Salesforce

AI Associate
Certification

Study Guide & Notes

Jo Crosier – September 2023

AI Associate Exam Overview
The AI Associate Credentialing Exam will focus on the fundamentals and About the Exam
capabilities of AI and data as they apply to CRM. The exam will be
intended for individuals from all backgrounds, including business leaders
and professionals who may have knowledge of AI, whether they are
⁇ Content: 40 multiple-choice questions
beginners or individuals with more experience. This certification will ⁇ Format: Select one correct answer from 3
validate foundational skills in responsible AI and data use, paving the
way for advanced technical certifications in the same domain. Time allotted to complete the exam: 70 minutes
Passing score: 65%
Exam Outline
💰 Registration fee: USD 75 plus applicable taxes
AI Fundamentals – 17% 🆓 Retake fee: Free
• Explain the basic principles and applications of AI within
Salesforce. 💻 Delivery options: Proctored exam delivered onsite
• Differentiate between the types of AI and their capabilities. at a testing center or in an online proctored
AI Capabilities in CRM – 8% environment. Click here for information on
• Identify CRM AI capabilities. scheduling an exam.
• Describe the benefits of AI as they apply to CRM.
Ethical Considerations of AI – 39% 📚 References: No hard-copy or online materials may
• Describe the ethical challenges of AI (e.g., human bias in be referenced during the exam.
machine learning, lack of transparency, etc.).
• Apply Salesforce's Trusted AI Principles to given scenarios. 👨🎓 Prerequisite: None
Data for AI – 36%
• Describe the importance of data quality.
• Describe the elements/components of data quality.
AI Associate Study guidance
Fast Track study plan Salesforce recommended preparation

5 hours 15 mins 26 hours

Prepare for your Salesforce Certified AI Associate Credential

Rather than completing the full recommended Trailmix, I suggest
that you can cover the information you need by completing the
Cert Prep: Salesforce AI Associate
following modules.

Artificial Intelligence Fundamentals

Generative AI Basics
Generative AI vs Predictive AI
Generative AI: 5 Guidelines for Responsible Development Other resources
AI Ethics Maturity Model
Ethical Data Use in Personalisation Technical Learning journey – complete the modules listed
Responsible Creation of Artificial Intelligence under ‘Scale Now with AI + Data + CRM’ and ‘Begin Your
Salesforce Einstein Basics Technical AI + Data + CRM Learning Journey
Data Quality
Is Your Data AI Ready?
Dinesh Yadav blog
Data Quality Dimensions Cheat Sheet
Cert Prep: Salesforce AI Associate
 AI Fundamentals Generative AI Capabilities
Exam weight 17% (7 Questions) Main Types of AI Capabilities Generative AI
Objectives Summarization – e.g. condensing a
Numeric Predictions Generative AI is technology that takes a
• Explain the basic principles and applications of AI within paragraph into one or two sentences
o probability, numeric values – e.g. next set of data and uses it to create something
Salesforce. quarter’s sales
• Differentiate between the types of AI and their new – like poetry, a physics explainer, an Translation – language translation using
capabilities. email to a client, an image, or new music – grammatical rules
Classifications when prompted by a human.
o identifying plants, phishing emails, Error correction – detect (and auto-correct)
fraudulent transactions errors in written text
Large Language Models (LLM) – AI models
trained on large amounts of data
Artificial Intelligence Technologies Robotic Navigation Question answering – ask a question – e.g.
o e.g., AI powered cars, assembly lines top 3 works of Shakespeare
One of the biggest LLMs is the GPT language
model, which stands for generative pre- Guided image generation – describe the
• Machine learning uses various Language Processing
image and AI will make it for you
trained transformer. In other words, a model
mathematical algorithms to get o Natural language processing –
that’s already been trained that can be used
insights from data and make translation, summarising documents, to generate text-related content. Text-to-speech
generative AI
predictions.
• Deep learning uses a specific type
of algorithm called a neural network Neural networks Common Concerns about Generative AI
to find associations between a set of Weight – how strongly an input should affect the output Structured data – e.g. spreadsheet. Hallucinations – predictions that diverge
inputs and outputs. Bias – adjustments to help get a more accurate estimate from expected results. Usually caused by
à Supervised learning incomplete or biased training data or if the
Node – unique combination of factors
• Natural language processing is a Layer – nodes connecting other nodes to each other Unstructured data - e.g. news article model was not designed well
technology that enables machines à Unsupervised learning Data security – businesses share data when
to take human language as an input Training AI by adding extra layers to find hidden fine tuning a model and when using the
model to process a request. Important that
and perform actions accordingly. meaning in data is what’s called deep learning.
Natural language understanding (NLU) this data is protected
• Computer vision is technology that Data processed from unstructured to
structured Plagiarism – as models are trained on
enables machines to interpret visual publicly available data they may learn to
information. Natural language generation (NLG) copy a particular style
• Robotics is a technology that Data processed from structured to User spoofing – easy to create a believable
unstructured online profile and interact with real users
enables machines to perform
Sustainability – models using large amounts
physical tasks. of computer power have big carbon
footprints
AI Fundamentals Semantic analysis

Generative AI vs Predictive AI
Sentiment analysis Intent analysis Context (discourse) analysis
Determines whether Virtual assistants, Understanding the surrounding
a piece of text customer support information to make sense of a
Generative AI Predictive AI
expresses a positive, systems or chatbots often piece of text. E.g. ”They had a
Creativity Creative and produces things Lacks the element of content negative or neutral use intent analysis to ball” context analysis can
that have never existed before creation
sentiment understand user requests determine if a dance party,
and provide appropriate sports equipment or a whole
Inferring the future Spots patterns but combines Uses historical and current data to responses or actions lot of fun
them into unique new forms spot patterns and extrapolate
potential futures

Different algorithms Uses complex algorithms and Uses statistical algorithms and Syntactic parsing
deep learning to generate new machine learning to analyse data
content based on the data it is and make predications
trained on • Segmentation: Larger texts are divided into smaller, meaningful chunks.
• Tokenization: Sentences are split into individual words, called tokens.
• Stemming: Words are reduced to their root form, or stem. For example breaking, breaks, or
unbreakable are all reduced to break. Stemming helps to reduce the variations of word forms,
but, depending on context, it may not lead to the most accurate stem. Look at these two
examples that use stemming:
GANs are made up of two neural networks: a “I’m going outside to rake leaves.”
generator and a discriminator. The two Natural Language Processing
Stem = leave
networks compete with each other, with the “He always leaves the key in the lock.”
generator creating an output based on some Natural language processing (NLP), is a field
Stem = leave
input and the discriminator trying to determine of artificial intelligence (AI) that combines
• Lemmatization: Similar to stemming, lemmatization reduces words to their root, but takes the
if the output is real or fake. The generator then computer science and linguistics to give
computers the ability to understand, part of speech into account to arrive at a much more valid root word, or lemma. Here are the
fine-tunes its output based on the same two examples using lemmatization:
discriminator’s feedback, and the cycle interpret, and generate human language in a
“I’m going outside to rake leaves.”
continues until it stumps the discriminator. way that’s meaningful and useful to humans.
Lemma = leaf
“He always leaves the key in the lock.”
Transformer models, like ChatGPT, (which
Parsing – breaking down text or speech
Lemma = leave
stands for Chat Generative Pretrained into smaller parts to classify them for NLP
• Part of speech tagging: Assigns grammatical labels or tags to each word based on its part of
Transformer), create outputs based on speech, such as a noun, adjective, verb, and so on. Part of speech tagging is an important function
sequential data (like sentences or paragraphs) Syntactic parsing - elements of natural
in NLP because it helps computers understand the syntax of a sentence.
rather than individual data points. This language are analyzed to identify the
• Named entity recognition (NER): Uses algorithms to identify and classify named entities–like
approach helps the model efficiently process underlying grammatical structure
people, dates, places, organizations, and so on–in text to help with tasks like answering questions
context and is why it’s used to generate or and information extraction.
translate text. Semantic parsing which derives meaning.
AI Fundamentals
AI The broad concept of having machines think and act like humans Hallucination A hallucination happens when generative AI analyzes the content we give it, but comes to an
erroneous conclusion and produces new content that doesn’t correspond to reality.
Artificial neural network A computer program that mimics the way human brains process
information. Large Language Model An LLM is a type of artificial intelligence that has been trained on a lot of text data. It’s like a
(LLM) really smart conversation partner that can create human-sounding text based on a given
Augmented intelligence The melding of people and computers to get the best of both worlds prompt. Some LLMs can answer questions, write essays, create poetry, and even generate code.

Machine Learning Machine learning is how computers can learn new things without being programmed to do
CRM with AI Generative AI can make CRM even more powerful — e.g. emails pre-written them. machine learning models learn from labeled data to make accurate predictions and
for sales teams, ecommerce product descriptions written based on images decisions.
alone, marketing campaign landing pages, contextual customer service
ticket replies, and more. Machine Learning Bias When computers are fed biased information, they make biased decisions. This can be the result
of a deliberate decision by the humans feeding the computer data, by accidentally incorporating
Deep learning Deep learning is an advanced form of AI that helps computers become biased data, or when the algorithm makes wrong assumptions during the learning process,
really good at recognizing complex patterns in data. It mimics the way our leading to biased results.
brain works by using what’s called layered neural networks
Model This is a program that’s been trained to recognize patterns in data. You could have a model that
Discriminator (in GAN) In a GAN, the discriminator is like a detective. When it’s shown pictures (or predicts the weather, translates languages, identifies pictures of cats, etc.
other data), it has to guess which are real and which are fake. The
discriminator’s job is to get better at telling real from fake, while the Natural Language NLP is a field of artificial intelligence that focuses on how computers can understand, interpret,
generator tries to get better at creating fakes. Processing (NLP) and generate human language. It’s the technology behind things like voice-activated virtual
assistants, language translation apps, and chatbots.
Ethical AI maturity model An Ethical AI maturity model is a framework that helps organizations assess
and enhance their ethical practices in using AI technologies. Prompt engineering It’s carefully crafting or choosing the input (prompt) that you give to a machine learning model
to get the best possible output.
Explainable AI (XAI) Explainable AI (XAI) should provide insight into what influenced the AI’s
results, which will help users to interpret (and trust!) its outputs. Sentiment analysis Sentiment analysis involves determining the emotional tone behind words to gain an
understanding of the attitudes, opinions, and emotions of a speaker or writer.
Generative Adversarial One of two deep learning models, GANs are made up of two neural
networks: a generator and a discriminator. The two networks compete with Supervised learning Supervised learning is when a model learns from examples. It’s like a teacher-student scenario:
Network (GAN)
each other, with the generator creating an output based on some input, and the teacher provides the student (the model) with questions and the correct answers. The
the discriminator trying to determine if the output is real or fake. The student studies these, and over time, learns to answer similar questions on their own.
generator then fine-tunes its output based on the discriminator’s feedback, Transformers are a type of deep learning model, and are especially useful for processing
Transformer
and the cycle continues until it stumps the discriminator. language. They’re really good at understanding the context of words in a sentence because they
Generative AI Generative AI is the field of artificial intelligence that focuses on creating create their outputs based on sequential data (like an ongoing conversation), not just individual
new content based on existing data data points (like a sentence without context).

Generator A generator is an AI-based software tool that creates new content from a Unsupervised learning Unsupervised learning is letting AI find hidden patterns in your data without any guidance. This
request or input. It will learn from any supplied training data, then create is all about allowing the computer to explore and discover interesting things on its own.
new information that mimics those patterns and characteristics. ChatGPT by
Validation In machine learning, validation is a step used to check how well a model is doing during or after
OpenAI is a well-known example of a text-based generator. the training process. The model is tested on a subset of data (the validation set) that it hasn’t
Generative pre-trained GPT is a neural network family that is trained to generate content. GPT seen during training, to ensure it’s actually learning and not just memorizing answers.
transformer models are pre-trained on a large amount of text data, which lets them In machine learning, ZPD is when models are trained on progressively more difficult tasks, so
Zone of Proximal
(GPT) generate clear and relevant text based on user prompts or queries. they will improve their ability to learn.
Development (ZPD)
 AI Capabilities in CRM
Exam weight 8% (3 Questions)
Objectives AI is already an integral part of the Customer 360 platform, and Einstein AI
technologies deliver nearly 200 billion predictions every day across Salesforce’s
• Identify CRM AI capabilities. business applications, including:
• Describe the benefits of AI as they apply to CRM.

• Sales, which uses AI insights to identify the best next steps and close deals faster.
Einstein allows all Salesforce users to: • Service, which uses AI to have human-like conversations and provide answers to
repetitive questions and tasks, freeing up agents to handle more complicated
• Discover insights that bring new clarity about your company’s customers. requests.
• Predict outcomes so your users can make decisions with confidence. • Marketing, which leverages AI to understand customer behaviour and personalize
the timing, targeting, and content of marketing activities.
• Recommend the best actions to make the most out of every engagement. • Commerce, which used AI to power highly personalized shopping experiences
• Automate routine tasks so your users can focus on customer success. and smarter ecommerce.

• New: Generate tailored content - from emails, to knowledge articles, to code.

Einstein Discovery Einstein Prediction Einstein Next Best Service AI Sales AI Commerce AI
Builder Action
• Enables businesses to explore • Personalize the end-to-end • Einstein-powered email: • Goals-based commerce: Set
patterns, relationships, and Einstein Prediction Builder Einstein Next Best Action service experience You can auto-create goals and get actionable,
correlations in historical data. is a simple point-click (NBA) allows you to use rules- • Improve agents' and mobile personalized and trusted AI insights
• Through the power of machine wizard that allows you to based and predictive models workers’ productivity impactful emails at scale • Einstein-powered product
learning and AI, Einstein can also make custom predictions to provide anyone in your • Keep information up to date, • Call summaries: descriptions: Create, localize,
on your non-encrypted business with intelligent, accurate, and available summarizing key and translate product
predict future outcomes, which
Salesforce data, fast. You contextual recommendations mentions, transcripts, and descriptions to build dynamic
allow business users to prioritize
can create predictions for and offers. Actions are Marketing AI action items automatically customer journeys.
their workloads and make data-
any part of your business— delivered at the moment of from previous calls. • Einstein-powered commerce
driven decisions.
across sales, service, maximum impact—surfacing • Content creation • Sales Assistant: This concierge: Reimagine product
Einstein Lead Scoring • Scaling AI central hub helps you discovery with a conversation-
marketing, commerce, IT, insights directly within
Uses data science and machine • Conversational interface prep for meetings and based interface that relies on
finance, and even HR—with Salesforce.
with technology. provides other selling your data to bring in-store
learning to discover your clicks, not code.
• Trusted content from your guidance. concierge experiences to life
business’s patterns of lead
CRM data on any digital channel.
conversion. Based on your • Einstein-created promotions
patterns, Einstein predicts which
of your current leads to prioritize
 Ethical Considerations of AI RA(TE)I
Exam weight 39% (16 Questions)
Potential risks to Generative AI
Objectives 1. Accuracy–a model might be able to create a new
sentence in the style of a famous writer, there isn’t any
• Describe the ethical challenges of AI (e.g., human way to know if the same sentence is factually true. That
bias in machine learning, lack of transparency, etc.). can be a problem when users assume that an AI's
• Apply Salesforce's Trusted AI Principles to given predictions are verified facts
scenarios. 2. Bias and Toxicity– Because human interactions can
involve a degree of toxicity—that is, harmful behaviour
like using slurs or espousing bigotry—AI replicates that
toxicity when not tuned to recognize and filter it.
3. Privacy and Safety– It’s easy to exploit the technology
to do huge amounts of damage very quickly. The models
have a tendency to “leak” their training data, exposing
private info about the people represented in it. And GAI
can even be used to create believable phishing emails or
replicate a voice to bypass security.
4. Disruption - Economic disruption, job and
responsibility changes, and sustainability concerns from
the intense computing power required for the models
to operate all have implications for the spaces we
share.

The biggest perceived risks of real-time Recommendations for Ethical Behavioral

personalization in marketing: Marketing

• Security events, like data breaches 1. Collect and Respect Preferences – Honour
• Data being collected, shared, or used in customer preferences and use only the data
unanticipated ways they’ve consented to sharing. Provide clear
• Personalizing interactions that feel invasive or controls to opt-in or out.
unwanted to consumers 2. Audience Targeting – Messaging should be
• Inadvertent bias introduced by relying on targeted based on consumer-expressed
demographic attributes for interactions instead interests, not demographics. View consumers
of behavioural and engagement data as they truly are: multi-dimensional individuals
with many varied idiosyncratic affinities. Highly
Responsible Marketing Principles individualized messaging is more effective than
targeting based only on demographic data.
• Be transparent about security
• Develop trust through consent and transparency 3. Frequency Capping – Overwhelming a customer
• Use data to personalise the experience with too many communications can subvert
• Prioritise behaviour-based intent over your brand. Frequent but unwanted messages Set of guidelines focused on the responsible development and implementation of generative AI.
demographic attributes can annoy your customers and drive them to
• Customers expect something in exchange
tune you out.
 Data for AI importance of data and data quality in AI.
Exam weight 36% (14 Questions)
Data types
Objectives • Training and performance: High-quality data ensures that the model learns accurate and
Qualitative representative patterns, leading to more reliable predictions and better decision-making.
• Describe the importance of data quality.
Can't be measured numerically, • Accuracy and bias: Biased or inaccurate data can lead to biased outcomes, reinforcing
• Describe the elements/components of data quality.
characteristics existing inequalities or perpetuating unfair practices. By ensuring data quality, organizations
• Nominal – categories that cannot be can strive for fairness and minimize discriminatory outcomes.
Data Quality Dimensions ranked – e.g. fruit
• Ordinal – Can be ranked – e.g. survey • Generalization and robustness: High-quality data ensures that the model learns relevant
responses, 5 point scale and diverse patterns, enabling it to make accurate predictions and handle new situations
Completeness effectively.
Validity Quantitative • Trust and transparency: Data quality is closely tied to the trustworthiness and
Can be measured numerically transparency of AI systems. Transparent data practices, along with data quality assurance,
Uniqueness • Discrete variables – you can count it help build trust and foster accountability.
Timeliness/Age individually – e.g. number of children in
• Data governance and compliance: Proper data quality measures are essential for
Consistency a household
• Continuous variables – e.g. time, maintaining data governance and compliance with regulatory requirements. Organizations
Accuracy temperature, distance, mass must ensure that the data used in AI systems adheres to privacy, security, and legal
standards.

Factors that determine data quality Types of data bias

Bad data is linked with:
Societal bias
• Missing records Association bias Reproduces past prejudice towards
• Lost revenue
• Duplicate records Data labelled according to stereotypes – historically marginalised groups
• Missing or inaccurate insights e.g. ’toys for girls’ – dolls, princesses
• No Data standards
• Wasted time and resources Survivorship bias
• Incomplete records Confirmation bias An algorithm focuses on the results of those
• Stale data • Inefficiency E.g. recommendations based on purchasing were selected, or who survived a certain
habits but the data influencing those purchases process, at the expense of those who were
• Slow info retrieval already reflect what people see and choose to
Steps to clean your data for AI: excluded.
• Poor customer service buy in the first place.
• Remove duplicates Interaction bias
• Reputational damage Automation bias
• Fix structural errors Humans create interaction bias when they
• Decreased adoption by reps Data set for training not representative. intentionally try to influence AI systems and
• Filter unwanted outliers E.g. Kodak calibrated skin tones, shadows, create biased results. E.g. when people
• Handle missing data light using fair skinned person intentionally try to teach chatbots bad
• Validate language.