Overview
Status first: the Salesforce Certified AI Associate exam is retired. Salesforce stopped delivering it on 1 May 2025 and retired the credential itself on 2 February 2026, so it can no longer be earned and existing holders now see it marked Retired on their Trailblazer profile. Salesforce now steers AI learners toward the Agentforce Specialist certification and the Agentblazer Status learning path.
This guide keeps the final outline for people who hold the credential, who studied for it, or who want the foundation before an Agentforce exam. AI Fundamentals had two objectives: explain the basic principles and applications of AI within Salesforce, and differentiate between the types of AI and their capabilities. That means knowing how machine learning finds patterns in historical records, how a model is trained and then used for predictions, and how predictive AI, generative AI, natural language processing and computer vision each produce a different kind of output.
How it's tested
Questions in this section were short and definitional, usually wrapped in a simple business scenario. A sales or service team wants a certain outcome, and you name the type of AI that produces it, or you explain in plain terms how a model learns and what it needs before it can predict anything.
-
Name the AI type
A use case such as forecasting churn, drafting an email or reading a scanned form is described, and you pick predictive, generative, NLP or computer vision.
-
How learning works
You explain that a model is trained on labeled historical examples and then applies the learned pattern to new records it has never seen.
-
Output shape
Options differ by whether the AI returns a score, a category, a number or new content, and only one output fits the stated goal.
A service manager wants each new case tagged with the most likely product area based on thousands of past cases already tagged by agents. The output is one label from a fixed set, and labeled history exists, so this is supervised classification, a predictive task. A generative model could write a summary of the case, but the requirement is a category, not new text, so predictive classification is the fit.
Objectives 2
Official objectives from the Salesforce outline, in outline order.
1.1 Explain the basic principles and applications of AI within Salesforce.What AI and machine learning are, and where Salesforce applies them.
- What it means
- AI is software that performs tasks that normally need human judgment, such as spotting patterns, making predictions or producing language. Machine learning is the main technique: instead of hand-written rules, a model learns relationships from historical data. Within Salesforce, these principles show up as scores on leads and opportunities, recommended next actions, case classification, and generated replies and summaries.
- What to understand
- Know the basic lifecycle: collect and prepare data, train a model on examples, evaluate it, deploy it, then monitor it as new data arrives. Understand that predictions are probabilities, not certainties, and that a model only knows what its training data showed it. Recognize that Salesforce places AI inside the record page and workflow, so users act on a score or draft where they already work rather than in a separate tool.
- How it shows up
- Expect plain-language questions on what a model is, why training data matters, and which business task an AI feature supports.
- Common confusion
- Treating an AI prediction as a certain fact rather than a probability the user should weigh.
1.2 Differentiate between the types of AI and their capabilities.Predictive versus generative AI, plus language and vision capabilities.
- What it means
- Different AI types produce different outputs. Predictive AI estimates a number, a likelihood or a category from past data. Generative AI creates new content such as text, code or images, usually through a large language model. Natural language processing lets systems read and interpret human language, and computer vision interprets images. Learning styles also differ: supervised, unsupervised and reinforcement learning.
- What to understand
- Match each type to its output. Classification returns a label, regression returns a numeric value, clustering groups similar records without labels, and generative models produce new content from a prompt. Understand that large language models predict the next likely word, which is why they can sound confident while being wrong. Know that many real features combine types, such as NLP reading an email and a predictive model scoring the sender.
- How it shows up
- Scenarios describe a goal and ask which AI type or learning approach fits; the output requested is the deciding clue.
- Common confusion
- Assuming generative AI replaces predictive AI, when each answers a different kind of question.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Machine learning
Models that learn patterns from data instead of following hand-coded rules.
- Remember
- Quality and quantity of training examples drive accuracy.
- Tested as
- Underlies nearly every Salesforce AI feature in the outline.
- Trap
- Calling any automation rule or Flow machine learning.
-
Predictive AI
Estimates likelihoods, values or categories from historical records.
- Remember
- Outputs a score, number or label, not new content.
- Tested as
- Lead scoring and churn likelihood are classic examples.
- Trap
- Picking predictive AI when the task is to write text.
-
Generative AI
Creates new text, images or code in response to a prompt.
- Remember
- Output can be fluent yet factually wrong.
- Tested as
- Email drafts and case summaries are generative tasks.
- Trap
- Assuming generated output is always grounded in company data.
-
Large language model (LLM)
A generative model trained on huge text corpora to predict likely next words.
- Remember
- It predicts plausible language; it does not look facts up unless grounded.
- Tested as
- Explains both the strengths and the hallucination risk of generative features.
- Trap
- Thinking an LLM queries the CRM by default.
-
Supervised learning
Training on examples that already carry the correct answer, or label.
- Remember
- Needs labeled history, such as won and lost opportunities.
- Tested as
- Most CRM scoring models are supervised.
- Trap
- Confusing it with clustering, which needs no labels.
-
Unsupervised learning
Finding structure, such as clusters, in unlabeled data.
- Remember
- No target field is required.
- Tested as
- Used for customer segmentation questions.
- Trap
- Expecting it to predict a specific yes/no outcome.
-
Reinforcement learning
Learning by trial and feedback toward a reward.
- Remember
- Improves through rewards and penalties over many attempts.
- Tested as
- Appears as a definition item rather than a CRM feature.
- Trap
- Mixing it up with users giving thumbs-up feedback on a single answer.
-
Natural language processing
Techniques that let software interpret and produce human language.
- Remember
- Covers intent detection, sentiment and entity extraction.
- Tested as
- Behind bots that understand typed questions.
- Trap
- Treating NLP and generative AI as the same thing.
-
Computer vision
AI that interprets images or video.
- Remember
- Classifies or detects objects in pictures.
- Tested as
- Fits scenarios like reading product photos or shelf images.
- Trap
- Choosing it for text-only tasks.
-
Training versus inference
Training builds the model; inference applies it to new records.
- Remember
- Inference runs every time a score is shown.
- Tested as
- Helps explain why stale models drift.
- Trap
- Thinking a model keeps learning automatically during inference.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Predictive AI vs generative AI
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Predictive AI | Forecasts an outcome or value from patterns in past data. | Returns a score, label or number. | Using it to draft a customer reply. |
| Generative AI | Produces new content such as text or images from a prompt. | Returns original content that did not exist before. | Using it to rank leads by conversion likelihood. |
| Combined use | A prediction selects the customer; generative AI writes the outreach. | Each type handles a different step. | Assuming one feature must do both jobs. |
Learning approaches
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Supervised | Learns from labeled examples. | Needs a known outcome field in history. | Applying it without any labeled outcomes. |
| Unsupervised | Finds groups or anomalies without labels. | No correct answer is supplied in training. | Expecting a specific prediction target. |
| Reinforcement | Learns from rewards over repeated actions. | Feedback arrives after actions, not as labels. | Equating it with ordinary supervised retraining. |
Prediction task types
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Classification | Assigns one of a set of categories. | Output is a label such as likely or unlikely to churn. | Calling it regression because a probability is shown. |
| Regression | Estimates a continuous number. | Output is a value such as expected deal amount. | Choosing it when the answer is a category. |
| Clustering | Groups records by similarity. | No predefined categories exist. | Treating clusters as validated segments without review. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Generative for everything
- What candidates pick
- A generative AI answer for a lead-ranking scenario.
- Why it looks right
- Generative AI dominates current Salesforce messaging.
- Why it's wrong
- Ranking leads needs a likelihood score from historical outcomes, which is a predictive task.
Clue to watch for The scenario asks for a score, ranking or likelihood.
-
Rules mistaken for learning
- What candidates pick
- Calling an if-then automation machine learning.
- Why it looks right
- Both automate a decision.
- Why it's wrong
- Rules are written by people and never learn from data; machine learning derives the pattern from examples.
Clue to watch for The logic is fixed and written by an admin.
-
Labels not needed
- What candidates pick
- Supervised learning for a dataset with no recorded outcomes.
- Why it looks right
- Supervised learning is the most familiar term.
- Why it's wrong
- Without a known outcome field there is nothing for a supervised model to learn to predict.
Clue to watch for The data has no outcome or target field.
-
LLM as a database
- What candidates pick
- Saying an LLM retrieves exact customer facts by itself.
- Why it looks right
- LLM answers read fluently and confidently.
- Why it's wrong
- A language model predicts likely text; it needs grounding in trusted data to state customer facts reliably.
Clue to watch for A question about why a generated answer was wrong.
-
Vision versus language
- What candidates pick
- Computer vision for interpreting customer emails.
- Why it looks right
- Both involve reading something.
- Why it's wrong
- Emails are text, so natural language processing is the relevant capability.
Clue to watch for The input is written language, not images.
Answer framework
Use these steps to classify any AI scenario in this section by the output it needs.
- Identify the output the business wants: a score, a category, a number, grouped records, or new content such as a draft or summary.
- If the output is new content, choose generative AI and note that it will need grounding and human review to be trustworthy.
- If the output is a label or likelihood, check whether labeled history exists; if it does, the answer is supervised predictive AI.
- If there are no labels and the goal is discovery, such as segments or outliers, choose unsupervised learning.
- Check the input type last: text points to natural language processing and images point to computer vision, regardless of the learning approach.
Practice question
Want more like this? The SF AI Associate practice test has questions for every domain.
Checklist
0 of 11
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Because the AI Associate exam is retired, study this section as a foundation rather than for a booking date. Spend about four to five hours here. Start with Salesforce's Trailhead AI Fundamentals content and the AI Associate cert-prep module, which remain useful even after the exam was withdrawn. Build a one-page table with three columns: AI type, typical output, and a CRM example. Fill rows for predictive classification, regression, clustering, generative text, NLP and computer vision.
Next, write a short explanation of the model lifecycle in your own words, including why labeled history is required for supervised learning and why models drift as data changes. Then practice sorting ten made-up business requests into the right AI type within a minute each. Finally, read a short primer on how large language models produce text so that hallucination makes sense, because Domain 3 and Domain 4 both build on it. If you are preparing for the Agentforce Specialist exam next, treat this section as the vocabulary check before moving on to agents, prompts and actions.
FAQ
How much of the Salesforce AI Associate exam was AI Fundamentals?
AI Fundamentals carried 17% of the exam, which works out to about seven of the 40 questions. It was the third-largest section, behind Ethical Considerations of AI at 39% and Data for AI at 36%.
Can I still take the Salesforce AI Associate exam?
No. Salesforce ended exam delivery on 1 May 2025, after closing registration on 31 March 2025, and retired the credential on 2 February 2026. It can no longer be earned.
What happens to people who already hold the AI Associate credential?
Reports of Salesforce's retirement FAQ say holders' Trailblazer profiles now show the credential as Retired and it no longer appears on the Certification Verification page. Your learning remains valid; the credential simply is not active.
What did Salesforce recommend instead?
Salesforce pointed AI learners to the Agentforce Specialist certification, formerly called AI Specialist, and to Agentblazer Status, a non-exam learning path. It did not name a one-to-one replacement for the AI Associate.
Is this material still worth studying?
Yes, as background. The difference between predictive and generative AI, the role of training data and the basics of language models are assumed knowledge in later Agentforce study.
Did this section require coding or math?
No. The objectives asked you to explain principles and differentiate AI types. Questions used plain business scenarios rather than formulas or code.
Which AI types should I know by name?
Predictive and generative AI, supervised, unsupervised and reinforcement learning, classification, regression and clustering, plus natural language processing, computer vision and large language models.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
Try practice questions that ask you to match business requests to predictive, generative, language or vision AI, and check your grasp of how models learn.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the SF AI Associate AI tutor Get any concept explained your way