Most exam guides only tell you what's covered. The official Salesforce's AI-201 exam guide does something more useful: it explicitly lists what you're not expected to know — fine-tuning or configuring LLMs, transformer architecture, Apex or Python basics, external AI tools, ROI calculations, white space analysis, or anything involving Marketing Cloud, Heroku, Commerce Cloud, MuleSoft, Tableau, Slack, Quip, or Industry Clouds. That's an unusually direct scope-setting move, and it matters because a lot of candidates walk in assuming this is a deep technical AI exam when it's actually a configuration-and-judgment exam about Salesforce's own Agentforce tooling.
Prompt Engineering and Agentforce Concepts are each weighted at 30%, meaning together they account for 60% of the entire exam — more than the other three domains combined. Prompt Engineering covers when to use Prompt Builder, which user roles manage and execute prompt templates, grounding techniques, and the full lifecycle of creating, activating, and executing a prompt template. Agentforce Concepts is broader still: how the reasoning engine actually powers an agent, working with standard versus custom topics and standard versus custom agent actions, managing agent adoption and user security, testing agents in Testing Center, and deploying an agent from sandbox to production. If you only have time to master two domains before exam day, these are the two — skipping ahead to the smaller cloud-specific domains before locking these down is a losing trade.
Agentforce and Data Cloud carries 20% and focuses on two specific mechanics: improving an agent's response accuracy using the Agentforce Data Library, and grounding agent responses with Data Cloud retrievers. It's a smaller domain than the top two, but it's not optional — retrieval grounding shows up as a recurring theme across scenario questions elsewhere in the exam too.
Agentforce and Service Cloud and Agentforce and Sales Cloud each carry 10%. Service Cloud content covers building an agent that answers from Knowledge articles, connecting an agent to a digital channel, and identifying the right generative AI feature for a given service scenario. Sales Cloud content is narrower and more name-specific: knowing when to reach for an SDR Agent versus a Sales Coach agent, and matching the correct generative AI feature to a sales scenario. These two domains are easy to under-study because they're each worth only 10% individually, but between them they're worth as much as the Data Cloud domain on its own.
Salesforce lists no formal prerequisite for AI-201, but the exam guide's audience description assumes working familiarity: navigating Salesforce tools and building custom actions or integrations, understanding Data Cloud retrievers, and having hands-on exposure to deploying agents from sandbox to production. In practice, most people who pass this exam already have some Salesforce platform experience going in — it's technically open to anyone, but it's not designed as a first Salesforce credential.
Sixty scored multiple-choice and multiple-select questions, plus up to five additional unscored questions mixed in that you won't be able to identify during the exam, in 105 minutes. You need 73% to pass. No printed or digital reference materials are allowed during the exam, and it's delivered either at a testing center or online under proctor supervision. This certification launched with a free registration promotion that ran through the end of 2025; expect a $200 exam fee and a $100 retake fee at this point, though Salesforce's fees and promotions shift, so confirm the current cost on Trailhead before you register rather than assuming an old number still holds.
Multiple-select questions are worth flagging specifically, since they trip up candidates used to single-answer formats: a question asking you to select all correct options is graded differently than a straightforward multiple-choice item, and guessing partial credit doesn't work the way it might on other platforms' exams.Get additional practice with our Certified AI Associate online practice test.
Given that Prompt Engineering and Agentforce Concepts carry 60% combined, don't treat the five domains as equal-priority checklist items — structure your practice time to mirror the weighting, spending roughly three practice sessions on those two domains for every one session on Service Cloud or Sales Cloud individually. When you hit questions on grounding techniques, resist the instinct to memorize a single "correct" grounding method; the exam tends to present scenarios where the right technique depends on the specific business requirement in the stem, not a fixed rule you can apply everywhere.
For the Sales Cloud domain specifically, drill the distinction between an SDR Agent and a Sales Coach agent until it's automatic — these are two specifically named Agentforce Sales Agents with different jobs, and confusing them under time pressure is an easy, avoidable point to lose in a domain that's already small.
The most common mistake is over-preparing on general generative AI theory because the certification has "AI" in its parent credential name. Salesforce's own exam guide rules out LLM fine-tuning, transformer architecture, and external AI tools explicitly — time spent there doesn't transfer to exam performance. The exam tests configuration judgment inside Salesforce's specific Agentforce framework, not AI theory in the abstract.
A second mistake is under-preparing for the deployment and testing mechanics inside Agentforce Concepts. It's easy to focus study time on "how agents work" conceptually and skip the more procedural pieces — using Testing Center to evaluate an agent, and the actual steps of deploying from sandbox to production — but these are explicitly named objectives worth points inside the highest-weighted domain.
A third mistake specific to this exam: treating the 10%-weighted Service Cloud and Sales Cloud domains as afterthoughts because they're individually small. Between them they're 20% of the exam — the same weight as the entire Data Cloud domain — and candidates who run out of study time and cut corners there are cutting a fifth of the exam, not a rounding error.
Build a working glossary of Agentforce's own vocabulary before you touch practice questions: Topic, Action, Instructions, Plan, Reason, Ground, and the Trust Layer are terms the exam assumes fluency in, and questions often test whether you can apply the right term correctly in a scenario rather than define it in isolation. Pair the official exam guide with actual hands-on time in a sandbox or Developer Edition org building a agent end to end — this exam rewards configuration experience more than reading comprehension, and the deployment and testing objectives in particular are hard to answer confidently without having clicked through the process yourself at least once.
Read every scenario question for its specific business requirement before jumping to an answer. A recurring exam pattern across Prompt Engineering, grounding, and generative AI feature selection is that the "correct" answer depends entirely on what the scenario actually needs — dynamic data fetching, static content, a specific cloud's feature set — not on which option sounds most advanced or comprehensive.
Work through the practice questions below weighted the way the exam is weighted — more time in Prompt Engineering and Agentforce Concepts, proportionally less in the three smaller domains — and pay attention to whether your misses cluster around scenario-based judgment calls or around specific named features and tools. The first points to needing more hands-on sandbox time; the second points to gaps in your Agentforce vocabulary that a focused glossary review can close quickly.
Last updated on Oct, 9 2026
