EMA AI Standard v1.0 · Exam Content Outline
What the EMX-AI examination assesses.
The authoritative description of the examination. Where marketing copy and this outline differ, the outline is correct.
The EMX-AI Exam Content Outline
- Questions
- 60
- Minutes
- 90
- Pass standard
- 75%
- Valid for
- 2 years
Sitting it requires a valid EMA-I, EMP-II, EME-III when the sitting starts. Each statement below is a proficiency indicator: what a certified holder can do. Its scope tag says at what altitude questions on it place the candidate.
AI-1Adoption strategy and staged bets
8 of 60 questions- AI-1.I1A certified AI Extension holder can decide where AI assistance is adopted from evidence about the team's own work, rather than from the market's enthusiasm or a mandate. team
- AI-1.I2A certified AI Extension holder can frame an AI initiative as a staged bet, with a stated hypothesis, a cost ceiling and a stop condition agreed before it starts. org
- AI-1.I3A certified AI Extension holder can sequence adoption so that review, data handling and measurement are in place before the scale arrives. org
- AI-1.I4A certified AI Extension holder can tell an executive what an AI investment can and cannot be expected to return, in the organisation's own terms, without overpromising or dismissing it. exec
- AI-1.I5A certified AI Extension holder can narrow or stop an AI initiative that is not meeting its hypothesis, and say so openly. org
Situations questions present
- The board asks engineering for an AI strategy by the end of the quarter.
- A vendor offers a steep discount on organisation-wide licences if the contract is signed this month.
- A pilot has run for six months with enthusiastic reports and no measured change in outcomes.
AI-2Governance, security and compliance
10 of 60 questions- AI-2.I1A certified AI Extension holder can trace where data goes when the team uses an AI tool or provider, and decide what may and may not be sent. team
- AI-2.I2A certified AI Extension holder can recognise prompt injection and excessive agency as design risks in any system that passes untrusted content to a model with tools, and require controls before release. team
- AI-2.I3A certified AI Extension holder can set AI usage policy in proportion to risk, so that low-risk use is easy and high-risk use is reviewed, rather than banning or permitting everything. org
- AI-2.I4A certified AI Extension holder can recognise when an AI use case may fall into a regulated category, such as a high-risk use under the EU AI Act, and involve legal or compliance early without making the legal judgment themselves. org
- AI-2.I5A certified AI Extension holder can treat the provenance, licensing and ownership questions raised by generated code as engineering risks with named owners. team
- AI-2.I6A certified AI Extension holder can keep an inventory of the AI systems in use and who owns each, so that accountability survives a change of team or vendor. exec
Situations questions present
- Engineers have been pasting customer records into a public chatbot to draft support replies.
- A support assistant with access to the ticketing system reads inbound customer email.
- Legal asks whether a new candidate-screening feature counts as a high-risk use.
AI-3AI-assisted software delivery
12 of 60 questions- AI-3.I1A certified AI Extension holder can set the conditions on which generated code is accepted, so that its author understands it, it is reviewed to the usual standard, and tests prove its behaviour. team
- AI-3.I2A certified AI Extension holder can decide what a coding agent may touch (repositories, secrets, networks, merge rights) by the damage it could do, and contain it accordingly. team
- AI-3.I3A certified AI Extension holder can adapt review practice when the volume of changes rises, so that review capacity, not generation speed, sets the pace of change. team
- AI-3.I4A certified AI Extension holder can keep a test strategy honest when tests are generated too, and recognise tests that assert what the code does rather than what it should do. team
- AI-3.I5A certified AI Extension holder can decide when an agent may work unattended, on which tasks, behind which pipeline checks and with which human checkpoint, and when it may not. org
- AI-3.I6A certified AI Extension holder can treat a team's fading understanding of its own codebase as a delivery risk, and protect its ability to debug and change what it ships. team
- AI-3.I7A certified AI Extension holder can put checks in the delivery pipeline (static analysis, dependency and secret scanning, policy rules) that apply to every change, whoever or whatever wrote it. org
Situations questions present
- A coding agent opened forty pull requests overnight and the team is unsure how to review them.
- The review queue has tripled since assistants were rolled out, and lead time is rising.
- A generated test suite reports high coverage and missed a defect that reached production.
- An engineer asks for an agent to be given merge rights so it can fix flaky tests on its own.
AI-4Building AI-powered products
12 of 60 questions- AI-4.I1A certified AI Extension holder can choose between prompting, retrieval, fine-tuning and a tool-using agent for a product need by what each fixes, what each costs and what each makes harder to change. team
- AI-4.I2A certified AI Extension holder can require an evaluation set that represents real use before a model or prompt change ships, and maintain it as an asset that goes stale. team
- AI-4.I3A certified AI Extension holder can judge evaluation evidence, including the limits of one model grading another, and tell an offline score from an outcome for users. team
- AI-4.I4A certified AI Extension holder can shape product behaviour around non-determinism, with safe failure, human review where an error is costly, and honesty with users about uncertainty. team
- AI-4.I5A certified AI Extension holder can weigh latency, unit cost and quality as one trade-off, and see how usage volume can turn a working prototype into an unaffordable product. org
- AI-4.I6A certified AI Extension holder can manage a model as a dependency, pinning versions, testing before upgrades, and keeping the option to change provider in proportion to the lock-in risk. org
- AI-4.I7A certified AI Extension holder can set honest delivery expectations for an AI feature, where the last share of quality often costs more than the first convincing demonstration. org
Situations questions present
- A product manager proposes fine-tuning a model to answer questions about prices that change every week.
- A demonstration impressed the board, and the team is asked to ship the feature in four weeks.
- The monthly model bill tripled in the first month after launch.
- The provider announces that the model version the product depends on will be withdrawn in ninety days.
AI-5Operating AI systems in production
10 of 60 questions- AI-5.I1A certified AI Extension holder can instrument an AI feature for quality, cost and safety signals as well as availability, so that a silent degradation becomes visible. team
- AI-5.I2A certified AI Extension holder can treat a change of model, prompt or retrieval index as a production change, with testing, staged rollout and a way back. team
- AI-5.I3A certified AI Extension holder can run an incident for an AI-specific failure (a false or harmful output, an injection, a runaway cost) with the discipline of an outage, including a blameless review. team
- AI-5.I4A certified AI Extension holder can design fallbacks and a kill switch so that the product degrades safely when the model or its provider fails. team
- AI-5.I5A certified AI Extension holder can put cost controls (budgets, alerts, rate limits) in place before cost becomes an incident. org
- AI-5.I6A certified AI Extension holder can notice when what a system now sees has drifted from what it was evaluated on, and decide when to evaluate it again. org
Situations questions present
- Users are sharing screenshots of a confidently false answer from the product.
- A provider outage has taken the product's main feature down.
- A prompt was changed in a vendor dashboard without review and support tickets rose the next day.
- A cost alert fires overnight because an agent is calling a model in a loop.
AI-6Capability, roles and measuring impact
8 of 60 questions- AI-6.I1A certified AI Extension holder can build a team's AI capability deliberately, through practice, shared conventions and time to learn, rather than by counting licences. team
- AI-6.I2A certified AI Extension holder can protect the development of early-career engineers when assistants do the work they used to learn from. org
- AI-6.I3A certified AI Extension holder can assess candidates and engineers on the judgment they show, with clear and fair rules about assistance during the assessment. team
- AI-6.I4A certified AI Extension holder can measure AI's effect in the team's own conditions with outcome measures, treat self-reported speed-ups as unreliable, and hold contested productivity claims as contested. org
- AI-6.I5A certified AI Extension holder can speak honestly about AI's effect on roles and headcount, without promising what the organisation cannot keep. exec
Situations questions present
- Leadership plans to hire a third fewer engineers next year on the strength of AI.
- Early-career engineers produce working changes quickly but cannot explain them in review.
- A survey says engineers feel twice as fast with assistants, and the delivery metrics have not moved.
- Several engineers ask in a team meeting whether their jobs are safe.
What the examination does not assess
- The baseline AI judgment of core competency TJ-5, AI-assisted engineering. The prerequisite ladder credential certifies it, and this examination assumes it.
- Writing code, prompts, configuration or queries. Questions present technical detail as context for a decision; they never ask the candidate to produce it.
- Machine-learning theory, training methods and model internals, beyond the mechanism a manager needs to judge a decision.
- Named products, models, vendors and their current figures (prices, context lengths, benchmark scores). Questions are vendor-neutral.
- Legal interpretation. Questions assess when to involve legal or compliance, not what an article of a regulation requires.
- Recall of models, frameworks, definitions or bibliography. No question asks the candidate to name a technique or define a term.