What the AIF-C01 exam measures

What you will be able to do

• Name the five AIF-C01 domains and their published weights
• Describe the exam format, length, delivery and fee
• Explain why 65 items on the paper produce a score from 50 of them
• Explain what a scaled score of 700 means and why it is not 70% correct
• State plainly what this guide is and where the authoritative source lives

The most useful sentence in the whole AIF-C01 blueprint is the one describing who the exam is for. AWS says the target candidate is familiar with Artificial Intelligence (AI), machine learning and generative AI technologies on AWS, with up to six months of exposure to them, and that this candidate uses AI/ML solutions rather than necessarily building them.

That single word sets the depth of every question on the paper. You are expected to know that retrieval augmented generation exists, what problem it solves, which AWS service provides it and when it beats fine-tuning. You are not expected to write the ingestion pipeline. Candidates who fail this exam usually fail it in one of two ways: they never learned the service surface at all, or they studied it as though it were an engineering exam and ran out of time before covering the breadth.

The credential at a glance#

There are no prerequisites, which is worth stating plainly because people assume otherwise: no AWS certification requires another one first, so AIF-C01 can be your first AWS exam. The recommended experience is a recommendation, not a gate.

Ninety minutes for 65 items is about eighty seconds each. That is comfortable for a definition item and tight for a scenario that describes a system in a short paragraph before asking which service fits, so the time pressure on this exam comes from reading rather than from thinking.

Sixty-five items, fifty of which count#

Fifteen of the 65 items are unscored. AWS includes them to trial questions for future versions of the exam, and they are not identified on the paper. There is no marking, no separate section and no way to tell which fifteen they were.

Two consequences follow. First, the published domain weights apply to the 50 scored items, not to all 65, so any arithmetic you do about how many questions a domain is worth is arithmetic on 50. Second, and more practically, a question that looks bizarre or unreasonably deep is quite possibly one of the fifteen. Answer it and move on rather than letting it cost you two minutes and your composure.

Four question formats, and no partial credit#

The exam uses four formats:

The rule that matters across all of them: there is no partial credit. A multiple response item with three correct answers gives you nothing for two of them. An ordering item with one pair swapped scores the same as one you never read. Matching is all or nothing across every pair.

This changes how you should handle the last two formats in particular. On an ordering or matching item, the fastest safe method is to place the items you are certain about first, then use structure to constrain the rest: in an AI/ML lifecycle ordering, data collection cannot follow evaluation, and deployment cannot precede training. You are rarely reasoning about five independent choices.

Five domains, five weights#

Read this comparatively rather than as a list. Domains 2 and 3 together are 52% of the scored content, so generative AI and the things you build with foundation models are more than half the exam. Domain 3 alone outweighs domain 1 by eight points, and it is the domain that carries retrieval augmented generation, prompt engineering, fine-tuning and evaluation.

Domains 4 and 5 are equal at 14% each and are the most predictable sections on the paper, because each has only two task statements behind it. That makes them the cheapest points available: a smaller body of material stands behind a seventh of the exam.

If the weights map evenly onto the 50 scored items, that is roughly ten items for domain 1, twelve for domain 2, fourteen for domain 3 and seven each for domains 4 and 5. Treat that as arithmetic rather than a promise. AWS publishes weights, not item counts.

What a scaled score of 700 means#

700 out of 1,000 is not 70% correct. This is the single most misread number in AWS certification, and it matters more here than on most exams because 700 sits exactly where a percentage would sit and looks like one.

The raw proportion of items you need is not published. Prep sites that state "about 70%" are guessing, and they are guessing at a number that would in any case move between forms, which is the entire reason scaling exists. You cannot take a percentage from any practice test, including ours, and conclude you are over the line.

Compensatory scoring is the reason you should not panic about a weak domain, and also the reason you should not abandon one. Nothing forces you to pass domain 5 on its own, so a shaky area cannot fail you by itself. But domain 5 is 14% of the paper, and simply skipping it in preparation means starting roughly seven scored items down out of fifty on that arithmetic, which is most of the margin most candidates have.

What comes back afterwards#

The result is reported as pass or fail, together with a table of section-level performance classifications rather than a per-domain percentage. That table is what you work from after a fail: it tells you whether you were close everywhere or lost one domain badly, and those two outcomes call for completely different second attempts.

The credential is valid for three years. AWS states that it is recertified by passing the current version of this exam or by passing a higher-level role-based exam, and its own page notes that AWS Certified Machine Learning Engineer, Associate recertifies it automatically.

Three years is a long time in this subject. Most of what follows in this guide describes services that have been renamed, superseded or closed to new customers within the last two years, which is exactly what the next lesson is about.