Generative AI Capabilities and Limitations for Business (AIF-C01)

Generative AI has four headline strengths that the exam groups under advantages. Adaptability means a single foundation model can handle many different tasks, such as summarizing, drafting, translating, and answering, without building a separate system for each. Because the model learned broad patterns during training, you can point it at a new job simply by changing the prompt.

Responsiveness means it produces useful output quickly, often in seconds, which makes it practical for interactive uses like chat and live assistance. Conversational capability is its ability to understand and reply in natural language, following the back-and-forth of a dialogue and keeping track of what was said. This is what lets non-technical users work with it just by typing plain questions. The ability to generate content is the defining advantage: it creates original text, images, code, and more on demand, rather than only classifying or scoring existing data.

Together these strengths explain why GenAI spreads so fast in business. One adaptable, conversational model can draft a report for marketing, answer a customer in support, and summarize a contract for legal, reducing the cost and time of work that used to need specialized tools or manual effort. For the exam, tie each advantage to a plain description: many tasks from one model is adaptability, fast answers is responsiveness, natural dialogue is conversational capability, and creating new material is content generation.

The same qualities that make generative AI powerful also create risks, and Task 2.2 weights these heavily. A hallucination is when a model states false or made-up information as if it were true and does so confidently; it is a frequently tested limitation and gets its own section next. Interpretability, also called explainability, is limited: these models are so complex that it is hard to explain why they produced a particular output, which is a problem in regulated settings where decisions must be justified.

Inaccuracy means the output can simply be wrong or low-quality, because the model predicts likely text rather than verifying facts. Two kinds of task are especially exposed. First, a model with no connection to live data or tools cannot know current figures such as today's exchange rate, a share price or this morning's stock level; it can only produce a plausible-looking number from its training data. Second, a language model generates digits as text rather than calculating them, so exact arithmetic, totals and percentages can come out wrong. Outputs like these must be checked against an authoritative source, or the model must be given that source or a tool, before anyone relies on them.

Nondeterminism means the same prompt can produce different answers on different runs, which makes behavior harder to test and guarantee. Beyond these four, two practical concerns appear at concept level: data privacy, since sensitive information sent in prompts or used for training must be protected, and cost, since running large models at scale adds up. The table contrasts the strengths and the weaknesses.

Telling look-alike problems apart

Several problems sound similar but have different causes and different fixes, so name each one precisely.

A hallucination is an output that is factually wrong or entirely fabricated, presented by the model with the same confident tone it uses for correct answers. The model might invent a statistic, cite a source that does not exist, or state a false fact, and because the wording sounds authoritative, a user can easily be misled. This happens because a generative model is built to produce plausible-sounding text based on patterns, not to check whether that text is true.

Hallucination is a frequently tested limitation because it carries real business risk. Imagine a bank deploys a GenAI assistant that confidently tells a customer about a fee policy that does not exist, or a healthcare chatbot that states an incorrect dosage. The answer looks trustworthy but is false, which can cause financial loss, legal exposure, or harm, and it erodes customer trust once the error is discovered.

What reduces hallucinations, and what does not

You cannot remove the possibility of hallucination entirely, but you can reduce it. The most direct mitigation is grounding: giving the model verified source documents at request time and instructing it to answer from them, usually through Retrieval Augmented Generation (RAG), which retrieves the relevant passages from a trusted knowledge base and adds them to the prompt. RAG is explained in depth in the Task 3.1 lesson on foundation-model design. Automated checks add a second layer: Amazon Bedrock Guardrails offers contextual grounding checks that score whether a response is supported by the supplied source and relevant to the question, and can block responses that fall below a threshold. These checks reduce ungrounded answers; they do not guarantee every answer is correct. Where a wrong answer could cause harm, such as medical, legal or financial advice, keep a human in the loop to review output before it reaches the user.

Three popular fixes do not solve the problem on their own:

• Lowering temperature makes the model choose its most likely words, so answers vary less from run to run. It reduces variation, not fabrication: if the most likely continuation is false, a low-temperature model states it consistently. (Temperature and other inference parameters are covered in Task 3.1.)
• Using a larger model can improve quality, but large models hallucinate too; size is not a substitute for grounding in your own facts.
• Fine-tuning on domain data improves the model's domain knowledge and style, but the model still generates rather than looks things up, so it can still invent details. It also bakes in a snapshot of facts that goes stale when policies change.

For Task 2.2, the essential recognition is the definition itself: when a scenario describes a model that "states false facts confidently" or "makes up information that sounds correct," the concept being tested is hallucination, and the mitigation that matches it is grounding plus verification, not a parameter change.

Three more limitations deserve a closer look because their trigger phrases show up in questions. Interpretability, also called explainability, is the degree to which a human can understand why a model produced a specific output. Large generative models are effectively black boxes: they have billions of internal parameters, so even their builders often cannot say exactly why a given answer appeared. When a question stresses that a model "cannot explain why it produced an output" or that a decision cannot be justified to a regulator, interpretability is the concept in play.

Asking the model to explain its reasoning does not fix this. The explanation is itself generated text: it can sound convincing while not reflecting how the answer was actually produced, and it can be wrong in the same ways any output can. Likewise, fine-tuning a model or lowering its temperature changes what it says, not how transparent it is. When a decision must be justified case by case, the reliable answer is to make that decision with a system whose logic can be shown, such as documented rules or an interpretable model, and use generative AI around it.

Nondeterminism means the model can give different answers to the identical prompt on separate runs. Traditional software is deterministic, so the same input yields the same output, but a generative model samples from probabilities, so its replies vary. This makes testing, quality assurance, and reproducibility harder, since you cannot always predict the exact response. The trigger phrase is "same prompt, different answers."

Two further concerns are practical rather than technical. Data privacy is a risk because prompts may contain sensitive or personal information, and that data must be handled, stored, and protected carefully, especially in regulated industries. Cost is a limitation because large models consume significant compute, and token-based pricing means heavy use can become expensive at scale. None of these rule out generative AI; they are trade-offs to manage, and recognizing each one by its description is what the exam asks of you.

Generative AI fits a task when the job is to create or transform language and content and a person or a check can catch the occasional mistake: drafting emails and letters, summarizing long documents, personalizing messages, answering questions from supplied documents, and brainstorming. It fits poorly when the task demands a guaranteed exact answer (calculating pay or tax, applying a fixed eligibility rule), a decision that must be justified case by case to a regulator or customer, or live business data the model cannot see. Ask a model with no connection to the inventory system how many units are in stock and it will produce a plausible guess, not the number; the task needs access to that data source, through retrieval or a tool, not generation alone.

The usual answer is not "GenAI or not" but a split workflow. Keep exact, rule-based or must-be-explained decisions in deterministic or interpretable systems, and use generative AI for the language around them. A payroll team, for example, calculates pay, tax and deductions with its exact payroll rules, then uses a generative model to draft a plain-language summary of each payslip for employees, with staff spot-checking the summaries before they go out. Each part uses the tool whose strengths match it.

Choosing a generative AI model is a balancing act: no single model wins on every dimension, so you match the model to the requirements of the use case. Task 2.2 lists the factors to weigh:

• Model type and capabilities — which inputs and outputs it supports and which tasks it handles well (see below).
• Performance requirements — the quality and accuracy the task demands; a legal summary needs higher accuracy than a casual brainstorming helper.
• Cost — the price per token and the total spend at your expected volume.
• Latency — how fast a response must come back; a live chat or voice assistant needs low latency, while a nightly batch job does not.
• Model complexity and size — larger models are often more capable but slower and more expensive, so a smaller model is often the better fit when it performs nearly as well on a simple, high-volume task.
• Constraints and compliance — regulatory, security, contractual or data-residency rules that the chosen model and its deployment must satisfy.

Capabilities: match the modality and the input size

"Multi-modal" covers different capabilities, so read which direction the scenario needs.

Compliance filters first, then you optimize

Compliance and hard constraints are not traded off against speed or price; they decide which models are on the shortlist at all. Apply them first, then compare the remaining candidates on quality, latency and cost. A model that scores best in testing but is only offered in a Region outside the required geography is out, however good it is. Data residency is about where inference physically runs, so only a different deployment (another model or Region, or an in-geography inference path) fixes a residency gap; changes that leave the processing location the same do not. Likewise, a healthcare organization in the United States subject to HIPAA needs a provider that will sign a business associate agreement (BAA) before it processes patient data; without that agreement in place, speed, quality or price cannot put a model back on the shortlist.

How you call a model matters too. With Amazon Bedrock, using a model in your own Region keeps inference there. Cross-Region inference can raise throughput by routing requests to other Regions: a geographic inference profile (for example an EU profile) keeps processing within that geography, while a global inference profile can route requests to supported Regions anywhere in the world, which can break a data-residency rule.

The other factors do trade off against one another. A bigger, more capable model may improve quality but hurt latency and cost; a cheaper, smaller model may be fast and affordable but less accurate. Scenarios usually state a priority, such as "lowest possible latency," "strict compliance," or "tight budget"; match that stated need to the selection factor it names.

A business measures the value of a generative AI project in business terms, such as revenue, customers and time saved, not just in technical ones. Task 2.2 lists several metrics you should recognize, and the table pairs each with what it tells you.

ARPU versus CLV is the near-twin to watch. ARPU is a per-period figure, so it moves when price or spend per period changes. CLV covers the whole relationship, so it also grows when customers simply stay longer. For example, a successful upsell that raises what each customer spends per month lifts ARPU (and CLV with it), whereas better retention can lift CLV even when monthly spend per user does not change.

Business metrics versus technical measurements. Throughput (requests or tokens per second), latency, perplexity and parameter count describe how a model runs or how well it predicts text; they are useful to engineers but do not show whether the project pays off. Leadership asking whether a GenAI feature is worth it wants business metrics such as ROI, conversion rate, ARPU, CLV and efficiency. Accuracy sits between the two: it is a quality measure, but it carries business meaning because wrong answers erode trust.

Match the metric to the stated goal. "Is the project worth the cost" points to ROI; "time saved per agent" points to efficiency; "share of visitors who buy" points to conversion rate; "revenue per user each month" points to ARPU; "how much a customer is worth over time" points to CLV; and "one model that holds up across several business areas" points to cross-domain performance.

Bring it together with a realistic decision. A mid-size online retailer is evaluating a generative AI customer-service assistant to answer shopper questions around the clock. The benefits are clear: it is adaptable enough to cover many question types, responsive enough to reply instantly, and conversational enough for customers to use without training, promising better efficiency and, if service improves, higher customer lifetime value.

The team also weighs the risks. The biggest is hallucination: if the assistant confidently states a wrong return policy or an incorrect price, the company could face refunds, complaints, and lost trust. Nondeterminism means two customers asking the same question might get different answers, which complicates quality control. Interpretability is limited, so if a customer disputes an answer, staff cannot easily explain why the model said it. And there are data-privacy obligations around the personal order details customers share.

The sensible conclusion is not "use it" or "avoid it" but "use it with guardrails." The retailer decides to ground the assistant in its verified policy documents through retrieval (RAG), connect it to the order system for live order status rather than letting it guess, keep a human in the loop for refunds and disputes, and measure success with ROI and conversion rate before expanding. This is exactly the balanced judgment Task 2.2 tests: recognizing the advantages, naming the specific limitations by their definitions, and weighing benefit against risk for the business problem at hand.