CompTIA Data+ DA0-002 · Free study guide

Objective 1.5 — Identify artificial intelligence concepts

Artificial intelligence appears in analytics as a family of capabilities, not one magic product. A system that generates a narrative, a model that recognizes patterns in images, and a bot that moves a report into a shared folder may all be described as automation, but they work in different ways and require different controls.

For the exam, identify what the system is doing before judging whether the term fits. Then remember that fluent output, high model accuracy, and reliable business meaning are separate claims.

Generative AI and large language models

Generative AI creates new content based on patterns learned from training data. It can produce text, code, images, audio, or structured output. In an analyst’s workflow, common uses include drafting SQL, explaining code, summarizing a report, proposing chart descriptions, and converting a plain-language request into a query.

A large language model, or LLM, is a model trained on large amounts of language data to predict and generate sequences of tokens. It can respond fluently because it has learned statistical patterns in language. Fluency does not prove that a statement is factual, that a query matches the organization’s metric definition, or that a cited source exists.

An LLM-generated query can execute successfully and still join at the wrong grain, omit null rows, or filter the wrong population. Verification must compare the query and result with the schema, business definition, and trusted reference data. Asking the same model whether its answer is correct is not independent validation.

Foundation models and deep learning

A foundation model is a broadly trained model that can support many downstream tasks. It is usually adapted through prompting, retrieval, fine-tuning, or another task-specific layer rather than trained from nothing for every use. Large language models are one kind of foundation model, but foundation models can also work with images, audio, or multiple data types.

Deep learning uses neural networks with many layers to learn representations from data. It supports capabilities such as speech recognition, image classification, language processing, and complex pattern detection. Deep learning is a technical approach; it is not synonymous with every AI system. A simple rule-based automation or linear model does not become deep learning because it runs automatically.

The broad training of a foundation model provides flexibility, while adaptation supplies task context. An organization that connects a model to approved documentation is not necessarily retraining the model. It may be retrieving relevant material and supplying that material with the request.

Natural language processing

Natural language processing, or NLP, is the field concerned with computers processing human language. NLP tasks include classification, translation, entity extraction, sentiment analysis, summarization, question answering, and turning natural-language requests into structured operations.

A BI feature that accepts “show quarterly revenue by region” and produces a chart is a natural-language interface. It may use an LLM, another NLP model, a rules engine, or a combination. The user experience alone does not reveal the exact implementation.

Language is ambiguous. “Active customer,” “last quarter,” and “revenue” can each have organization-specific meanings. A responsible NLP system connects those phrases to governed definitions and exposes how the request was interpreted.

Robotic process automation and automated reporting

Robotic process automation, or RPA, automates repetitive interactions with applications by following defined steps. An RPA bot might log in to a legacy system, download an export, rename it, place it in an approved location, and trigger a downstream process. RPA is useful when a stable interface exists but no practical API is available.

RPA is usually deterministic: given the same state and rules, it performs the same steps. It differs from generative AI, which produces probabilistic output. An RPA bot can still fail when a button moves, a page loads slowly, or an unexpected dialog appears. Logging, exception handling, and human escalation remain necessary.

Automated reporting schedules the collection, calculation, rendering, and delivery of reports. It may combine queries, business rules, BI software, RPA, and AI-generated narrative. Automation improves consistency and timeliness, but it can also distribute an error faster. Every automated report needs an accountable owner, source checks, calculation tests, access controls, and monitoring for stale or failed refreshes.

Responsible use in analytics

AI controls should match the harm a failure could cause.

• Verification compares generated claims with source data, known totals, tested logic, and independent review.
• Privacy keeps personal, confidential, or regulated data out of unapproved external services. Use an approved environment, redacted schema, or synthetic sample when possible.
• Security applies least privilege. A query assistant should use read-only access to approved views rather than an unrestricted administrative account.
• Bias review checks whether training data, labels, or outputs disadvantage a group or reproduce historical decisions without justification.
• Transparency records the model, prompt or configuration, source context, output, reviewer, and known limitations when those details matter to the decision.
• Human oversight assigns a person authority to approve, reject, correct, or stop the system. A nominal reviewer who cannot inspect the evidence is not meaningful oversight.

AI should not receive more authority merely because it is convenient. The analyst and organization remain responsible for what is published or acted upon.

Scenario: automating a weekly performance report

A team currently downloads a legacy export, calculates KPIs, writes a summary, and emails a report every Monday. It proposes an RPA bot for the download, a scheduled data pipeline for calculations, an LLM for the narrative, and a BI portal for delivery.

The RPA bot fits the stable, repetitive interface step, especially if no API exists. Its run should log each action and stop on an unexpected screen. KPI calculations should remain deterministic, tested code tied to certified definitions. The LLM may draft a narrative, but specific numbers and causal claims must be validated against the calculated output. The model should receive only the minimum approved data. A human owner reviews the narrative, while portal permissions ensure recipients see only authorized information.

If the model invents a reason for a revenue change, a polished paragraph does not make that claim acceptable. The report should state observed evidence and distinguish a hypothesis from a demonstrated cause.

Exam traps

• Assuming all AI is generative AI or all models are LLMs.
• Calling a deterministic screen-clicking bot deep learning.
• Treating a foundation model as a model trained for only one narrow task.
• Assuming natural-language querying understands an organization’s definitions automatically.
• Publishing an LLM-generated number because the output sounds confident.
• Asking a model to verify itself instead of checking source data.
• Giving an AI assistant write access when read-only approved views meet the task.
• Believing automation removes the need for ownership, monitoring, or access control.

Readiness checklist

• I can distinguish generative AI, LLMs, foundation models, and deep learning.
• I can recognize common NLP tasks and natural-language analytical interfaces.
• I can explain how RPA differs from probabilistic AI.
• I can describe the parts and controls of automated reporting.
• I can identify hallucination, privacy, security, and bias risks in a scenario.
• I can choose independent verification rather than model self-confirmation.
• I can apply least privilege and data minimization to an AI workflow.
• I can state where meaningful human oversight belongs before an output affects a decision.