An AI agent is an AI system that decides its own next step, picks the right tool, and runs a multi-step task on a person's behalf, like booking a meeting or chasing an invoice, without being scripted step-by-step.
The cleanest way to read it is the goal-versus-steps line. In a workflow you write the steps; in an agent you give it a goal and a toolbelt and it picks the steps itself. Anthropic draws that line in Building Effective Agents: a "workflow" is a system where an LLM and tools are orchestrated through pre-defined code paths, an "agent" is a system where the LLM "dynamically directs its own processes and tool usage, maintaining control over how they accomplish tasks". Same underlying model, different control surface. An agent sits one rung above AI Automation, which sits above Business Process Automation, but the rung that matters is the one where the software stops being told what to do next.
That tradeoff is the point. Agents handle work whose path isn't known in advance, the lead came in via SMS asking something the runbook doesn't cover, and they handle exceptions a script would drop. They also cost more, run slower, and fail in less predictable ways than a fixed workflow, which is why Anthropic's recommendation is "finding the simplest solution possible, and only increasing complexity when needed". Australian investment is following the same curve: the Australian Bureau of Statistics reports business R&D spend on AI hit $668.3 million in 2023-24, a 142% increase since 2021-22, the fastest-growing R&D category in the country. McKinsey's State of AI 2025 calls agents the next adoption wave, but notes enterprise tangible-EBIT impact is still rare, the technology is ahead of the playbooks.
A concrete one. A buyers agency runs an inbox agent built with Claude plus a calendar, CRM and Gmail tool. A prospect emails "can we chat next week", the agent reads the prior thread, checks the principal's calendar, proposes two slots, and replies, no human in the middle until the prospect picks one. The same shape is now landing inside consumer products too, ChatGPT Search is an LLM browsing the web for the user without being told which sites to visit. Scoping which jobs are actually agent-shaped is the work we run inside AI Strategy & Training; the longer Australian-pricing explainer lives at what is an AI agent for business.