THE SHORT ANSWER

Traditional automation follows predefined rules. A conversational assistant mainly responds to a person. An AI-enabled workflow uses models inside a path designed by software. An agent lets a model choose some next steps or tools within defined limits. Real systems can combine all four patterns.

Compare who controls the path

Four system patterns
PatternWho chooses the next step?Typical strengthMain control need
Deterministic automationRules and codeConsistency for known casesCorrect rules and exception handling
Conversational assistantUsually the user through dialogueExplaining, drafting and answeringClear boundaries on advice and data
AI-enabled workflowPredefined code, with model judgments inside stepsFlexible interpretation within a repeatable processValidate every model output
AI agentThe model chooses among allowed actionsAdapting when the path cannot be fully predictedPermissions, observation, limits and escalation

Evidence & context: Anthropic

These categories are a spectrum

A support system may use rules to identify a customer, a model to classify the request, a fixed workflow for refunds and an agent only for unusual investigation. Calling the whole product an agent hides the more useful question: which decisions are variable and which remain deterministic?

This page differs from AI Agent vs Chatbot. That introduction distinguishes interfaces and basic behaviour. Here the focus is how work is orchestrated behind the interface.

Choose architecture from the task

  • Payroll calculation: deterministic rules, audits and explicit exceptions usually matter more than flexible planning.
  • Research brief: an assistant can help a person explore without taking external actions.
  • Lead enrichment: a fixed workflow can call a model for classification and validate the output schema.
  • Incident investigation: an agent may select diagnostic tools as evidence changes, while write actions stay approval-gated.

These are illustrative patterns. Consequences, data access and organisational controls can change the appropriate design.

Prefer the simplest pattern that meets the need

An agent adds flexibility, but also more possible paths to test, observe and secure. Sophistication is not the number of model calls. It is the fit between variability, control and the desired outcome.

Evidence & context: Anthropic

Sources & further reading

  1. Building effective agents

    Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.

  2. Using tools

    OpenAI Developers. Official documentation showing how models can be given built-in, function and remote tools. Checked 13 September 2026; product-specific tool names and availability can change.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗