessment and date of next review

Section B: Context

  • What decision or task does the AI support?
  • Who uses it, and how many people?
  • Who is affected by its outputs?
  • What data goes in? Tick: public, internal, personal, sensitive personal, client confidential, financial
  • Does it act on its own, or does a human approve each output?

Section C: Tiering

  • Score for each of the five factors (1 to 3)
  • Total score and resulting tier: low, medium, high
  • Override applied? Yes or no, and why
  • Possible EU AI Act category: prohibited, high-risk, transparency, minimal, not applicable

Section D: Risk register

For each relevant category: risk description, likelihood (1 to 3), impact (1 to 3), existing controls, additional controls needed, residual risk, owner.

Section E: Evidence

  • Vendor documentation reviewed (security, data processing agreement, model card)
  • Test results with sample size and pass threshold
  • Bias test results, for high-risk cases
  • Human review procedure
  • Monitoring metrics and alert thresholds

Section F: Decision

  • Approve, approve with conditions, or reject
  • Conditions and deadlines
  • Approver name and role
  • Signature and date

The template's value comes from discipline, not complexity. One clean row per use case, updated on schedule, is worth more than a beautiful document no one maintains.

If you want this template adapted to your industry, your tool stack and the regulations that apply in your markets, that is exactly the kind of work I do with leadership teams. A short advisory session, requested through the consultation page on this site, is usually enough to tell you which of your current AI use cases need attention first.

AI Vendor Risk Assessment: Questions to Ask Before You Sign

Most companies do not build models. They buy them, inside SaaS products, copilots and APIs. That makes AI vendor risk assessment one of the most important parts of the process, and one of the most neglected. Traditional vendor questionnaires ask about encryption and uptime but miss the questions that matter for AI.

Ask every AI vendor these twelve questions:

  1. Do you use our inputs or outputs to train or improve your models? Is that off by default, and is it in the contract?
  2. Which third-party models do you use, and will you notify us before changing them?
  3. Where is our data processed and stored, and which sub-processors have access?
  4. How long do you retain prompts, files and outputs, and can we delete them on request?
  5. How do you protect against prompt injection and data exfiltration through the AI features?
  6. What role-based access controls exist for AI features, and can we restrict them by team?
  7. Do you log AI interactions, and can we export those logs for audit?
  8. What testing do you do for accuracy and bias, and can you share results?
  9. Do you hold ISO 27001, SOC 2 or ISO 42001, and does the scope cover the AI features?
  10. How will you support our obligations under the EU AI Act and data protection law?
  11. Who owns the outputs, and do you offer indemnity for intellectual property claims?
  12. What happens to our data and our workflows if we end the contract?

A vendor that answers these clearly, in writing, is a lower risk than a vendor with a better demo and vague answers. When the answers are weak, the choice between building and buying changes too, a trade-off covered in our analysis of build vs buy for AI software.

Self-Assessment: How Mature Is Your AI Risk Process?

Before you design anything new, score where you are today. Answer each question yes or no.

Visibility

  1. Do you have a single inventory of the AI tools and use cases running in the company?
  2. Do you know which of them process personal or client confidential data?
  3. Do you have a way to detect AI tools that were adopted without approval?

Process

  1. Is there a written, tiered process for approving new AI use cases?
  2. Can a low-risk use case be approved in under a week?
  3. Does every AI use case have a named business owner?
  4. Is there a review date for each approved use case?

Controls

  1. Do you test outputs on a sample before launch, with a pass threshold?
  2. Are AI features covered by role-based access controls?
  3. Do contracts with AI vendors address training on your data and model changes?

People

  1. Have employees received practical guidance on what data they may put into AI tools?
  2. Does a senior executive own AI risk, with the authority to stop a project?

Scoring: count your yes answers.

  • 0 to 4: Exposed. You are likely running more AI than you know, with no evidence of control. Start with the inventory and a one-page acceptable use rule.
  • 5 to 8: Developing. You have pieces in place but gaps that an incident or a client audit would expose. Formalize tiering and vendor review.
  • 9 to 12: Managed. You have a working process. Focus on monitoring, periodic review and making the approval path faster.

If you want a broader view that goes beyond risk, the AI readiness assessment scores data, skills, leadership and infrastructure as well.

How the Assessment Plays Out in Real Projects

Theory is easy. Here is how tiering changed real decisions in projects I have been involved in. The companies are anonymized by sector.

A sports distribution company. We used AI to personalize marketing and product recommendations, which contributed to a 30% increase in sales. The assessment placed the content generation part in the low tier: internal teams drafted copy, humans approved every campaign. The personalization engine, which used customer purchase history, landed in the medium tier because of the personal data involved. That meant a data review, a check of the vendor's data processing terms and a rule that no sensitive categories would be inferred. None of this slowed the project meaningfully. It made the legal team an ally instead of a blocker.

A hotel. Revenue grew from 9 million to 10 million with AI-supported pricing and revenue management. Pricing touches customers directly, so the model scored medium on exposure and impact. The control that mattered most was a human approval band: the system could move prices within agreed limits automatically, and anything outside the band went to the revenue manager. That single control removed the scenario that worried the owners most, a runaway price on a peak weekend.

A medical center. AI helped increase capacity by 20%, mainly through scheduling and administrative automation. Here the tiering did its most important work. Scheduling optimization based on appointment types scored medium. Any idea that touched clinical triage or used health data to prioritize patients scored high and was kept out of scope for the first phase. Separating the two let the center capture the operational gain quickly while keeping the sensitive questions for a slower, properly governed second phase.

An agritourism business. Guest numbers doubled, with AI used in marketing, booking communication and reviews management. Almost everything sat in the low tier. The assessment took an afternoon. That is the point: a proportionate process does not burden small, low-risk projects.

The common lesson is that tiering does not slow companies down. It tells them where speed is safe and where care is worth the time.

Monitoring After Launch: The Part Everyone Forgets

An AI risk assessment is not finished when the system goes live. Models, data, users and vendors change. Your monitoring should match the tier.

For low-risk use cases: an annual check that the tool is still approved, still used as described and still free of sensitive data.

For medium-risk use cases:

  • A quality sample every month or quarter, reviewed against the original threshold
  • Tracking of user overrides: if humans reject 40% of drafts, the system is not working
  • A check of vendor changes: new model versions, new sub-processors, new terms
  • A review every six months, or sooner if the scope expands

For high-risk use cases:

  • Continuous metrics on accuracy, error rates and, where relevant, outcomes across groups
  • Alert thresholds that trigger a human investigation
  • A tested kill switch and a manual fallback process
  • Incident logging, with root cause analysis for every significant failure
  • A formal review every quarter and after any material change

Agents deserve special attention. An AI agent that can call tools, send messages or change records has a larger blast radius than a model that only writes text. Permissions, spending limits and action logs are the controls that matter most, as covered in our guide to AI agent governance.

Who Owns AI Risk Assessment Inside the Company

A process without owners is a document. These are the roles that make it work.

Executive sponsor. A C-level leader, often the COO, CIO or general counsel, who owns AI risk overall, resolves disputes and has the authority to stop a project.

Business owner of each use case. The person accountable for the outcome. They fill in the template, own the risk register and answer for the system in production.

Risk reviewers. Security, privacy and legal specialists who review medium and high-risk cases. In smaller companies this can be one person plus external advisors.

AI inventory keeper. Someone, often in IT or the operating team, who maintains the register, chases review dates and spots unregistered tools.

Employees. Everyone who uses AI tools, trained to know what data they may use and how to flag a new use case.

In companies with a dedicated AI team, these roles often sit inside a center of excellence. Our guide to the AI operating model explains how to set that up without creating a bottleneck.

Your 30/60/90-Day Roadmap

Here is how to go from no process to a working AI risk assessment in one quarter.

Days 1 to 30: See what you have

  1. Appoint the executive sponsor and the inventory keeper.
  2. Build the AI inventory. Survey department heads, check expense reports and software subscriptions, and review network or identity logs for AI services.
  3. Publish a one-page rule on what data may never go into AI tools, effective immediately.
  4. Create the approved tools list with the tools that are already widely used and pass a basic vendor check.
  5. Draft the tiering model using the five factors above, adapted to your sector.

Deliverable at day 30: a complete inventory with a first, rough tier for every use case.

Days 31 to 60: Assess what matters

  1. Finalize the template and test it on five real use cases across the three tiers.
  2. Run full assessments on every use case currently scored high, and on the medium-risk cases with personal data.
  3. Send the twelve vendor questions to the vendors behind those use cases.
  4. Fix the urgent gaps: access controls, data processing agreements, human review steps that exist on paper but not in practice.
  5. Train managers on how to use the tiering model in under an hour.

Deliverable at day 60: assessed and documented high-risk and key medium-risk use cases, with a remediation list.

Days 61 to 90: Make it routine

  1. Embed the process in procurement and project approval, so no AI tool is bought or built without a tier.
  2. Set review dates for every approved use case.
  3. Launch monitoring for the medium and high-risk systems.
  4. Run a short AI literacy session for all staff on safe use, data rules and how to request a new tool.
  5. Report to the board or leadership team with the inventory, tiers, open risks and next steps.

Deliverable at day 90: a repeatable process, a living register and a first leadership report.

This roadmap works for companies from about 50 to a few thousand employees. Larger groups need the same steps run by business unit, with a central team setting the standard.

The Most Common Mistakes to Avoid

Copying a framework built for banks. A 40-control assessment designed for a regulated financial institution will crush a 200-person company. Start light and add depth only where the tier demands it.

Letting legal own the whole process. Legal review is essential for high-risk cases, but if every case starts with legal, teams will avoid the process. The business owner should drive the assessment.

Assessing tools instead of use cases. Approving "Copilot" or "ChatGPT Enterprise" as a tool is necessary but not sufficient. The same tool used for internal notes and for candidate screening carries very different risks.

Ignoring the data already shared. Many companies discover during the inventory that confidential data has been going into consumer AI tools for months. Deal with that honestly: assess what was shared, notify where required and fix the controls.

Treating the assessment as a one-time project. The first inventory is the easy part. Keeping it current as teams adopt new tools every month is where the process proves its value.

Skipping the literacy part. Most AI risk comes from well-meaning employees who do not know what they should not paste into a chat window. Practical guidance prevents more incidents than any technical control. Pair the assessment with an AI acceptable use policy that people can read in five minutes.

Turning Risk Assessment Into a Competitive Advantage

Most companies see AI risk assessment as a cost. The better way to see it is as a sales asset and a speed asset.

It wins deals. Enterprise buyers now ask suppliers how they govern AI. A supplier that can show its AI inventory, its tiering model and its vendor controls answers the security questionnaire in a day instead of a month, and looks more trustworthy than competitors who improvise.

It speeds up adoption. When teams know that a low-risk use case gets approved the same day, they bring ideas to the official process instead of hiding them. Visibility goes up, shadow AI goes down, and the company learns faster.

It protects the projects that matter. By concentrating scrutiny on the few high-risk cases, you give them the testing and oversight they need to reach production rather than being abandoned after the pilot.

It prepares you for regulation. The EU AI Act's high-risk obligations apply from 2 December 2027 for Annex III systems. Companies that already have an inventory and a documented assessment process will adapt their existing work. Those that do not will start from zero under pressure.

If you are deciding how to structure AI risk assessment in your company, or you need a second opinion on whether a specific use case is as safe as the project team believes, an outside view is often the fastest route. You can request an advisory conversation through the consultation request page: bring your AI inventory, even an incomplete one, and we will work out where your real exposure sits.

FAQ

What is an AI risk assessment?

An AI risk assessment is a structured review of a specific AI use case before and after it goes live. It identifies what could go wrong, such as wrong outputs, data leaks, biased decisions or security abuse, estimates how likely and how severe each risk is, and defines controls with a named owner. It assesses use cases, not tools, because the same model can be low risk in one context and high risk in another. It should be repeated whenever the model, data, vendor or scope changes.

How does an AI risk assessment framework with low, medium and high tiers work?

A tiered AI risk assessment framework scores each use case on a few factors, typically decision impact, data sensitivity, affected people, reversibility and exposure. Each factor gets a score from 1 to 3, and the total places the use case in a low, medium or high tier. Low-risk cases get a quick check and same-day approval. Medium-risk cases need a data review, quality testing and human review. High-risk cases need a full impact assessment, legal review, monitoring and executive sign-off.

What should an AI risk assessment template include?

A practical AI risk assessment template includes six sections: identification (use case, owners, vendor, status, review date), context (users, affected people, data types, level of autonomy), tiering (factor scores, total, tier, possible EU AI Act category), a risk register (risk, likelihood, impact, controls, residual risk, owner), evidence (vendor documents, test results, review procedures, monitoring) and a decision with conditions and an approver. One row per use case in a shared register works better than long standalone documents.

How long does an AI risk assessment take?

It depends on the tier. A low-risk use case, such as drafting internal emails with an approved tool, should take less than a day. A medium-risk use case, such as a support assistant drafting replies from customer data, usually takes one to two weeks including a data review and a sample test. A high-risk use case, such as candidate screening or credit decisions, typically needs four to eight weeks because it requires bias testing, legal review and monitoring design.

Is an AI risk assessment required by the EU AI Act?

The EU AI Act requires specific risk management, documentation and human oversight for high-risk AI systems, with obligations for Annex III systems applying from 2 December 2027 and for Annex I systems from 2 August 2028. It also imposes transparency duties under Article 50, applicable since 2 August 2026, and asks companies under Article 4 to take measures that support the AI literacy of their staff. A tiered internal assessment is the most practical way to identify which of your systems fall under which obligations.

What is an AI vendor risk assessment?

An AI vendor risk assessment evaluates the AI-specific risks of a third-party product before you buy or renew it. Beyond standard security questions, it asks whether the vendor trains models on your data, which third-party models it relies on, where data is processed and retained, how it prevents prompt injection, what access controls and audit logs exist, which certifications cover the AI features, who owns outputs and what happens to your data when the contract ends.

Who should be responsible for AI risk assessment in a company?

Responsibility should be shared but clearly assigned. A senior executive owns AI risk overall and can stop a project. Each use case has a business owner who completes the assessment and answers for the system in production. Security, privacy and legal specialists review medium and high-risk cases. An inventory keeper maintains the register and review dates. Employees are trained to recognize and report new AI use cases.