AI agents vs chatbots vs RPA: which one your process needs
Use RPA when the process is fixed and runs through stable screens, a chatbot when the job is answering questions, and an AI agent when the work involves reading messy text and deciding which system to act in. Many real deployments combine them: a chatbot on the front, an agent behind it, RPA for legacy screens without an API.

Vendors use "agent" for everything now, which makes buying decisions harder than they should be. The useful question isn't which technology is newest but what kind of work the process involves.
| Aspect | Scripted bot | RPA | AI chatbot | AI agent |
|---|---|---|---|---|
| Input it handles | button presses | fixed screens and fields | free text and voice | free text, documents, system data |
| What it does | follows a menu | clicks and types like a person | answers questions | takes actions across systems |
| Where it breaks | anything off-menu | a changed screen layout | questions outside its data | unclear permissions, no log |
| Typical fit | FAQ menus | legacy systems without an API | support, pre-sales | order intake, reconciliation, reporting |
RPA
Robotic process automation records how a person clicks through an interface and replays it. It's reliable for repetitive, unchanging processes, such as copying fields from one legacy system to another every night. It's brittle when screens change, and it can't read a customer's email and work out what they want. In our market RPA is mostly useful for old 1C configurations without web services.
Chatbots
A modern AI chatbot understands free text, typos and mixed Russian-Uzbek messages and answers from your data: price list, stock, delivery terms. It talks. It doesn't usually do anything in your systems beyond maybe creating a lead.
AI agents
An agent has tools (CRM search, invoice creation, email) and decides which to call to finish a task. Take a Telegram message saying "40 bags of M400 cement to Chilanzar by Friday, invoice our LLC". The agent checks stock, prices delivery, opens a deal, issues the invoice and sets a follow-up task if payment doesn't arrive. Anthropic's "Building effective agents" draws the line between workflows that follow predefined steps and agents that choose their own; most business deployments sit closer to the workflow end, and that's usually right.
Failure modes to plan for
Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027 because of cost and unclear value. The fix is boring: one process, a metric written down before the start, minimal permissions, a log of every action and a week in shadow mode where a person approves each step.
Can one system be all three?
Often it is. The customer sees a chatbot, an agent acts behind it, and RPA covers a legacy screen without an API.
Which is cheapest?
A scripted bot. But the cheapest option that breaks on real customer messages costs more in lost orders.
How much does an agent cost?
Our single-agent pilots start at $3,900, plus model usage. See the cost breakdown article.
More on pricing: How much does an AI agent cost. Service page: AI agents for business.


