Artificial intelligence (AI) lets computers learn from data to make decisions or predictions, without being explicitly programmed for every single situation. For someone studying for the Salesforce AI Associate exam, understanding what AI is — and why ethical rules matter when using it — is the foundation for every other topic. Businesses use AI to speed up tasks, find patterns, and automate choices, but without ethical guardrails, those same tools can discriminate, mislead, or invade privacy.
Jump to a section
A simple way to picture Introduction to Artificial Intelligence and Ethical Principles
A real estate agent looks at thousands of house listings every day. The agent learns which features — granite countertops, a south-facing garden, good school catchment — tend to make a house sell quickly for a high price. Over time, the agent gets better at predicting which new listing will attract offers above the asking price. This learning from past examples is training. The agent’s gut feeling about a new listing is the predicted output. Now imagine a second agent who only looks at houses in one wealthy neighbourhood and ignores all others. That agent’s predictions will be biased — she’ll overvalue every house in that neighbourhood and miss bargains elsewhere. Her bias harms buyers who can’t afford that area. The ethical problem is clear: the data the agent trained on (only expensive houses) made her unfair. In business, if a company builds an AI tool to screen job applicants using only past hires (who were mostly men from one university), the AI will learn that pattern and unfairly reject women and graduates from other universities. The company must check the data for fairness, just as a good agency would train its agents on diverse listings. The core idea is simple: AI is pattern-learning at scale, and ethical principles are the rules we put in place so the patterns don’t harm people.
Artificial intelligence is a broad term for computer systems that can perform tasks which normally require human intelligence. These tasks include recognising speech, making decisions, translating languages, and spotting patterns in data. The key idea is that the system ‘learns’ from examples rather than following fixed, step-by-step instructions written by a programmer.
For a complete beginner, think of it this way: in traditional programming, you give the computer a recipe and it follows it exactly. If you change the ingredients, the recipe breaks. With AI, you feed the computer thousands of recipes and let it figure out what makes a good cake. The computer builds its own rules based on the patterns it sees. That’s called ‘machine learning’, which is a subset of AI.
The core components of AI in business are:
Data: The raw information the system learns from. This could be customer purchase histories, images, text documents, or sensor readings. The quality and quantity of data directly determine how well the AI performs.
Model: The mathematical structure that captures the patterns learned from the data. Think of it as the ‘brain’ of the AI. Once trained, the model can make predictions on new, unseen data.
Training: The process of showing the model many examples so it can adjust its internal parameters to reduce errors. For instance, training a model to recognise cats means showing it millions of labelled cat pictures until it learns what features (whiskers, pointy ears, fur) usually appear in a cat image.
Inference: When the trained model makes a prediction or decision on new data. This is the ‘live’ use of the AI — for example, a credit card company using a trained model to decide instantly whether a transaction is fraudulent.
Algorithm: The set of mathematical rules the model uses to learn from data. Different algorithms suit different problems — some are good at sorting emails into spam and not spam, others at recommending movies you might like.
Why does a business use AI? The main reasons are speed, scale, and consistency. A human can review 100 insurance claims in a day; an AI can review 100,000 in a minute without getting tired. AI also finds patterns humans would miss — like subtle correlations between shopping habits and product preferences that help a retailer stock the right items.
Now, ethical implications arise because AI systems are not inherently fair, transparent, or private. The model reflects the biases present in its training data. If a loan-approval model learns from historical data where a bank rejected more applications from people of a certain ethnicity, the model will likely continue that discriminatory pattern. This is ‘algorithmic bias’. Other ethical concerns include:
Privacy: AI often needs huge amounts of personal data to work well. Collecting and storing that data without proper consent or security can violate people’s privacy rights.
Transparency: Many AI models, especially complex ones called ‘deep neural networks’, act like a black box — even the engineers who build them cannot fully explain why a particular decision was made. In regulated industries like healthcare or finance, this lack of transparency is a serious problem.
Accountability: If an AI makes a harmful decision — for example, misdiagnosing a patient or approving a risky loan — who is responsible? The developer? The company that deployed it? The AI itself cannot be held accountable, so humans must stay in the loop.
Job displacement: AI can automate tasks once done by people, leading to job losses. However, it also creates new roles — like AI ethics officers and data labelers — so the net effect on employment is complex.
The Salesforce AI Associate exam expects you to recognise that AI is not magic or a single technology. It is a collection of tools that learn from data. Ethical principles are not optional add-ons; they are fundamental to building trustworthy AI systems. The exam will test whether you understand that a biased training dataset produces a biased model, regardless of how clever the algorithm is. You must also know that ethical AI requires human oversight, transparency, and fairness from the very beginning of the project lifecycle.
Define the business problem and goal
Before any AI work begins, the team identifies what problem they want to solve — for example, reducing customer churn, automating claim approvals, or personalising product recommendations. This step ensures the AI project has a clear, measurable objective and that the problem is actually suitable for AI (not just a simple rule-based task).
Gather and prepare training data
The team collects historical data related to the problem. This data must be cleaned (remove duplicates, fix errors) and checked for bias. For ethical AI, the team ensures the data represents all relevant groups fairly. If the data is imbalanced (e.g., 90% approvals from one demographic), they must balance it or acknowledge the limitation.
Train the AI model using an algorithm
A machine learning algorithm is selected based on the problem type (classification, regression, etc.). The model is trained on the prepared data, meaning it adjusts its internal parameters to reduce prediction errors. The team splits the data into training and test sets to check that the model generalises well to new examples.
Conduct an ethical review and fairness audit
The model is evaluated using fairness metrics (e.g., demographic parity, equal opportunity). If the model shows bias, the team must revise the training data, change the algorithm, or add rules to override biased outputs. This step is not optional; it must happen before any real-world use, and documentation of the review is kept for auditors.
Deploy the model with human oversight
The AI is put into production, but not left to run unsupervised. A human-in-the-loop design ensures that low-confidence decisions, high-stakes decisions, or flagged cases are reviewed by a person. The team also sets up monitoring to track the model's accuracy and fairness over time.
Monitor, maintain, and update the model
After deployment, the team continuously checks for model drift (when real-world patterns change) and for new bias complaints. If performance drops or fairness issues emerge, the model is retrained with new data or retired. This step closes the loop, making the AI lifecycle truly ongoing.
An IT professional working with AI in a business context typically starts by understanding the problem the business wants to solve. Let us walk through a concrete scenario: a mid-sized insurance company wants to automate its claims processing. The goal is to reduce the time it takes to approve or reject a claim from several days to a few minutes.
Step one — data gathering. The IT professional works with claims handlers to collect historical claim records. These records include details like claim amount, type of damage (car accident, storm damage, theft), customer history, and the final outcome (approved or denied). The IT professional checks this data for completeness and fairness. For example, are there significantly more claims from customers in affluent areas? If so, the data might be biased, and the model might learn to be more lenient with those customers. The IT professional flags this and works with the business to balance the dataset before training begins.
Step two — model selection and training. The team chooses an algorithm suitable for classification (approve/deny). They split the historical data into a training set (used to teach the model) and a test set (used to check accuracy). The IT professional monitors the training process, ensuring the model does not ‘overfit’ — that is, memorise the training data perfectly but fail on new claims. They use techniques like cross-validation to test the model’s performance on multiple subsets of the data.
Step three — ethical review. Before deployment, the team runs the model against a set of ‘fairness metrics’. For example:
Demographic parity: Does the approval rate differ significantly across customer groups (age, postcode, ethnicity)?
Equal opportunity: Is the model equally good at spotting legitimate claims (true positives) for all groups?
Explainability: Can the model produce a reason for each decision (e.g., ‘claim amount exceeded historical average by 300%’)?
If the model fails any fairness check, the IT professional must iterate — rebalance the data, try a different algorithm, or add explicit rules to override the model’s output in certain cases.
Step four — deployment with a human-in-the-loop. The IT professional sets up the system so that low-risk, straightforward claims are approved automatically by the AI. Complex or high-amount claims are flagged for a human adjuster to review. The AI also provides a confidence score — if its confidence is below a threshold (say 90%), the claim goes to a person. This prevents the AI from making decisions it is unsure about.
Step five — monitoring and maintenance. After deployment, the IT professional tracks the model’s performance weekly. They look for ‘drift’ — when real-world conditions change (e.g., a new type of storm damage appears) and the model’s accuracy drops. They also monitor for complaints from customers who feel unfairly treated. If patterns of bias emerge, the model is retrained or retired.
In this scenario, the IT professional is not just a coder; they are a bridge between the business goal (speed) and ethical responsibility (fair treatment of customers). They must communicate clearly with non-technical stakeholders about what the AI can and cannot do, and they must document every step for auditors. The AI Associate exam tests whether you understand this full lifecycle, from data gathering to ongoing monitoring, and that ethical considerations must be embedded in every phase — not added as an afterthought.
The Salesforce AI Associate exam tests your understanding of AI fundamentals and ethical principles in a very direct way. The questions are not about deep technical details; they focus on definitions, business implications, and recognising ethical pitfalls. Here is exactly what you need to know.
First, you must be able to define what AI is and what it is not. Exam questions often present a scenario and ask whether the scenario describes AI, machine learning, or traditional programming. The trap is that traditional rule-based systems (like a spreadsheet formula that calculates sales tax) are not AI. The correct answer always involves the system learning from data and improving without being explicitly reprogrammed.
Second, you must distinguish between types of AI. The exam introduces three categories:
Narrow AI (or weak AI): Designed to perform a specific task, like recommending films or translating text. This is the only type of AI that exists today.
General AI (or strong AI): A hypothetical AI that can perform any intellectual task a human can. It does not exist.
Super AI: An even more hypothetical AI that surpasses human intelligence in every field. It is purely theoretical.
A common trap question asks: ‘Which type of AI is used in Salesforce Einstein?’ The correct answer is narrow AI — because Einstein performs specific tasks like predicting lead conversion scores or recommending next best actions, not general reasoning.
Third, ethical principles are a major focus. The exam tests your ability to identify when an AI system is being used unethically. Memorise these five core principles:
Fairness: The AI should not discriminate against individuals or groups. Bias in training data is the most common cause of unfair outcomes.
Accountability: There must be a human responsible for the AI’s decisions. The exam will ask: ‘Who is ultimately responsible for an AI’s harmful action?’ The answer is always the organisation that deployed it, not the AI itself.
Transparency: The AI’s decision-making process should be explainable to users and regulators. ‘Black box’ models that cannot explain decisions are risky.
Privacy: Personal data used to train or run the AI must be collected and stored with proper consent and security measures.
Human oversight: AI should assist humans, not replace them entirely, especially in high-stakes decisions like hiring or loan approvals.
The exam loves to present a scenario where a company builds an AI using only one source of data — for example, training a hiring model only on resumes of current employees (who are mostly from one demographic). The correct answer will identify that the model lacks fairness due to biased training data. A distractor option might say ‘the model lacks speed’ or ‘the model lacks storage’ — both are wrong.
Another frequent question type asks about the ‘AI lifecycle’. You will see a list of steps (data collection, training, deployment, monitoring) and need to identify where ethical review fits. The correct answer is: ethical review should happen at every stage, not just at the end. A trap answer might say ‘ethical review is only needed during deployment’.
Finally, memorise that Salesforce Einstein uses ‘predictive AI’ (a type of narrow AI that makes predictions) and ‘generative AI’ (which creates content like text or images). The exam may ask to identify which is which. Predictive AI answers ‘what will happen?’ (e.g., which customer is likely to buy), while generative AI answers ‘what content can I create?’ (e.g., write an email draft). Both require ethical oversight, but generative AI introduces additional risks like generating false or harmful content.
To summarise the exam focus: know the definition of AI (learning from data), the three types (with narrow AI as the only real one), the five ethical principles, and the necessity of human oversight and bias mitigation throughout the lifecycle. Practise with scenario-based questions where you must spot bias or lack of transparency.
Artificial intelligence is a computer system that learns patterns from data to make predictions or decisions without being explicitly programmed for every scenario.
Machine learning is a subset of AI where the system improves its performance over time by being exposed to more examples.
Today, only narrow AI (designed for specific tasks) exists — general AI that can think like a human does not yet exist.
Bias in AI comes from biased training data; the AI will reproduce and amplify the prejudices present in the historical examples it learns from.
Ethical principles for AI — fairness, accountability, transparency, privacy, and human oversight — must be applied at every stage of the AI lifecycle, not just at the end.
Human oversight is essential: AI should assist human decision-making, not replace it entirely, especially in high-stakes areas like hiring, lending, and healthcare.
Salesforce Einstein uses narrow AI for predictive tasks (e.g., scoring leads) and generative tasks (e.g., writing email drafts), and both require ethical guardrails.
Explainability matters: if an AI cannot justify its decision, it should not be used in regulated or high-impact settings.
The organisation that deploys an AI is accountable for its actions — the AI itself cannot be held responsible.
Data quality and diversity are more important than data quantity for building a fair and accurate AI model.
These come up on the exam all the time. Here's how to tell them apart.
Artificial Intelligence (AI)
Learns patterns from data without explicit rules for each situation.
Can handle complex, unstructured inputs like images or natural language.
Requires large amounts of historical data to train effectively.
Traditional Programming
Follows fixed, step-by-step instructions written by a programmer.
Works best with structured, predictable inputs and clear logic.
Needs no data to function; the programmer encodes all decision logic upfront.
Narrow AI
Designed for one specific task (e.g., translation, recommendation).
Exists today and is used in products like Salesforce Einstein.
Cannot transfer knowledge to tasks outside its training domain.
General AI
Can perform any intellectual task a human can, across domains.
Does not exist yet — it is a theoretical goal of AI research.
Would require general reasoning, common sense, and adaptability far beyond current systems.
Algorithmic Bias
Causes the AI to produce systematically unfair outcomes for certain groups.
Is often invisible until the model is used on real, diverse data.
Requires fairness metrics and ethical review to detect and mitigate.
Data Quality Issue
Refers to incomplete, duplicate, or erroneous data that harms overall model accuracy.
Is usually detected during data cleaning and preparation.
Can often be fixed by improving data collection methods or removing bad records.
Mistake
AI and machine learning are the same thing — you can use the terms interchangeably.
Correct
Machine learning is a subset of AI. Not all AI uses machine learning. For example, early AI systems used rule-based logic (if-then statements) without learning from data.
Marketers and news articles often use 'AI' as a buzzword for anything smart, including plain automation. Beginners hear 'AI' everywhere and assume it is all the same.
Mistake
If an AI is trained on good data, it will automatically be fair and unbiased.
Correct
Even high-quality data can contain historical biases that become baked into the AI's decisions. Fairness must be actively checked and corrected — it is not automatic.
People assume that computers are objective. In reality, an AI learns the patterns in its training data, including any human biases present. The phrase 'garbage in, garbage out' applies to bias as well as accuracy.
Mistake
Ethical principles for AI are optional guidelines that companies follow if they want to look good.
Correct
Ethical principles are increasingly required by law (like the EU AI Act) and by platform rules (like Salesforce's own AI Ethics guidelines). Ignoring them can lead to fines, lawsuits, and reputational damage.
Many beginners see ethics as a 'soft' topic compared to coding. They do not realise that regulatory bodies now enforce ethical AI practices with real penalties.
Mistake
AI can think and reason like a human, so it understands the context of its decisions.
Correct
Today's AI (narrow AI) does not understand anything. It finds statistical correlations in data. It cannot reason, feel empathy, or grasp the real-world meaning of its outputs.
Movies and sci-fi books portray AI as conscious beings. Beginners transfer that expectation to real systems, leading to unrealistic fears or trust.
Mistake
The more data you feed an AI, the better it will perform, regardless of the data's quality.
Correct
More data helps only if it is relevant, diverse, and clean. Adding irrelevant or duplicate data can confuse the model and reduce performance. Data quality matters more than quantity.
The phrase 'big data' makes people think volume alone solves problems. Beginners often overlook that noisy or biased data leads to poor AI.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
No. The AI Associate exam is conceptual. You need to understand definitions, ethical principles, and real-world scenarios — not write code. Focus on memorising terms like narrow AI, bias, and transparency.
Predictive AI forecasts what will happen next (e.g., which customer is likely to churn). Generative AI creates new content — text, images, code — that didn't exist before (e.g., writing a sales email). Both are types of narrow AI.
No, because fairness is a human value that depends on context. Different groups may define fairness differently. AI can be made less biased, but perfect fairness is an ideal to strive for, not a guarantee.
The organisation that built and deployed the AI is legally and ethically responsible. The AI itself is just a tool. This is why human oversight and documentation are critical.
A black box model is an AI whose internal decision-making process is so complex that even its creators cannot explain why it reached a particular output. This violates the transparency principle and is risky in regulated industries like healthcare or finance.
Yes. Salesforce publishes AI Ethics principles that align with industry standards: fairness, accountability, transparency, privacy, and human oversight. The exam expects you to be familiar with these values.
You've finished Introduction to Artificial Intelligence and Ethical Principles. Continue through the AI Associate study guide to build a complete picture of the exam.
Done with this chapter?