Exam program: AI governance and data protection

IAE-EXAM-AI · version 1.0

Sections, format, sample tasks and the scoring rubric of the IAE written exam in AI Governance and Data Protection.

Version 1.0Effective 2 November 2026Exam in English or Russian

The exam is not a legal qualification and does not test the law of any single country.

01. Who is the exam for?

For specialists in AI governance, data protection, compliance, risk and internal audit, and for AI product and ML team leads, with a bachelor's degree or higher and at least 3 years of experience (5 years without a degree). The exam tests work decisions on AI systems and personal data: inventory and classification of systems, assessment and testing before launch, human oversight and monitoring, breach response, and the terms of contracts with model vendors.

The exam is not a legal qualification, does not give the right to practice law and does not test the law of any single country. GDPR, CCPA/CPRA and the EU AI Act are used as reference frameworks of international practice: the exam tests how their principles are applied to a situation, not article numbers or exact wording. An answer may rely on another framework or on the law of the candidate's own country if the requirement is named and the decision is justified.

02. Format

  • 5 questions, one per section S1-S5, 10 minutes each.
  • 2 cases of 25 minutes from the case set, for example: a pre-launch review of an AI CV screener with given error rates by group, a personal data breach with given numbers of records and timelines, a review of a contract with a language model vendor, a monitoring alert on model drift.
  • 1 question about the project from the candidate's questionnaire, 10 minutes: an AI system or a data protection decision the candidate worked on personally.
  • Up to 2 hours in total, in one sitting. Pasting is disabled, and each candidate gets their own figures and circumstances in the tasks.
  • Deadlines and thresholds that the calculation depends on (breach notification deadline, monitoring thresholds, an internal risk scale) are given in the task, usually as company policy.

03. Sections and weights

  • S1. AI inventory and risk classification - 15 %.
  • S2. Pre-launch assessment and testing - 20 %.
  • S3. Data protection - 25 %.
  • S4. Human oversight, monitoring and incidents - 20 %.
  • S5. Vendors, contracts and governance practice - 20 %, including the question about the candidate's own project.

04. Section topics

S1. AI inventory and risk classification.

  • The AI inventory: what goes into it, how to find unregistered services, a sample audit of the inventory.
  • Risk classes under a reference framework: prohibited practices, high risk, transparency obligations, minimal risk; an internal risk scale and its weaknesses.
  • Roles of the parties: provider and deployer of an AI system, controller and processor of data.
  • A substantial change to a model and use of a model for a new purpose.
  • Telling people they are interacting with AI; staff use of public AI services.

S2. Pre-launch assessment and testing.

  • An impact assessment before launch: content, participants, who approves it.
  • Selection rates and error rates by group, the ratio of rates, the four-fifths rule as a reference point.
  • Uncertainty of estimates in small groups, the confidence interval of a rate, the required sample size.
  • Choosing a decision threshold, the trade-off between accuracy and the gap between groups.
  • Launch criteria, residual risk and the launch decision.

S3. Data protection.

  • Lawful basis for processing and compatibility of a new purpose with the original one.
  • Minimisation of data and model features, retention periods by record category.
  • Data subject requests: the backlog, response deadlines, deletion of data and trained models.
  • Cross-border transfer: choosing a mechanism, a transfer risk assessment, supplementary measures.
  • Consent and opt-out for sharing data with advertising platforms, anonymisation and pseudonymisation.

S4. Human oversight, monitoring and incidents.

  • Real human oversight: time per decision, signs that reviewers agree with the model without checking.
  • Monitoring when outcomes arrive late, data drift and the population stability index (PSI), appeals as a signal.
  • The decision to stop a model or keep it running, AI system logs and their retention periods.
  • The size of a sample check of AI assistant answers for a given precision.
  • Personal data breaches: the timeline, when the notification deadline starts, risk assessment, whom to notify and how many people.

S5. Vendors, contracts and governance practice.

  • Checking an AI vendor before purchase.
  • Terms of a contract with a model vendor: training on customer data, model version changes, sub-processors and processing locations, incident notification, the right to audit, exit from the contract.
  • The AI committee and approval of projects by risk level, launching before the assessment is finished.
  • An AI risk report to the board, staff training by role.
  • The candidate's own project: their role, decisions, figures and outcome.

05. Sample tasks

Below are one question and one case as the candidate sees them, with the rubric the examiner uses to score them. These tasks are not part of the exam. In the exam, each candidate gets their own figures and circumstances.

Question (section S3, 10 minutes). A company wants to replace office access badges with face recognition at the turnstiles. The office has 1,200 employees, and each passes a turnstile 4 times a day on average. According to the vendor, the system wrongly refuses 0.5 % of passes by employees with lighter skin tones and 2 % of passes by employees with darker skin tones; the company has 900 and 300 such employees. HR proposes to obtain employee consent in an addendum to the employment contract.

  • 1) How many false refusals per day will there be in each group, and how many times a month (20 working days) will one employee of each group be refused on average?
  • 2) What requirements for processing biometric data must be met, and can the company rely on employee consent?
  • 3) Should the system be launched, and on what conditions?

Rubric for the question.

  • Correctness (40 points): 18 false refusals per day in the first group (3,600 passes x 0.5 %) and 24 in the second (1,200 x 2 %); 0.4 and 1.6 refusals per employee per month, that is, 4 times as often; the second group is 25 % of employees and 57 % of refusals. Biometric data used for identification is treated as a special category. The answer explains why employee consent is usually not considered freely given and why it cannot be relied on without an equivalent alternative (the badge). It names an impact assessment before launch, where and how long templates are stored, their deletion when an employee leaves, and a ban on the vendor using the data for training.
  • Reasoning (25 points): the decision follows from the figures and from proportionality: badges already do the job, so a more intrusive method needs separate justification. Both not launching and a voluntary mode with the badge as an alternative are accepted if conditions are stated, for example a test on the company's own employees and an acceptable gap in refusal rates between groups.
  • Risks and constraints (20 points): the discriminatory effect on the second group; a leaked template cannot be fixed by changing a password; use of the system for time tracking without a new assessment; the vendor's figures were not measured on the company's employees; matching a face against the whole database (1:N) carries more risk than matching against a template on the employee's card (1:1).
  • Clarity (15 points): the answer is structured in points, conclusions and figures are stated explicitly, no padding. Grammar and style do not reduce the score if the meaning is clear.
  • An answer whose calculation or conclusion does not rest on the figures in the candidate's own task gets no more than 19 points for Correctness, even if the general reasoning is right.

Case (section S2, 25 minutes). A grocery delivery service receives 90,000 refund requests a month for spoiled goods; the average refund is $18. Today every request is checked by an employee, and one check costs $1.50. The company wants to launch a model that uses the product photo and the customer's order history to approve 70 % of requests on its own, reject 10 % and pass 20 % to an employee. On a validation sample of 2,000 requests that employees later reviewed, 15 % of automatic rejections turned out to be valid claims and 4 % of automatic approvals turned out to be fraudulent. The image recognition model comes from an external vendor: under its terms, photos are kept for 30 days and may be used to improve its models. Some photos show customers' faces, documents and addresses.

  • 1) Calculate per month: the number of false rejections, the amount paid on fraudulent requests and the savings on checks compared with the current process.
  • 2) Can automatic rejections stay? Propose a human oversight procedure and calculate the staff workload and the savings under it.
  • 3) What conditions on photos and customer data will you set for the vendor and for your own team?
  • 4) Should the model be launched, and on what conditions? Name monitoring metrics with thresholds.

Rubric for the case.

  • Correctness (40 points: item 1 - 12, item 2 - 10, item 3 - 9, item 4 - 9). Item 1: 9,000 automatic rejections, of which 1,350 are false, so $24,300 is not paid to customers; 63,000 automatic approvals, of which 2,520 are fraudulent, $45,360; checks cost $135,000 today and $27,000 with the model, savings $108,000. Item 2: if all rejections go to employees, they check 27,000 requests (30 % of today's volume), checks cost $40,500, savings $94,500; the extra $13,500 a month is $10 per false rejection prevented. Item 3: an instruction to photograph only the product, cropping or blurring faces and documents before sending to the vendor, a ban on the vendor training on the photos, retention periods at the vendor and at the company, processing locations and sub-processors. Item 4: a decision with conditions, for example a pilot on part of the requests and stopping automatic decisions when a threshold is exceeded.
  • Reasoning (25 points): the decision on automatic rejections rests on the 15 % error rate and on the fact that a rejection affects the customer's right to a refund; the chosen oversight procedure is compared with another option by cost; the answer notes that the savings ($94,500) exceed the payments on fraudulent approvals ($45,360), but the level of fraud that employees miss today needs to be measured.
  • Risks and constraints (20 points): telling the customer about the automated decision and the right to human review; checking errors by customer group (new and repeat customers, delivery areas); metrics with thresholds: the share of false rejections in a weekly sample, the share of fraud in a random sample of approvals, the share of appeals and reversed decisions, a shift in the approval rate; a model version change at the vendor.
  • Clarity (15 points): the answer is structured in points, conclusions and figures are stated explicitly, no padding. Grammar and style do not reduce the score if the meaning is clear.
  • A generic answer that does not work through the figures of this case and does not choose between options gets no more than 19 points for Correctness and no more than 10 for Reasoning.

06. Scoring rubric

Each task is scored on four criteria; the total is converted to 100 points using the section weights.

  • Correctness of the solution - 40 %: the calculation and conclusion based on the figures and circumstances of the candidate's own task.
  • Reasoning behind decisions - 25 %: why this option was chosen over another, based on the figures and circumstances of the task.
  • Risks and constraints - 20 %: consequences for people, data, vendors, what can go wrong after launch.
  • Clarity - 15 %: structure and precise wording. Grammar and style do not lower the score if the meaning is clear.

A reference to a rule without applying it to the situation adds no points. A reference to another framework or law counts the same as a reference to GDPR, CCPA/CPRA or the EU AI Act if the requirement is stated correctly.

07. Result and certificate

  • Pass mark - 70 of 100.
  • Result in 5 business days. Those who pass receive an IAE Certified Professional certificate stating the field, valid for 3 years; scores by section are in a separate report.
  • The certificate and the public IAE registry show the field AI Governance and Data Protection, program IAE-EXAM-AI and its version, the certificate number and the validity date.
  • The certificate confirms the exam result. It is not a license, a legal qualification or confirmation that an organization complies with the law.
  • Retake 30 days after the first failed attempt and 60 days after the second, with no more than 3 attempts in 12 months.
  • Full terms: the Exam terms page.
  • The Credentialing Council reviews the program at least once a year; tasks are updated when the reference frameworks change.

08. What does the exam not cover?

  • Article numbers, exact wording and effective dates of rules; the law of a particular country and case law.
  • Legal characterisation and drafting of legal documents: a vendor contract is assessed by its terms and risks, not by legal drafting technique.
  • Writing code, training and tuning models: the tasks require calculations on given figures, not programming.
  • Technical configuration of security tools, cryptography, penetration testing.
  • Certification audit procedures under management system standards.
  • Detailed sector rules (medical devices, banking supervision of models, insurance regulation): tasks include sector situations, but answers are scored against general principles.

09. Preparation

  • Work through the sample tasks and the rubric on this page.
  • Reference frameworks at the level of principles: GDPR (lawful basis, data subject rights, impact assessment, breach notification, cross-border transfer), CCPA/CPRA (consumer rights, opting out of sale and sharing), the EU AI Act (risk classes, provider and deployer obligations, human oversight, transparency). NIST AI RMF and ISO/IEC 42001 are useful as models of an AI risk management system.
  • Calculations that appear in the tasks: selection rates and their ratio between groups; false rejection and false positive rates by group; the 95 % confidence interval of a rate and the required sample size (n = 1.96² x p x (1 - p) / e²); deadlines in hours and days from the moment the company became aware of an event; cost and effort (number of cases x time x rate). A calculator is enough for the calculations.
  • Structure your answer as: calculation on your own figures, decision, conditions, risks. General statements without the task's figures get no more than 19 of 40 points for Correctness.
  • The practice exam on the website gives tasks of the same format from a separate set; they are not repeated in the exam.
  • Reference materials are allowed during the exam. You write the answers yourself: the exam terms prohibit text from AI services and help from other people.

Exam

Book the exam

The questionnaire takes 15 minutes. Admission decision in 1 business day.

Candidate questionnaire

Received

Admission decision within one business day

FieldAI governance and data protection
DurationUp to 2 hours
Result5 business days