Domain 2 of 5 · Chapter 2 of 3

GenAI Capabilities & Limitations

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing AWS Certified AI Practitioner AIF-C01 premium course — no separate purchase.

Included in this chapter:

  • Capabilities that create business value
  • Limitations, mitigations, and model selection
  • Exam-pattern recognition

Choosing a GenAI model by business need: what to weigh

Selection criterionLarger / higher-capability FMSmaller / distilled FMWhy it matters to the business
Accuracy / qualityGenerally higher on complex, cross-domain tasksAdequate for narrow, well-scoped tasksDrives task success, customer satisfaction, and trust
LatencySlower per responseFaster, better for real-time UXAffects conversion rate and user experience
CostHigher per token (more spend at scale)Lower per tokenDirectly drives ROI under token-based pricing
ModalityMore likely multimodal (text, image, etc.)Often single-modalityMust match the input/output the use case needs
Customization & complianceMore fine-tuning / grounding optionsSimpler, fewer knobsDetermines fit for regulated or domain-specific needs

Decision tree

Need fully auditable,explainable decisions?YesInterpretable / classic MLFMs are low-interpretability; not GenAINoNeed exact, deterministic,reproducible output?YesRules engine / classic ML;or lower temperature toward 0NoNeed verifiable, up-to-datefacts (beyond knowledge cutoff)?YesGround with RAGretrieval + cite sources, curbs hallucinationNoRisk of harmful or wrongoutput with low error tolerance?YesGuardrails + human reviewconstrain output before it shipsNoGenAI fitsFM via Bedrock;pick model on cost vs accuracyAlways: judge value by business metrics (ROI, cost per interaction, conversion)and choose the cheaper non-GenAI path when it meets the requirement

Cheat sheet

  • GenAI's three advantages: adaptability, responsiveness, simplicity
  • One foundation model generalizes across tasks with no per-task retraining
  • Invoke a pretrained FM through one managed API, no infrastructure or training
  • Hallucination: a fluent, confident answer that is factually wrong
  • Knowledge cutoff: the model knows nothing after its training date
  • Nondeterminism: the same prompt can return different answers
  • Temperature trades off deterministic vs. random output
  • Interpretability: an FM is a low-interpretability black box
  • Foundation models inherit training bias and degrade as data drifts
  • No single best model: selection balances competing factors
  • A smaller or distilled model can be the right business answer
  • Token-based pricing ties spend to model choice and prompt length
  • Modality must match what the use case produces and consumes
  • Compliance and regulatory needs can rule a model out outright
  • GenAI value is measured by business metrics, not model scores
  • Model-quality scores answer 'is it good'; business metrics answer 'is it worth it'

Unlock with Premium — includes all practice exams and the complete study guide.

References

  1. What is Generative AI?
  2. What is Amazon Bedrock?
  3. Amazon Bedrock
  4. Amazon Bedrock Knowledge Bases
  5. Inference parameters for foundation models
  6. AWS Certified AI Practitioner (AIF-C01) Exam Guide
  7. Amazon Bedrock Pricing