Predictive AI
PurposeEstimates a class, probability, score, ranking, or numeric outcome from observed data.
Best fitStable, bounded prediction problems with measurable ground truth.
Insurance examplesSeverity estimation, fraud propensity, triage, churn, risk scoring, recovery likelihood.
Primary outputScore, class, probability, forecast, or rank.
Knowledge and dataLearned statistical relationships in training data and engineered features.
Change methodRetrain or recalibrate when data, relationships, rules, or objectives change.
EvaluationDiscrimination, calibration, error, stability, fairness, and business outcome.
OperationsFeature pipelines, training, registry, deployment, monitoring, retraining, rollback.
Key risksBias, proxy discrimination, drift, data leakage, opacity, false positives/negatives.
Choose whenChoose when the primary question is what is likely to happen or which case belongs in a defined class.
Generative AI
PurposeCreates or transforms text, images, audio, code, or other content from instructions and context.
Best fitDrafting, summarization, extraction, conversational assistance, and knowledge synthesis.
Insurance examplesClaim summaries, correspondence drafts, policy Q&A, document extraction, adjuster assistance.
Primary outputNatural-language or multimodal content; sometimes structured output.
Knowledge and dataFoundation-model training plus prompt context, retrieval, tools, or fine-tuning.
Change methodUpdate prompts, retrieval, model, validators, or fine-tuning; re-evaluate behavior.
EvaluationGrounding, material factual accuracy, instruction following, usefulness, safety, and consistency.
OperationsPrompt/version control, retrieval, evaluations, safety filters, cost/latency, monitoring.
Key risksUnsupported content, prompt injection, sensitive data disclosure, inconsistency, overreliance.
Choose whenChoose when the primary objective is to create, transform, explain, or converse.
Agentic AI
PurposePursues a goal through planning, tool use, state, iteration, and action within defined authority.
Best fitMulti-step workflows requiring coordination across systems and conditional action.
Insurance examplesClaim intake orchestration, evidence gathering, task routing, recovery workflows, service operations.
Primary outputA completed workflow, tool actions, state changes, communications, and evidence.
Knowledge and dataModels plus instructions, retrieval, memory, tools, policies, and current system state.
Change methodUpdate the complete harness: goals, tools, permissions, plans, policies, memory, models, and controls.
EvaluationTask success, action correctness, authority compliance, side effects, recovery, cost, and human rescue.
OperationsRun traces, tool permissions, approvals, state, retries, limits, observability, rollback, incidents.
Key risksUnauthorized action, compounding errors, loops, hidden side effects, persistence, cascading harm.
Choose whenChoose only when autonomous or semi-autonomous action creates enough value to justify stronger controls.