[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"cheat-sheet---en":3,"domain-info---en":3,"topic-info----en":3,"prev-aws-certified-ai-practitioner-ai-ml-fundamentals-practical-ai-use-cases-when-to-use-ai-and-when-not-en":4,"lesson-aws-certified-ai-practitioner-ai-ml-fundamentals-practical-ai-use-cases-when-to-use-ai-and-when-not-en":18,"next-aws-certified-ai-practitioner-ai-ml-fundamentals-practical-ai-use-cases-when-to-use-ai-and-when-not-en":324},null,{"locked":5,"reason":6,"meta":7,"item":3},true,"paywall",{"title":8,"description":9,"isFree":10,"estimatedMinutes":11,"difficulty":12,"learningObjectives":13},"Types of Inferencing","Once a model is trained, inference is how it serves predictions. Compare real-time, batch, asynchronous, and serverless inference and match each to latency, payload, and traffic needs.",false,15,"beginner",[14,15,16,17],"Define inference and distinguish it from training","Compare real-time, batch, asynchronous, and serverless inference","Match an inference type to latency, payload, and traffic requirements","Map the inference types to Amazon SageMaker options",{"locked":10,"reason":3,"meta":19,"item":27},{"title":20,"description":21,"isFree":5,"estimatedMinutes":11,"difficulty":12,"learningObjectives":22},"When to Use AI and When Not To","How to tell when machine learning genuinely adds value and when a plain rule, a lookup, or a human is the better answer, using cost-benefit and the prediction-versus-exact-outcome test.",[23,24,25,26],"Recognize the three patterns where AI and ML add real value: assisting decisions, scaling a solution, and automating repetitive work","Apply a cost-benefit view that counts data, training, and ongoing maintenance, not just the model","Use the prediction-versus-exact-outcome test to reject ML when a guaranteed result is required","Identify the common signals that a problem does not need machine learning at all",{"id":28,"title":20,"body":29,"description":21,"difficulty":12,"estimatedMinutes":11,"extension":272,"infographics":273,"isFree":5,"learningObjectives":274,"meta":275,"navigation":5,"path":276,"quiz":277,"seo":321,"stem":322,"__hash__":323},"courses/courses/aws-certified-ai-practitioner/en/domains/01-ai-ml-fundamentals/02-practical-ai-use-cases/01-when-to-use-ai-and-when-not.md",{"type":30,"value":31,"toc":261},"minimark",[32,36,39,44,47,55,62,69,72,76,79,82,85,89,100,103,109,113,116,149,153,234,237,241,244,258],[33,34,35],"p",{},"A product manager says, \"Let's add AI to it.\" The team nods, but the useful question comes next: add AI to what, and why? Some problems are transformed by machine learning. Others get slower, more expensive, and less reliable when you bolt a model onto them. The AI Practitioner exam tests whether you can tell the two apart, and so does every real budget meeting. This lesson gives you the judgment to say yes for the right reasons and no without hesitation.",[33,37,38],{},"The previous topic taught you what AI and ML are and how models learn. This one starts the practical half: given a business problem, is machine learning even the right tool? Get this decision wrong and no amount of good modeling saves the project.",[40,41,43],"h2",{"id":42},"where-ai-and-ml-earn-their-keep","Where AI and ML earn their keep",[33,45,46],{},"Machine learning adds value in three recurring patterns. When a problem fits one of them, ML is worth considering. When it fits none, be suspicious.",[33,48,49,50,54],{},"The first pattern is ",[51,52,53],"strong",{},"assisting human decisions",". The model does not decide; it surfaces a prediction, and a person acts on it. A radiologist still signs off on the scan, but a model flags the three regions most likely to need a closer look. A loan officer still approves the loan, but a model scores the default risk first. The value is a faster, better-informed human, not a replaced one.",[33,56,57,58,61],{},"The second pattern is ",[51,59,60],{},"scaling a solution beyond human capacity",". Some tasks are simple for one item and impossible for millions. Reading one product review is easy; reading 4 million reviews a day to track sentiment is not. A model that classifies each review in milliseconds turns an impossible task into a routine one. The signal here is volume that no team could staff.",[33,63,64,65,68],{},"The third pattern is ",[51,66,67],{},"automating repetitive judgment",". Not repetitive mechanics, which a plain script already handles, but repetitive decisions that used to need a person: routing a support ticket to the right queue, tagging photos by content, transcribing calls. When the same fuzzy decision happens over and over and a small error rate is tolerable, automation with ML frees people for work that actually needs them.",[33,70,71],{},"Notice what unites the three: the pattern is complex or high-volume, and a good-but-imperfect answer is genuinely useful. Hold that idea, because the cases where ML fails are exactly where it breaks.",[40,73,75],{"id":74},"the-cost-side-of-the-ledger","The cost side of the ledger",[33,77,78],{},"The exam explicitly lists cost-benefit analysis as a reason to say no, and the trap is counting only half the cost. Teams price the first training run and stop there. A model in production is a living system with recurring costs that often dwarf the build.",[33,80,81],{},"You pay to collect and label data before you can train anything, and labeling is frequently the biggest line item. You pay to train, and for large models that bill is real. Then, after launch, the ongoing costs begin: monitoring the model so you notice when its accuracy slips, retraining it on fresh data as the world changes (a fraud model trained on last year's fraud goes stale), and the engineering time to keep the whole pipeline running. A model that took two weeks to build can take years to maintain.",[33,83,84],{},"So the real question is never \"can ML do this?\" It is \"does the value beat the full lifetime cost?\" A model that boosts conversion by 0.1% may be a brilliant technical result and a terrible investment if it costs three engineers to keep alive. Weigh the benefit against data, training, and maintenance together, not against the training run alone.",[40,86,88],{"id":87},"when-a-prediction-is-the-wrong-shape-of-answer","When a prediction is the wrong shape of answer",[33,90,91,92,95,96,99],{},"This is the highest-value idea in the lesson and the one the exam probes most. Machine learning produces ",[51,93,94],{},"predictions",", and predictions are probabilistic: usually right, occasionally wrong, never guaranteed. Whenever the task demands a ",[51,97,98],{},"specific, exact, provable outcome",", a prediction is the wrong shape of answer no matter how accurate the model gets.",[33,101,102],{},"Calculating sales tax is the clean example. Tax follows published rules, and the figure on the invoice must be exactly correct and defensible in an audit. A model that is \"99.7% accurate\" on tax is a broken system, because the 0.3% is a legal problem. The right tool is a rule: look up the rate, multiply, done. The same logic covers computing an account balance, applying a fixed discount, checking a password, or enforcing a hard business policy. Known logic plus a required exact answer equals write the rule, not train the model.",[33,104,105,106],{},"Contrast that with forecasting next quarter's sales. No rule can produce the exact number, an approximate prediction is genuinely useful, and being close is a win. That is ML's home ground. The test to carry into the exam: ",[51,107,108],{},"if the task needs one provably correct answer from known logic, reject ML; if it needs a good estimate of something uncertain, ML is a candidate.",[40,110,112],{"id":111},"other-times-to-say-no","Other times to say no",[33,114,115],{},"Beyond the exact-outcome rule, four signals mean a problem does not need ML.",[117,118,119,131,137,143],"ul",{},[120,121,122,125,126,130],"li",{},[51,123,124],{},"The logic is simple and known."," If you can write the rule in a few lines and it will not change, write it. A model to decide whether a number is even is absurd; ",[127,128,129],"code",{},"n % 2 == 0"," is exact, free, and permanent.",[120,132,133,136],{},[51,134,135],{},"You do not have enough quality data."," ML learns from examples. Forty partial records will not teach a reliable churn model. No data, no pattern, no model, regardless of how good the use case sounds.",[120,138,139,142],{},[51,140,141],{},"Errors are intolerable and uncheckable."," If a wrong output causes real harm and no human reviews it, the probabilistic nature of ML is a liability. Either add human review or use a deterministic method.",[120,144,145,148],{},[51,146,147],{},"You cannot define success."," If you cannot state what a good prediction looks like or measure it, you cannot train or evaluate a model. A fuzzy goal produces a fuzzy, untrustworthy system.",[40,150,152],{"id":151},"a-quick-decision-guide","A quick decision guide",[154,155,156,172],"table",{},[157,158,159],"thead",{},[160,161,162,166,169],"tr",{},[163,164,165],"th",{},"Ask",[163,167,168],{},"If yes",[163,170,171],{},"Points toward",[173,174,175,186,195,204,214,223],"tbody",{},[160,176,177,181,183],{},[178,179,180],"td",{},"Is the pattern too complex or changing to write as rules?",[178,182],{},[178,184,185],{},"ML",[160,187,188,191,193],{},[178,189,190],{},"Is an approximate, occasionally-wrong answer acceptable?",[178,192],{},[178,194,185],{},[160,196,197,200,202],{},[178,198,199],{},"Do you have enough relevant, reasonably clean data?",[178,201],{},[178,203,185],{},[160,205,206,209,211],{},[178,207,208],{},"Must the output be exactly correct and auditable every time?",[178,210],{},[178,212,213],{},"rules, not ML",[160,215,216,219,221],{},[178,217,218],{},"Is the logic simple, known, and stable?",[178,220],{},[178,222,213],{},[160,224,225,228,231],{},[178,226,227],{},"Can you clearly define and measure a good result?",[178,229,230],{},"If no",[178,232,233],{},"neither yet, fix the goal first",[33,235,236],{},"Machine learning needs the top three to lean yes and the exact-outcome trap to be absent. Miss any of those and the honest answer is that ML is the wrong tool here.",[40,238,240],{"id":239},"exam-tips","Exam tips",[33,242,243],{},"The exam sets traps that reward the exact-outcome rule and punish \"AI everywhere\" enthusiasm.",[117,245,246,249,252,255],{},[120,247,248],{},"The strongest keyword for rejecting ML is \"a specific outcome is needed\" or \"an exact/precise result.\" When you see it, the answer is that ML is not appropriate; pick the rule-based or deterministic option.",[120,250,251],{},"\"Cost-benefit\" in a stem usually signals that the correct answer weighs ongoing maintenance and data costs, not just accuracy. Watch for an option that mentions the total lifetime cost.",[120,253,254],{},"Value patterns to recognize as pro-ML: \"assist decision making,\" \"scale,\" \"automate,\" \"handle high volume,\" \"detect patterns humans miss.\"",[120,256,257],{},"Do not be seduced by \"the task is repetitive.\" Repetition with simple fixed logic is a script, not a model. The exam plants this to see if you conflate automation with ML.",[33,259,260],{},"The one rule to carry forward: machine learning trades exactness for the ability to handle complexity and scale, so it fits when the pattern is hard and an estimate is useful, and it fails when the answer must be exactly right from known logic. Once you have decided ML is the right tool, the next question is which technique fits the problem, and that is where the next lesson on regression, classification, and clustering begins.",{"title":262,"searchDepth":263,"depth":263,"links":264},"",3,[265,267,268,269,270,271],{"id":42,"depth":266,"text":43},2,{"id":74,"depth":266,"text":75},{"id":87,"depth":266,"text":88},{"id":111,"depth":266,"text":112},{"id":151,"depth":266,"text":152},{"id":239,"depth":266,"text":240},"md",[],[23,24,25,26],{},"/courses/aws-certified-ai-practitioner/en/domains/01-ai-ml-fundamentals/02-practical-ai-use-cases/01-when-to-use-ai-and-when-not",{"passingScore":278,"questions":279},70,[280,289,299,307,315],{"question":281,"type":282,"options":283,"correctAnswer":285,"explanation":288},"A finance team must calculate the exact sales tax owed on every invoice, following published tax rates. Why is machine learning the wrong tool here?","single",[284,285,286,287],"Machine learning cannot process numbers, only text and images","The task needs a guaranteed exact outcome from known rules, not a probabilistic prediction","Machine learning is always more expensive than any alternative","Tax calculations change too slowly for a model to learn them","Tax is a fixed, known rule where the answer must be provably correct every time, so you write the rule. ML produces probabilistic predictions that are occasionally wrong, which is unacceptable for a legally exact figure. ML handles numbers fine, and cost is not the deciding factor here; the need for an exact, auditable outcome is.",{"question":290,"type":291,"options":292,"correctAnswers":297,"explanation":298},"Which of the following are patterns where AI/ML typically adds value? (Select TWO.)","multiple",[293,294,295,296],"Assisting human decisions by surfacing predictions a person then acts on","Scaling a task to millions of items that no team could handle by hand","Producing a single legally binding figure that must never be wrong","Replacing a three-line business rule that already works perfectly",[293,294],"AI/ML earns its cost when it helps people decide, scales beyond human capacity, or automates repetitive judgment. A legally binding exact figure calls for a rule, not a prediction, and replacing a working three-line rule adds cost and risk with no benefit.",{"question":300,"type":282,"options":301,"correctAnswer":303,"explanation":306},"A startup wants to predict customer churn but has only 40 historical customer records, half of them incomplete. What is the most likely problem with using ML here?",[302,303,304,305],"ML models cannot predict churn under any circumstances","There is not enough quality data for a model to learn a reliable pattern","Churn prediction requires reinforcement learning, which is banned on small data","The data must be images for ML to work","ML learns patterns from data, so too few or too dirty examples produce an unreliable model. Churn is a valid ML use case at scale, but 40 partial records cannot support a trustworthy model. The learning type and data format are not the blocker here; data quantity and quality are.",{"question":308,"type":282,"options":309,"correctAnswer":311,"explanation":314},"In a cost-benefit analysis of an ML project, which costs are commonly underestimated?",[310,311,312,313],"Only the one-time cost of training the model","The ongoing costs of data collection, monitoring, retraining, and maintenance after launch","Nothing, because managed AI services are always free","Only the salary of the person who writes the first version","A model is not a one-time build. It needs fresh data, monitoring for drift, periodic retraining, and engineering upkeep, and these recurring costs often dwarf the initial training. Teams that count only the first training run underestimate the true cost and misjudge the return.",{"question":316,"type":282,"options":317,"correctAnswer":319,"explanation":320},"AI/ML is the right choice whenever a task is repetitive, regardless of whether the underlying logic is simple and fixed.",[318,319],"True","False","Repetition alone does not justify ML. If the logic is simple, known, and fixed, a plain rule or script is cheaper, exact, and easier to maintain. ML pays off when the pattern is complex or changing and an approximate prediction is genuinely useful.",{"title":20,"description":21},"courses/aws-certified-ai-practitioner/en/domains/01-ai-ml-fundamentals/02-practical-ai-use-cases/01-when-to-use-ai-and-when-not","2PCORCFhbwGMMTsByrHmeot2qA99dkXxkaKsu8iKmkE",{"locked":5,"reason":6,"meta":325,"item":3},{"title":326,"description":327,"isFree":10,"estimatedMinutes":328,"difficulty":12,"learningObjectives":329},"Regression, Classification, and Clustering","The three workhorse ML techniques and how to pick the right one for a use case, using the shape of the answer you need: a number, a category, or a set of natural groups.",16,[330,331,332,333],"Match a use case to regression, classification, or clustering by the shape of the answer it needs","Distinguish regression from classification even when the same subject can be framed either way","Explain when clustering fits: no labels exist and you want to discover groups","Avoid the common traps of confusing a number with a category and a known label with a discovered group"]