Reach for a chatbot when the work is a single conversation, RPA when the work is a fixed sequence against a stable screen, and a long-running agent when the work is a multi-step goal that unfolds over days across channels. Most teams mix all three; the mistake is using one where another fits.
What each tool is good at
- Chatbots: instant answers inside one session. Low cost, no memory, no follow-through.
- RPA: deterministic, high-volume steps on systems that do not change. Brittle when screens or paths shift.
- Long-running agents: goals that require memory, judgment, waiting, and outreach across weeks.
A simple decision framework
Ask three questions. Does the work span more than one sitting? Does success depend on reacting to things that happen between actions? Does it cross more than one channel or system? If you answer yes to two or more, a long-running agent is the right shape. If you answer no to all three, a chatbot or an RPA script is cheaper and simpler.
Where the line actually falls
The dividing line is not intelligence, it is time and state. RPA and chatbots both assume the world holds still. Agents assume it does not — that people go quiet, portals change, and queues take hours — and are built to keep the job moving anyway.