How AI Evolved from Rule-Based Systems to Generative and Agentic AI

Artificial intelligence has not evolved through a simple sequence in which one technology replaces another. Instead, capabilities have accumulated. Systems based on rules and statistical prediction remain widely useful, generative models add the ability to produce text, images, audio, video and code, and agentic systems connect models to tools so they can pursue goals through multiple steps.

The clearest way to understand the progression is as a shift from deciding, to creating, to acting. The risks change with it: prediction systems must be tested for accuracy and fairness; generative systems must be checked for truthfulness, safety and provenance; agents must also be constrained by permissions, approval gates and audit trails.

A seven-decade evolution

The modern field took shape in the 1950s, after foundational work on machine reasoning, game playing and neural networks. Research later moved through symbolic systems and expert systems, machine learning in the 1990s and deep learning in the 2010s.

These periods are best understood as overlapping approaches. A contemporary product may use a predictive model to classify a request, a generative model to draft a response and conventional software controls to decide whether an action is allowed.

1. Traditional AI: rules, classification and prediction

Traditional AI includes two broad approaches:

• Symbolic or rule-based systems, in which people encode logic, domain knowledge and decision procedures.
• Machine-learning systems, which learn statistical relationships from historical data to classify inputs or estimate outcomes.

Early game-playing programs demonstrated how machines could search through possible moves in a bounded environment. Later systems applied similar ideas, along with statistical learning, to practical tasks such as:

• detecting unusual or fraud-like transactions;
• assessing credit and default risk;
• estimating healthcare risks such as deterioration or readmission;
• ranking products, videos, music and posts in recommendation engines; and
• combining speech recognition, intent classification and predefined workflows in voice assistants such as Siri.

These systems are often strong when the task is narrow, repeated and measurable. They can process large volumes of cases quickly and produce consistent outputs. Rule-based systems may also be easier to inspect than complex models because their decision logic is explicit.

The limits of prediction

Traditional AI does not automatically understand why an outcome occurs. A model can identify a statistical association without establishing causation, and a system trained on biased historical data can reproduce or amplify that bias. Rules can also be brittle outside the conditions for which they were designed, while complex predictive models may be difficult to explain.

Performance can deteriorate when the real world changes—a problem commonly described as data drift. In high-stakes settings such as lending, employment, insurance, policing and healthcare, responsible deployment therefore requires data-quality checks, bias testing, appropriate explanations, human review, monitoring and routes for appeal.

2. Generative AI: creating new content

Generative AI learns patterns from large collections of text, images, audio, video or code and uses those patterns to produce new outputs. Large language models generate text sequentially, while multimodal systems can work across formats.

This changed the user experience of AI. Instead of selecting from a fixed set of commands or receiving a classification score, people can describe an objective in ordinary language and ask a system to draft, transform or explain material.

Where generative AI is used

Common applications include:

• Customer service: drafting replies, summarising conversations, translating material and retrieving information from knowledge bases.
• Legal work: summarising documents, extracting clauses, comparing authorities and preparing first drafts. Human professionals still need to verify primary sources and legal claims.
• Education: explaining concepts, creating practice questions and giving feedback.
• Creative work: brainstorming, drafting copy and generating images, music, video concepts and software code.
• Programming: explaining code, writing tests, suggesting fixes and accelerating routine development.

The main benefits are speed, lower drafting costs, natural-language interaction and the ability to produce multiple personalised iterations. But generating a plausible answer is not the same as exercising reliable expert judgement.

The main risks

Hallucinations and overconfidence. Large language models can produce fabricated, inaccurate, biased or oversimplified information in a confident tone. Research on grounding and evaluation identifies hallucinations, harmful content, manipulation, copyright infringement, privacy and security as issues that require testing and monitoring.

Learning and academic integrity. A large school field experiment found that AI tutoring improved performance while the tool was available, but students who relied on it without suitable safeguards could perform worse when access was removed, suggesting reduced skill acquisition. Teacher-designed hints were more protective than simply giving away answers. Generative tools can also make plagiarism easier or obscure whether a student has mastered the work.

Copyright and provenance. Questions about training-data provenance, stylistic imitation and outputs that reproduce protected material remain contested. Education guidance identifies intellectual-property infringement as a significant risk alongside inaccurate, biased or unsafe content.

Harmful and manipulative content. Generative systems can produce misinformation, harassment, sexual or violent material and dangerous instructions. Safety guidance recommends testing, safeguards, activity logging and monitoring rather than assuming that a model will always follow its intended behaviour.

Youth safety and wellbeing. Research on generative AI risks for young people identifies concerns involving mental wellbeing, behavioural and social development, toxicity, privacy and misuse or exploitation. These risks are especially serious when users treat a chatbot as a trusted substitute for qualified human support.

Privacy and confidentiality. Prompts may contain personal information, trade secrets, client material or legally privileged content. Organisations need clear rules about which data may be entered into which services, and users should treat unapproved systems as unsuitable for sensitive information.

Practical controls for generative AI

A reasonable baseline is to treat model output as a draft rather than evidence. Important factual claims should be checked against authoritative sources, and human review should be required for medical, legal, financial, employment and safety-critical decisions.

Useful controls include retrieval grounded in approved source material, citations, red-team testing, content safeguards, age-appropriate design, audit logs, user disclosure and policies covering data handling, academic use, provenance and intellectual-property clearance. Government guidance similarly highlights inaccurate or biased content, privacy breaches, IP infringement and academic-integrity challenges.

3. Agentic AI: systems that can act

Agentic AI adds an action layer. An agent can interpret a goal, break it into steps, call tools such as browsers, code repositories, databases or messaging systems, inspect intermediate results and continue working within the permissions it has been given.

That makes an agent different from a chatbot even when both use a large language model. A chatbot may draft an email; an agent may search for information, prepare the message, send it and update another system. A coding assistant may explain a bug; an agent may investigate a codebase, edit files, run tests and prepare a change for review.

Named examples in the current market include software-development agents such as Google Jules, recruiting workflow tools such as LinkedIn Hiring Assistant and personal-agent frameworks such as OpenClaw. OpenClaw is described by its project materials as a personal, open-source assistant that runs on a user’s devices and connects through channels and tools; what it can actually do depends on the integrations and permissions enabled by the user.

Why agents can be useful

Agentic systems can:

• complete workflows rather than isolated drafting tasks;
• reduce routine coordination and administration;
• connect information across fragmented business systems; and
• operate continuously within defined limits.

This could make complex workflows more accessible to smaller organisations and individual users. It also changes the central question from “Is the answer accurate?” to “What is the system allowed to do, and how can its actions be checked?”

Why autonomy increases the risk surface

A generative model can be wrong in an answer. An agent can turn a wrong inference, a malicious instruction or a compromised tool into an external action.

For example, an agent with financial access could make an unauthorised payment, purchase, transfer or cancellation. An agent with email, browser or cloud access could be manipulated by a malicious webpage or document into exposing credentials or sending confidential files. Excessive permissions, weak identity controls and poorly separated duties can magnify a small model error into a serious incident.

Agent accountability is also harder to assign when an outcome emerges from interactions among the model, instructions, data, tools and other agents. The system’s operator therefore needs visibility not only into the final result, but also into the proposed action, the evidence behind it and the permissions used.

Governance must scale with capability

The same basic principles apply across all three stages: assign an accountable owner, limit the system to a defined purpose, protect data, test before deployment, monitor performance, prepare for incidents and provide meaningful human review or override for consequential decisions.

Agentic systems need additional controls because they can affect external systems. Practical safeguards include:

• least-privilege and time-limited credentials;
• separate accounts and clear identity for each agent;
• explicit boundaries covering what the agent may read, write, send, buy, delete or deploy;
• confirmation gates for irreversible or high-impact actions;
• approved-tool lists, sandboxing and protection against prompt injection;
• transaction limits and anomaly detection;
• detailed, tamper-resistant action logs;
• segregation of duties and proportional human approval; and
• a rapid pause or kill switch.

Singapore’s Model AI Governance Framework for Agentic AI, launched on 22 January 2026, presents guidance for organisations deploying these systems rather than a standalone AI statute. Its approach places human and organisational accountability at the centre and recommends risk management and controls proportionate to an agent’s level of autonomy.

For financial services, the Monetary Authority of Singapore’s SAFR paper describes a runtime-governance approach in which a structured record captures what an agent intends to do, why it intends to do it and what evidence supports the proposed action.

The key takeaway

AI’s evolution is not simply a march from “old AI” to “new AI.” Rule-based and predictive systems remain the right choice for many bounded decisions. Generative AI adds flexible production and natural-language interaction. Agentic AI adds the ability to pursue goals through tools and external actions.

The governance challenge grows with that capability. Prediction demands attention to accuracy, bias and drift. Generation demands verification, safety, privacy and provenance. Action demands all of those measures plus permission controls, approval gates, accountability and recovery plans. The most reliable AI products will therefore combine model capability with disciplined system design and human oversight.