How to Prepare for an AI Certification: Calculate the Exam Date First, Then the Study Order

🌏 中文版

Two things decide whether you pass a certification, and both happen before you open any study material: which day you book the exam for, and which chapter you study first. Most prep guides skip both and start teaching at chapter one.

This piece breaks those two decisions into four steps, across AWS, Google, Microsoft, NVIDIA, Databricks, and Taiwan's iPAS. It is the first post of the AI Certification Prep series: the other posts take one certification each and break down its exam guide, while this one covers what you have to work out first whichever exam you take. The rules and weights here are the official versions as of October 2026.

If you're still choosing which exam to take, start with what AI certifications engineers can take in 2026.

Step 1: Work Backwards From "What Happens If I Fail"

The exam date isn't "whenever I feel ready." It's calculated — and the calculation depends entirely on one thing: when you'd be allowed to sit it again.

Vendors differ far too much to carry one rule over to another:

Fourteen days is the industry standard — AWS, Google, Databricks, and NVIDIA all land on that number. The real outliers sit at both ends: Microsoft's 24 hours and iPAS's six months.

The AWS After Testing policy states:

If you fail an exam, you must wait 14 calendar days before you are eligible to retake the exam. There is no limit on exam attempts. However, you must pay the full registration fee for each exam attempt.

Google's exam terms escalate:

If you don't pass the exam, you can take it again after 14 days. If you don't pass the second time, you must wait 60 days before you can take it a third time. If you don't pass the third time, you must wait 365 days before taking it a fourth time.

Microsoft is the most forgiving, but only once. The official retake policy states:

If you don't pass an exam the first time, you must wait 24 hours before retaking it. A 14-day waiting period is imposed between all subsequent attempts (up to 5).

Databricks' exam terms read "There is a 14-day wait between all attempts. Payment is required each time you take an exam," and NVIDIA's certification FAQ is also 14 days, capped at "no more than five times per year."

Three Deadline Structures, Three Calculations

1. Sit any time, but with an external deadline — a time-limited discount code, a voucher expiry, an expense-claim cutoff. Work backwards:

latest first attempt = deadline − retake waiting period − score reporting buffer

Using AWS, with a deadline of passing by December 31:

The score buffer exists because the same AWS page notes "Final results will be posted to your AWS Certification Account within five business days." If the deadline requires passing rather than merely sitting the exam, those five business days belong in the math.

So the first attempt goes on 12/13, not "sometime in late December." One week separates "a second chance if I fail" from "out of luck if I fail."

2. Fixed sessions (iPAS). There are no days to subtract; the buffer is measured in sessions. Per the official exam information (in Mandarin), the 2026 intermediate exam runs twice — May 23 and November 14. Fail one and the next opportunity is six months away.

That changes the strategy entirely: there's no "postpone a week" option, so the stop-line decision has to move ahead of the registration deadline rather than sitting a week before the exam.

3. Sit any time, no external deadline. Easy to assume no math is needed, but a retake still costs full price — AWS and Google both say so explicitly. Write the total cost of one retake at the top of the plan; you'll use it when deciding whether to move the date earlier.

Step 2: Take an Official Practice Set Before Studying Anything

This is the least intuitive step and the most commonly skipped: before you study anything, work through free official questions.

The point isn't practice, it's diagnosis. What you want isn't the score, it's the table of how many questions you missed per domain — that table decides how the following weeks get allocated. Without it, your study order is just someone else's recommendation copied down.

Vendors are unequal here, and it directly affects whether you can do this at all:

iPAS is by far the most generous here, and most people don't know it: the learning resources page publishes complete past papers as PDFs for all three intermediate subjects, alongside official study guides and errata sheets. Using a recent real paper as your diagnostic beats any mock exam.

The Microsoft cell is a practical trap. The official note reads "Practice Assessments are usually available within 8 weeks of the exam being out of beta and generally available," so a freshly revised certification has a gap. The fallback is to walk the exam guide's task statements one by one and ask whether you could do each.

Step 3: Order by Weight, Not by Domain Number

Domain numbering reflects the exam writer's taxonomy, not your learning sequence. Order by weight.

Official weightings for seven certifications, from their current exam guides:

Study AIF-C01 by number and week one covers the 20% domain; study by weight and week one covers the 28% domain. The difference matters for one reason: if work derails the plan halfway through, which domain do you want already finished?

There's a second rule: put memorization-heavy domains last. On AIF-C01, responsible AI and security governance total 28% and read like compliance documentation, which engineers instinctively skim — yet they're pure memorization and the highest return per hour. Schedule them in the two weeks closest to the exam, while recall is freshest.

The third rule comes from the scoring model, and it differs by certification. The AWS exam guide states:

The exam uses a compensatory scoring model, which means that you do not need to achieve a passing score in each section. You need to pass only the overall exam.

Passing scores differ by exam: 700 for AIF-C01, 720 for MLA-C02, and 750 for AIP-C01 (out of 1,000). Since only the total counts, a weak domain doesn't need to be brought to full strength. If the diagnostic shows a 14% domain is especially poor, the right move isn't spending a week repairing it — it's putting that time into the 28% domain, where the same hours return more points.

The iPAS intermediate exam does not work this way. Subject 1 and your chosen second subject must each reach 70; failing either fails the whole thing. That's conjunctive, not compensatory: the weak subject has to be brought to the threshold, with no room to compensate from the stronger one. Applying the lean-toward-weight strategy here will sink you.

Step 4: Write Your Stop-Lines in Advance

The last item in the plan isn't the exam. It's the three lines that decide whether to sit it — written down before you start studying.

For a five-week plan:

• Week 4 practice exam below 55% → the gap is too large; postpone two weeks and rework the two highest-weight domains
• Week 5 practice exam below 70% → postpone one week and repair only what the practice exam flagged
• Week 5 practice exam at or above 80% → sit it on the original date, and stop studying

Why write them in advance: self-assessment three days before an exam is almost always too optimistic. By then you've invested five weeks, and sunk cost pushes you toward sitting it anyway. A number written down beforehand is the only thing that counters that.

The third line matters just as much. Past 80%, more practice questions return very little, and the extra anxiety costs you on exam day.

Fixed-session certifications need all three lines moved earlier. iPAS has no "postpone a week" option, so the decision point belongs before the registration deadline: use your practice score then to decide whether to register for that session at all, rather than registering and agonizing afterwards. The fee isn't refundable, but what you save is six months of misplaced expectation.

Worked Example: A Five-Week AIF-C01 Plan

Combining all four steps, assuming 7 hours a week and an exam date of 12/13:

Week 0 puts booking ahead of downloading material because seat availability and price are external conditions — finding a problem a week earlier buys a week to adjust.

That "concept → AWS service name" table deserves a note. Anyone who has already built RAG or LLM applications largely has the conceptual layer of the first two domains; what's missing is the vendor's naming. Managed foundation models are Amazon Bedrock, managed RAG is Bedrock Knowledge Bases, output filtering is Guardrails, bias detection is SageMaker Clarify, drift monitoring is Model Monitor, and human review is Amazon Augmented AI — the last four are named explicitly in the exam guide's responsible AI domain. For those readers, the weekly output should be a mapping table rather than notes: the mode is "I know this one, what does AWS call it," not learning concepts from scratch.

How Preparation Differs by Vendor

The four steps are universal; each vendor has its own thing to watch.

AWS: the official four-step flow starts with the exam guide and the free 20-question practice set, fills gaps with courses and labs, and ends with the Official Practice Exam as a readiness check (that last item needs a subscription). Work through the free tier before deciding whether to subscribe. What separates the three exams is hands-on exposure, not years on a résumé. For the domain-by-domain breakdown see AIF-C01, MLA-C02, and AIP-C01.

Google (PMLE): the main risk isn't difficulty, it's stale material. The certification page states that "This exam was updated to reflect the transition from Vertex AI to Gemini Enterprise Agent Platform," so everything published before mid-2026 uses the old vocabulary — take notes in the new names only. Google also notes that "the exam does not directly assess coding skill," so time belongs on architectural tradeoffs and service selection. The domain-by-domain breakdown is in the PMLE preparation path.

Microsoft (AI-103): the most free coursework — four Microsoft Learn learning paths, which are the material for the four-day instructor-led course AI-103T00. The practice assessment is on AI Skills Navigator and needs a sign-in to launch. If you plan to go further, AI-500 requires earning AI-103 first — they aren't parallel credentials. The series covers seven Microsoft exams; the trade-offs among four of them are in choosing among Microsoft's AI certifications.

NVIDIA (NCA-GENL): 50–60 questions in 60 minutes, about a minute each — this one tests fluency, not depth. The certification page lists recommended training under each weighted area, so the preparation path is to work down the list and tick boxes. One constraint to know upfront, though: the official FAQ states that "NVIDIA certification exams are pass/fail. You won't receive a score." There's no score report, so failing leaves you without a weakness breakdown and a second attempt runs on memory alone — which raises the value of diagnosing yourself properly beforehand. The trade-offs among the four are in choosing among NVIDIA's four.

Databricks: weight concentrates in application development (30%) and assembling and deploying (22%), so more than half is implementation. Preparation means building one RAG chain end to end on the platform — vector search, model serving, lifecycle management, data governance — rather than reading documentation. The official recommendation is 6+ months of hands-on experience.

iPAS intermediate: three characteristics shape preparation. First, complete past papers are public, so diagnose with real questions. Second, subjects 2 and 3 have included code-interpretation questions since the second 2025 session; the official notice (in Mandarin) says code-type questions are projected at roughly 25% of the paper, "in the form of single-choice questions and question sets" — note the question sets, where one fragment feeds several questions and one misreading costs you all of them. It's a question type where reading isn't the same as knowing; print the snippets and trace the values by hand. Third, check the intermediate subject-exemption rules in the official downloads section before you start: they may remove one subject from your workload entirely.

Common Mistakes

Using outdated material. The most expensive mistake, because you discover it at the test center. Ten minutes comparing service names against the official exam guide beats another hundred practice questions.

Buying a course before reading the exam guide. Backwards. Free resources are enough in most cases, and the only thing usually worth paying for is a full-length practice exam.

Treating years of experience as an eligibility requirement. The AWS Before Testing policy states that "All AWS Certifications may be earned without completing specific prerequisites" — the "2 years recommended" lines are difficulty labels, not gates. The reverse also holds: don't skip the diagnostic because your experience looks sufficient on paper.

Carrying one vendor's retake rule over to another. Microsoft is 24 hours, AWS is 14 days, Google jumps to 60 days on the second failure, and iPAS is six months. The same shrug — "I'll just take it again" — costs two hundred times more in one case than the other.

Not checking exam languages. Language lists can differ between certifications from the same vendor, and reading speed affects the outcome more than an extra week of study.

Where This Method Stops Working

It allocates time; it doesn't create knowledge. If the diagnostic shows you're unfamiliar with the entire subject area, that isn't a scheduling problem — build hands-on experience first rather than compressing five weeks into three.

That applies doubly to certifications testing operational detail: AWS's MLA-C02 expects you to have used Amazon SageMaker, and the Databricks exam expects you to have deployed on the platform. Grinding question banks to pass one of those has poor returns, and the value of the credential was in that hands-on familiarity to begin with.

What to Read Next

• Still choosing an exam: start with the vendor's chooser — AWS, Microsoft, NVIDIA, or Claude.
• Already decided: go straight to that exam's preparation path; the series page lists them all by vendor.
• Preparing for two or more: read the cross-exam comparisons to see what carries over — multi-agent architecture, RAG and retrieval evaluation, prompt and context engineering, cost and latency, and AI governance frameworks.

The series has no single-exam preparation path for Databricks or iPAS; their specs are in the overview table.

References

• AWS Certification: After Testing policy (retake waiting period, passing standards, compensatory scoring, score reporting)
• AWS Certification: Before Testing policy (no prerequisites)
• AWS Certified AI Practitioner (AIF-C01) certification page
• AIF-C01 official exam guide (domain weights, 700 passing score, compensatory scoring)
• AWS MLA-C02 official exam guide (domain weights, 720 passing score)
• AWS CLF-C02 official exam guide PDF
• AWS exam preparation overview (free practice question sets and Exam Prep courses)
• Google Cloud Exam Terms & Conditions (retake waiting periods of 14 / 60 / 365 days)
• Google Cloud Certification: Retake Policy
• Google Professional ML Engineer official exam guide PDF
• Google Professional ML Engineer certification page (exam update notice)
• Microsoft retake policy (24 hours for the first retake, 14 days thereafter, five per year)
• Microsoft AI-103 study guide (five skill-area weights)
• Microsoft AI-103 certification page (where the practice assessment lives)
• Microsoft course AI-103T00
• NVIDIA NCA-GENL certification page (Exam Blueprint)
• NVIDIA Certification FAQ (14-day retake wait, five per year, pass/fail with no score)
• Databricks Certified Generative AI Engineer Associate (domain weights)
• Databricks Certification Terms & Conditions (14-day retake wait, payment each time)
• iPAS AI Application Planner learning resources (past papers and official study guides) (in Mandarin)
• iPAS AI Application Planner exam information (sessions, subjects, passing conditions) (in Mandarin)
• Notice on iPAS intermediate code question weighting (in Mandarin)

Related on this site

• What AI certifications engineers can take in 2026
• Which of the three AWS certifications
• Choosing among Microsoft's AI certifications
• Choosing among NVIDIA's four
• Choosing among the four Claude certifications
• Which AI courses to take in 2026 (in Mandarin)