Building Simple Ai Workflows

• A. A single prompt sent to an AI model once
• B. A sequence of connected steps in which AI and other tools process inputs to produce a desired output ✓
• C. A type of neural network architecture
• D. A programming language used only for machine learning

Why: An AI workflow chains together multiple steps, such as receiving input, processing it with an AI model, and passing results to other tools or actions, to accomplish a task automatically.

• A. Typing a single question into a chatbot and reading the answer
• B. Manually writing an email without any tools
• C. Automatically reading incoming customer emails, classifying them by topic with an AI model, and routing each one to the right team ✓
• D. Storing files in a folder on your computer

Why: This example includes a trigger (incoming email), an AI processing step (classification), and an action (routing), which together form a workflow rather than a one-off interaction.

• A. To permanently delete the workflow's data
• B. To start the workflow when a specific event occurs, such as receiving a form submission ✓
• C. To train the AI model on new data
• D. To format the final output as a PDF

Why: A trigger is the event that initiates a workflow, such as a new email arriving, a file being uploaded, or a scheduled time being reached.

• A. It makes the workflow run faster on any hardware
• B. It ensures each step knows what data to expect and what format to produce, so the steps connect reliably ✓
• C. It is required by law for all AI systems
• D. It removes the need for testing

Why: Clear inputs and outputs act as contracts between steps. When each step knows the expected data format, the workflow connects smoothly and errors are easier to find.

• A. Start with the most complex multi-agent system you can imagine
• B. Choose a single, repetitive task and map out its steps before adding AI ✓
• C. Purchase the most expensive AI model available
• D. Write as much code as possible before defining the goal

Why: Beginners should start small with a clear, repetitive task. Mapping the steps first helps identify where AI adds value and keeps the project manageable.

• A. Model training
• B. Data ingestion ✓
• C. Model deployment
• D. Performance monitoring

Why: Data ingestion is the stage where raw data is gathered from sources such as databases, APIs, or files and brought into the pipeline for further processing.

• A. To deploy the model to end users
• B. To clean and transform raw data into a format suitable for model training ✓
• C. To measure how accurate the final predictions are
• D. To store the trained model in a registry

Why: Preprocessing handles tasks like removing duplicates, handling missing values, normalizing values, and encoding categories so that the model can learn effectively from the data.

• A. The final output the model produces
• B. An individual measurable input variable used by the model to make predictions ✓
• C. The software library used to write the code
• D. The server where the model runs

Why: Features are the input variables, such as age, income, or word counts, that the model uses to learn patterns and generate predictions.

• A. It makes training faster by reducing data size
• B. It provides an unbiased estimate of how the model performs on data it has not seen ✓
• C. It replaces the need for a training dataset
• D. It is required to deploy the model to production

Why: Evaluating on data the model was not trained on reveals whether it generalizes well, whereas testing on training data can give misleadingly optimistic results.

• A. Data collection
• B. Feature engineering
• C. Model deployment ✓
• D. Hyperparameter tuning

Why: Deployment makes the trained model available to users or other systems, often through an API, so it can generate predictions on new inputs in real time or in batches.

• A. It makes the files smaller so they load faster on any device
• B. Messy or inconsistent data can lead the AI to produce inaccurate or unreliable outputs ✓
• C. AI models can only read data stored in spreadsheets
• D. Cleaning data removes the need for any testing of the workflow

Why: AI systems learn patterns from and act on the data they receive. Errors, duplicates, missing values, or inconsistent formats can propagate through the workflow and degrade the quality of results, so cleaning data improves reliability.

• A. Delete the entire dataset because it is incomplete
• B. Ignore the missing values and hope the AI handles them automatically
• C. Fill the gaps with a reasonable default such as the median, or flag them for review, depending on the context ✓
• D. Replace every missing value with the number zero regardless of the column's meaning

Why: Missing values should be handled thoughtfully. Imputing with a statistic like the median or marking them explicitly preserves useful rows, while blindly using zero can distort meaning, for example treating an unknown age as zero years old.

• A. Translating all text into a single language
• B. Converting text to a consistent format, such as lowercase and standardized spacing and punctuation ✓
• C. Removing all words that appear more than once
• D. Encrypting text so only the AI can read it

Why: Normalization standardizes text so that variations like 'New York', 'new york', and 'New York' are treated as the same value. This reduces noise and helps the workflow match and group records correctly.

• A. Duplicates always contain viruses that corrupt the AI model
• B. Duplicates can overweight certain examples, skew results, and inflate counts or statistics ✓
• C. AI platforms charge extra for each duplicate row, so they must be deleted
• D. Duplicates are required for the AI to verify that the data is authentic

Why: When the same record appears multiple times, it has disproportionate influence on summaries, analyses, and any learned patterns, which can bias outputs. Identifying and resolving duplicates keeps the data representative.

• A. Sending the full raw file so the AI can decide what is important
• B. Removing or masking personally identifiable information and checking the format is consistent ✓
• C. Converting the file to an image so the AI cannot read personal details
• D. Adding extra random rows to hide the real customer records

Why: Removing or masking personal identifiers such as names, emails, and phone numbers protects privacy, and verifying consistent formatting ensures the AI receives clean, usable input. Both are standard preparation steps before processing.

• A. To make the model train itself on new data automatically
• B. To pass inputs to the model and use its outputs to complete a specific job ✓
• C. To permanently store every response the model has ever generated
• D. To replace the need for any human review of results

Why: A workflow connects a model to a defined task by sending it an input, receiving its output, and using that output in the next step. Training and storage are separate concerns, and human review is often still needed.

• A. The hardware the model runs on
• B. The billing limit set for API usage
• C. The instructions and input text given to the model ✓
• D. The final result returned after processing

Why: A prompt is the text sent to the model, usually containing instructions and the input data it should work with. The model's reply is the output, not the prompt.

• A. It makes the model respond faster in every case
• B. It lets later steps parse the result reliably without manual cleanup ✓
• C. It removes the need to write any prompt at all
• D. It guarantees the model's answer is always factually correct

Why: A predictable structure lets downstream code read specific fields directly. Structured formats improve reliability of parsing but do not guarantee factual accuracy or eliminate the need for good prompts.

• A. Image generation
• B. Text classification ✓
• C. Database backup
• D. Audio transcription

Why: Assigning a review to one of a fixed set of labels is classification. The input is text and the output is a category, so it is not image, storage, or audio work.

• A. Remove all instructions from the prompt so the model is free to decide
• B. Validate the output and retry or flag it when it does not match the expected format ✓
• C. Run the workflow only once and never check the results
• D. Increase the model's temperature to maximum

Why: Checking outputs against the expected format and retrying or flagging failures catches errors before they propagate. Removing instructions or raising randomness tends to make outputs less predictable, not more.

• A. To make each step run completely independently with no shared data
• B. To pass the output of one step as input to the next step, producing a final result ✓
• C. To replace the need for any input data from the user
• D. To run every step at the same time regardless of dependencies

Why: Chaining connects steps so that each step builds on the previous one. The output of one step becomes the input of the next, which allows a complex task to be broken into manageable parts.

• A. Translate the article first, then summarize the translation
• B. Summarize the article first, then translate the summary ✓
• C. Translate and summarize at the same time with no connection between them
• D. Skip the summary and translate only the title

Why: The translation step needs the summary as its input, so the summary must be produced first. Reversing the order would mean translating the full article and then summarizing it, which is a different workflow.

• A. All later steps still run using empty or incorrect data, which may produce misleading results ✓
• B. The error is automatically fixed and the workflow continues without any issue
• C. The workflow always skips to the final step and returns the original input
• D. The earlier steps are rerun automatically from the beginning

Why: Because each step depends on the previous step's output, an error can propagate forward. Good workflows check outputs and handle failures so that bad data does not silently flow into later steps.

• A. It allows you to inspect, debug, and reuse intermediate results without rerunning the entire workflow ✓
• B. It prevents the AI model from ever generating incorrect text
• C. It reduces the number of words the user is allowed to type
• D. It removes the need to define the order of the steps

Why: Saving intermediate outputs makes it easier to see where a problem occurred and to reuse results. It helps with debugging and can save time and cost by avoiding repeated work.

• A. Draft the reply, because replies are always the first thing written
• B. Extract the main complaint, because the reply depends on knowing what the complaint is ✓
• C. Both steps should run at the same time without using each other's output
• D. Neither step is needed because the email already contains the reply

Why: The drafting step relies on the extracted complaint to write a relevant response. Placing extraction first ensures the second step has the information it needs.

• A. To make the workflow run faster on the server
• B. To check that outputs are accurate, consistent, and useful for the intended task ✓
• C. To reduce the number of words the model generates
• D. To remove the need for any human review in the future

Why: Testing verifies that outputs meet quality expectations for the real task, catching errors, inconsistencies, and gaps before users depend on the workflow.

• A. Reading the output once and assuming it is correct if it looks fluent
• B. Checking whether the output matches the required structure, such as valid fields or JSON syntax ✓
• C. Counting how many times the word 'the' appears
• D. Comparing the output only to the length of the input

Why: Format-based checks are objective and easy to automate, so they reliably catch structural failures that fluent-looking text can hide.

• A. A single example always produces the same output, so more tests are unnecessary
• B. Edge cases, unusual inputs, and varied phrasing can reveal failures that a single example would miss ✓
• C. Testing with many inputs is required by law for all AI systems
• D. Multiple inputs make the model learn from your data permanently

Why: Real users send varied and unexpected inputs, so testing across a range of cases, including edge cases, reveals weaknesses that a single happy-path example hides.

• A. A set of expensive premium model outputs used for marketing
• B. A collection of inputs paired with known-good expected outputs used to measure performance over time ✓
• C. A backup copy of the model's training weights
• D. A list of the most popular user questions with no answers attached

Why: A golden dataset provides a fixed benchmark of inputs with verified expected results, so you can rerun it after changes and detect regressions.

• A. Deploy the change immediately since the new prompt is clearly better
• B. Rerun your test cases and compare the results against the previous version ✓
• C. Delete the old test cases because they no longer apply
• D. Ask the model whether its new output is correct and accept its answer

Why: Rerunning the same tests and comparing results to the previous version shows whether the change improved outputs or introduced new failures.

• A. Deleting all test data to free up storage
• B. Testing the workflow with representative inputs and confirming the outputs meet expectations
• C. Publicly announcing the launch before testing begins
• D. Removing all logging so the system runs faster

• A. Logs replace the need for any monitoring or alerts
• B. Logs allow you to trace errors, track performance, and diagnose problems when the workflow behaves unexpectedly
• C. Logs automatically improve the model's accuracy over time
• D. Logs are only useful during the development phase and should be turned off after deployment

• A. The software becomes more efficient with age
• B. The input data changes in ways that differ from the data used during development, a problem often called data drift
• C. The workflow automatically stops learning after deployment
• D. Users always submit identical inputs, causing the system to slow down

• A. Setting it up once and never checking it again
• B. Monitoring outputs, reviewing errors, updating prompts or logic, and retesting after changes
• C. Replacing the entire workflow every week
• D. Only fixing problems after users complain publicly

• A. Versioning makes the workflow run faster on all hardware
• B. It lets you compare results, and roll back to a previously working version if a change causes problems
• C. Versioning is required by law for every AI system
• D. It prevents the workflow from ever producing errors