Independent AI skills certification
VeriLevel
Get certified

Level 3 of 3

VeriLevel Expert

Design AI workflows and agents others can trust

Expert certifies that you can design AI-powered workflows, automations and agents for a team or for clients, and that you can make them reliable, measurable and safe. It tests design judgment, not product trivia.

Get Expert, $99See exam content
Questions
40
Time limit
60 min
Pass mark
70%
Attempts
4
Price
$99

Who it is for

  • AI leads, automation specialists and solution consultants
  • Product and operations managers responsible for AI rollouts
  • Developers and analysts building AI assistants or pipelines
  • Agency owners and freelancers delivering AI automation to clients
  • Risk, security and compliance professionals who review AI systems

No prior certification is required. Level 2 or equivalent hands-on experience is recommended.

What you will demonstrate

  1. Write system-level instructions, output contracts and tool descriptions for pipelines and agents
  2. Build evaluation sets, choose graders and use evaluations as release gates
  3. Pick the simplest workflow pattern that works and place human review where risk concentrates
  4. Enforce data minimization, retrieval permissions and isolation between clients
  5. Design accountable, reviewable decision paths for high-stakes uses
  6. Constrain agents with least privilege, approvals, budgets, monitoring and injection defenses

Exam content outline

The Expert exam covers 6 domains. Each attempt draws questions from every domain in the proportions below, so you see a balanced exam every time.

1

Instructions and context

15%about 6 questions
  • Write role, scope, decision rules and an escalate-or-unknown path for automated use
  • Separate trusted instructions from untrusted data without relying on wording as security
  • Define structured output contracts that downstream steps can validate
  • Design diverse examples that include edge and negative cases
  • Budget context per step and version prompts as releases
2

Evaluating output

20%about 8 questions
  • Build representative evaluation sets weighted toward hard and high-risk cases
  • Combine deterministic checks, model-based graders and human calibration
  • Control for position, verbosity and self-preference bias in model graders
  • Evaluate retrieval and generation separately in grounded systems
  • Evaluate agents on outcome and trajectory across repeated runs
  • Monitor live traffic with risk-based review sampling and drift detection
3

Everyday workflows

20%about 8 questions
  • Start with the simplest design and add steps only when evaluations show gains
  • Apply chaining, routing, parallel and evaluator patterns with clear stop rules
  • Use deterministic code for calculations, lookups and dates
  • Place human review before external sending and irreversible actions
  • Manage cost and latency per successful outcome, not per call
  • Roll out with shadow runs, pilots, exit criteria and a rollback plan
4

Data and privacy

10%about 4 questions
  • Send each step only the fields it needs and redact before model calls
  • Treat logs and traces as sensitive data stores
  • Enforce the requesting user’s access rights at retrieval time
  • Isolate memory and data between clients or tenants
5

Responsible use

10%about 4 questions
  • Assign named owners for automated decision paths and their failure modes
  • Design meaningful review with evidence, appeal paths and bias testing
  • Counter automation bias and approval fatigue, and track override rates
  • Tier use cases by risk and match controls to the tier
6

AI agents

25%about 10 questions
  • Decide when a predefined workflow beats an agent
  • Scope tools, credentials, destinations, spend and rate limits
  • Show the exact action and evidence in every approval request
  • Break the combination of private data, untrusted content and an outbound channel
  • Validate model output before other systems act on it
  • Set step, time and cost budgets, checkpoints, escalation and a kill switch
  • Trace tool calls and approvals, and stage autonomy from suggest to act within limits

Sample questions

These examples show the style and difficulty of Expert questions. They are not part of the exam question bank.

Sample question 1

An agent reads incoming supplier emails, looks up internal pricing and can send replies on its own. Security testing shows that a crafted email can make it reveal internal prices. Which change most directly removes the risk?

  1. AAdd a stronger instruction telling the agent never to share internal prices
  2. BRemove its ability to send external messages and route drafts to a person for approvalCorrect answer
  3. CSwitch to a larger model that follows instructions more reliably
  4. DShorten the emails the agent reads so injected text is less likely

Why it is correct. The danger comes from private data, untrusted content and an outbound channel in one system. Removing autonomous sending breaks that combination, while wording or model changes cannot guarantee protection.

Sample question 2

Your team changes the prompt of a production assistant that classifies customer tickets. Before the change goes live, what gives the strongest evidence that it is safe to release?

  1. AA few manual spot checks by the prompt author
  2. BAsking a model to rate whether the new prompt is clearer
  3. CRunning the existing evaluation set, including past failures, and comparing results per category with the current versionCorrect answer
  4. DReleasing to all users and watching for complaints

Why it is correct. A versioned evaluation set with past failures, compared per category, shows regressions that spot checks and model opinions miss, before users are affected.

Exam format and rules

Everything runs online. There is no account to create, no appointment and no camera monitoring.

Delivery
Online in your browser, from any computer
Questions
40 multiple-choice questions, mostly workplace scenarios
Time limit
60 minutes, the timer cannot be paused
Pass mark
70%, scored automatically against a fixed answer key
Attempts
4 per purchase: your first attempt and 3 free retakes within 12 months
Results
Right after each attempt, with a breakdown by domain
Certificate
PDF by email when you pass, with a public verification page
Validity
24 months from the issue date
Accommodations
Extra time and other adjustments at no extra cost
Refunds
Full refund any time before your first attempt

Read the full Exam Policy

How to prepare

You do not need a course. These steps use your own work to close the gaps the exam looks for.

  1. 1

    Audit one real system

    Take an AI workflow you built or use and write down its review points, permissions and failure modes.

  2. 2

    Build a small evaluation set

    Collect 30 to 50 real inputs, write reference answers and run two prompt versions against them.

  3. 3

    Threat-model an agent

    List what private data, untrusted content and outbound channels it touches, and remove one.

  4. 4

    Write a rollout plan

    Define pilot scope, exit criteria, owner and rollback steps for your next AI change.

Expert questions

More answers about exams and certificates are on the main FAQ.

Is Expert a technical exam?

It tests design decisions that technical and non-technical leads both make. You will not write code or need to know specific frameworks, but you should understand how pipelines, evaluations and agent permissions work.

Do I need Practitioner first?

It is recommended but not required. Expert assumes you already work confidently with AI and focuses on designing systems for others.

Why is the agents domain so large?

Agents that act on email, files and payments carry the highest risk in real deployments, so they make up about a quarter of the exam.

Get VeriLevel Expert

Pay once, receive your exam link by email and start whenever you are ready. Your certificate arrives the moment you pass.

  • No account needed
  • 4 attempts included
  • Full refund before you start

$99USD, one-time payment

Get ExpertCompare all levels

Other levels