Thesis: “AI” is not one thing. It spans a wide spectrum—from basic analytics that explain the past, to predictive and prescriptive systems that optimize the present, up to Generative AI and autonomous agents that produce content and orchestrate actions. Treating this landscape as a single category leads to mismatched expectations, wasted spend, and governance gaps. A clear capability map lets you invest with intent.

---

1) A practical taxonomy (C-level view)

Layer	What it does	Typical questions	Core techniques	Business value (examples)
Descriptive Analytics	Summarizes what happened	What happened? Where? How much?	SQL, dashboards, basic stats	KPI tracking, regulatory reports, operational visibility
Diagnostic Analytics	Explains why it happened	Why did X move? Which drivers?	Multivariate analysis, causal inference, SHAP	Root-cause of churn/cost spikes; driver analysis for sales
Predictive AI/ML	Forecasts what will happen	What will happen next?	Regression, tree ensembles, time series, deep nets	Demand forecasting, risk scoring, fraud propensity
Prescriptive AI/Optimization	Recommends what to do	What’s the best action now?	Reinforcement learning, optimization, causal uplift	Pricing & promotion optimization, next-best-action
Cognitive AI (NLP/CV)	Understands text, speech, images	What does this mean?	Transformers, embeddings, ASR, OCR	Claims triage from documents, KYC doc parsing
Generative AI	Produces content or code	Draft, summarize, translate?	LLMs, diffusion, instruction tuning, RAG	Assisted writing, code co-pilot, contract review
Agentic AI	Executes multi-step goals via tools	Can it do it for me?	LLM planners, tool-calling, feedback loops	Automated support flows, back-office automations

Reality check: these layers stack. Better descriptive/diagnostic data ≈ better predictive models ≈ safer generative outputs.

---

2) Choosing the right capability for the job

• If you need situational awareness → start with descriptive/diagnostic. You’ll de-risk downstream ML and GenAI.
• If outcomes are measurable and repetitive → predictive/prescriptive will yield hard ROI (conversion, loss ratio, throughput).
• If work is text- or image-heavy → cognitive + generative boost throughput (summaries, drafting, extraction).
• If processes cross multiple systems → agentic AI can orchestrate steps via APIs and workflows—once guardrails exist.

---

3) Illustrative scenarios across the spectrum

• Customer operations

• Descriptive: weekly CSAT, AHT, backlog.

• Diagnostic: drivers of dissatisfaction by issue type/segment.

• Predictive: churn propensity; next-best-offer.

• Prescriptive: optimal retention action given customer value.

• Cognitive/GenAI: summarize long email threads; propose replies.

• Agentic: auto-create tickets, check entitlements, schedule callbacks.

• Finance & risk

• Descriptive: close/variance analysis; liquidity dashboards.

• Predictive: cash-flow forecasting; delinquency risk.

• Prescriptive: hedging or collections strategy optimization.

• Cognitive/GenAI: policy summarization; IFRS note drafting.

• Agentic: assemble regulator packs, validate figures, file to portals.

• Healthcare

• Descriptive: readmission rates; length-of-stay by cohort.

• Predictive: readmission risk; no-show likelihood.

• Prescriptive: care-path optimization (resource constraints).

• Cognitive/GenAI: clinical note summarization; patient comms.

• Agentic: pre-visit planning, order sets, benefits verification.

---

4) Architecture patterns that scale across types

• Data Design first: canonical entities, data contracts, lineage. Garbage in → hallucinations and model drift out.

• Two rails for intelligence:

• Structured rail (features, labels, feature store → ML).

• Unstructured rail (documents, vectors, RAG → LLM).

• Orchestration & tools: workflow engine, model registry, API gateway, tool-calling for agents.

• Trust layer: policy-as-code, PII controls, red-teaming, eval harness (LLM + ML), human-in-the-loop.

• FinOps & SRE-AI: cost per inference/query, SLOs, error budgets, rollback plans.

---

5) Governance and risk by maturity level

Capability	Key risks	Essential controls
Descriptive/Diagnostic	Misleading metrics	Data contracts, metric definitions, lineage, access control
Predictive/Prescriptive	Bias, drift, leakage	Model cards, drift monitors, explainability, A/B guards, rollback
Cognitive/GenAI	Hallucinations, sensitive data exposure	RAG with citations, prompt/access policies, PII redaction, logging
Agentic AI	Unsafe actions, tool misuse	Allow-list tools, simulation sandboxes, step-approval, guardrails (policies & KGs)

---

6) KPIs to keep everyone honest

• Analytics/ML: time-to-insight, AUC/PR, lift vs. baseline, PSI/JS drift, cost per prediction.
• GenAI: faithfulness/factuality, citation rate, red-team score, cost per 1k tokens, tool-call success.
• Agentic: end-to-end task success, human handoff rate, error recoveries, cycle time vs. SOP.

---

7) Investment roadmap (quarterly increments)

Q1 — Foundation & quick wins

• Clean KPIs (descriptive), driver analysis (diagnostic).
• First predictive model in production (1 narrow use case).
• GenAI pilot with RAG + citations on internal docs.

Q2 — Scale and harden

• Feature store + evaluation harness (ML).
• Expand GenAI to 2–3 departments; implement prompt/PII policies.
• Start agentic pilots with tool allow-list and step approvals.

Q3 — Optimize and automate

• Prescriptive optimization for one core process (pricing, staffing, routing).
• GenAI cost/performance tuning; telemetry dashboards.
• Agentic workflows for routine back-office tasks.

Q4 — Industrialize

• Cross-domain models (uplift/causal), model registry + CI/CD.
• GenAI guardrails with knowledge graphs; enterprise eval suite.
• Agentic automation with measurable SLAs and rollbacks.

---

8) Common pitfalls—and how to avoid them

• “Jump to GenAI” without data basics → fix data contracts, metrics, lineage first.
• Endless POCs → define production-grade exit criteria (SLOs, ROI, controls).
• Monotechnologies → prefer composable architecture; avoid vendor lock-in.
• Paper governance → encode policies and evaluations in code (pipelines, gateways, CI).