2.1 AI Taxonomy & Core Technologies

2.1 AI Taxonomy and Core Technologies

Core principle: AI terms describe fields, methods, and capabilities that overlap. Use them to understand what a system can do and how to evaluate it, not as a marketing ladder or a reason to surrender human judgment.

---

A Useful, Non-Nested Map

Artificial intelligence (AI) is the broad field of building systems that perform tasks associated with perception, language, prediction, planning, search, or action. AI also includes rule-based and search methods that are not machine learning.

Machine learning (ML) learns patterns from data to perform a task. Deep learning is a family of ML methods based on multilayer neural networks. Many current generative models use deep learning, but not all generative systems must be described by one architecture, and deep learning includes many non-generative systems.

Generative AI creates samples—text, code, images, audio, video, or structured data—based on learned distributions and current context. Agentic AI describes systems that pursue a goal through a control loop, often using a generative model plus memory, retrieval, planning, external tools, and observations. Agentic is not simply a smaller box inside generative: an agent may combine language models, deterministic software, search, rules, and predictive models.

The Scrum.org primer also names supervised, unsupervised, reinforcement, diffusion, RAG, fine-tuning, tokens, LLMs, multimodal models, ANI, AGI, and GPT. Know what each term helps explain without assuming the list is a strict hierarchy.

Learning Paradigms

These categories can combine. A system may use supervised fine-tuning, preference optimization, retrieval, and an autoregressive generator in one product.

Predictive and Generative Tasks

A predictive model estimates a class, value, risk, or probability from features. A generative model creates a sample conditioned on input. The distinction is about the task, not whether one family is deterministic and the other random. Both can be stochastic, badly calibrated, biased, or wrong. An LLM with a calculator or forecasting tool can help explain a forecast, but free-form text generation alone is not evidence of a valid probability.

For Scrum work:

• use verified delivery data and an appropriate forecasting method when estimating likely ranges;
• use a generative model to draft questions, alternatives, code, tests, or summaries;
• validate each output with measures appropriate to the product and risk.

ANI, AGI, and ASI

Artificial Narrow Intelligence (ANI) commonly refers to systems with bounded capabilities rather than human-level general competence. Artificial General Intelligence (AGI) and Artificial Superintelligence (ASI) have no single operational definition accepted across the field. They are not useful labels for deciding whether a current output is correct. Evaluate observed capability, failure modes, authority, evidence, and context.

A fluent model can perform many tasks and still fail on a simple factual or causal question. Calling it intelligent does not make it a Scrum Team member, a legal person, or an accountable professional.

Agentic Systems and Bounded Autonomy

An agent may plan steps, call a search service, edit files, run tests, and retry after an error. Tool feedback can improve results, but self-reflection is not independent assurance. The system can choose the wrong goal, misuse a tool, repeat a bad action, expose data, or optimize a proxy. Controls include least-privilege access, action limits, approval for consequential operations, observable logs, deterministic validators, rollback, and a stop condition.

The needed oversight is risk-based. Automatically formatting public text is different from deploying code, contacting a customer, or making an employment recommendation.

Economic Potential Is Not a Guaranteed Benefit

Generative AI can reduce effort, expand exploration, or enable a new service. It can also add review load, errors, licensing cost, security risk, or rework. Treat value as a hypothesis: define the user or team outcome, compare with a baseline, count verification cost and failures, and inspect what changed. Prompt volume and generated lines of code are activity measures, not product value.

Scrum Accountabilities Stay Human

The Product Owner remains accountable for maximizing value and effective Product Backlog management. Developers create the Sprint Backlog, instill quality through the Definition of Done, adapt their plan, and hold one another accountable. The Scrum Master establishes Scrum and supports team effectiveness. AI can supply evidence or options and may automate bounded steps; it cannot hold one of these accountabilities. Legal and contractual duties remain separate questions governed by applicable agreements and law.