AI Agent

A chatbot responds to a message. An AI agent completes a task. The distinction is functional: a chatbot is a question-answer interface; an agent perceives a goal, decides what steps are required, executes those steps using available tools, and produces an outcome — an extracted invoice, a classified support ticket, a completed ERP entry.

This requires more than a language model. An agent needs a set of tools it can call (APIs, database queries, document extractors), a memory mechanism to maintain context across steps, and a decision layer that determines which tools to use in what order. In enterprise deployments, it also needs a governance layer that constrains what the agent can and cannot do.

Retrieval agents answer questions grounded in organisational knowledge — policies, product documentation, case history. They retrieve, synthesise, and respond with citations, without taking action in external systems. Knowledge assistants and internal Q&A tools are typically built on retrieval agents.

Workflow agents execute multi-step tasks: extract data from a document, validate it against a system record, decide on an outcome, and write the result to an ERP or CRM. These agents take real-world actions and require explicit governance: defined permissions, audit trails, and human checkpoints for low-confidence or high-stakes decisions. Accounts payable automation, order processing, and onboarding workflows are typically built on workflow agents.

Consumer AI agent frameworks prioritise capability and speed of development. Enterprise production requirements add three additional constraints. Predictability: enterprise agents need defined, testable behaviour — not emergent reasoning that is hard to debug or explain to a compliance team. Governance: every action the agent takes must be authorised, logged, and reviewable. Reliability: production agents must degrade gracefully on unexpected inputs rather than hallucinating outputs that downstream systems act on.

Meeting these requirements is what separates a demo agent — impressive, fast to build, works in controlled conditions — from a production agent that can be deployed to thousands of users handling sensitive transactions.