Agents vs. RPA vs. conversational assistants

An enterprise agent reasons over a goal, uses tools and executes actions in your systems of record. RPA replays a fixed sequence of clicks. A conversational assistant answers questions and stops there. The three overlap in the demo and diverge in production, which is why Gartner coined the term agent washing: of thousands of vendors claiming agents, about 130 ship the real thing (Jun 2025).

A definition you can test

Something is an agent if it does three things. It reads the state of the world: a ticket, a call, a record. It decides what to do next under a policy. And it acts on a system with consequences: issues a refund, schedules a call, posts an invoice. Remove any of the three and you have something useful, but not an agent.

RPA: fast to start, brittle to keep

RPA automated the 2010s: copy this field, click that button, paste into SAP. It is deterministic, auditable and cheap per run. It also breaks the moment a screen changes, cannot handle a case it has not seen and needs a human queue for every exception. Most RPA estates end up as a maintenance budget. Agents do not replace RPA where the process is rigid and the input is clean. They replace the human queue around it.

Conversational assistants: a front door with no house behind it

The first wave of enterprise GenAI was assistants: answer the FAQ, summarize the document, draft the email. Useful, and measurable only in satisfaction. Menlo Ventures found that only 16% of enterprise GenAI deployments in 2025 qualified as real agents. The rest could talk but could not act. When a customer asks where the order is, the assistant explains the policy. The agent looks up the order in the OMS, sees the carrier exception and reschedules the delivery.

The comparison, honestly

• Input: RPA needs structured, identical inputs. Assistants take free text. Agents take free text, documents and events, and look up what is missing.
• Decision: RPA follows a script. Assistants pick an answer. Agents choose among actions under a policy, with a human approval step where you define it.
• Action: RPA acts on screens. Assistants do not act. Agents act through APIs on the system of record, with every write logged.
• Change tolerance: RPA breaks on a UI change. Assistants degrade silently. Agents need evals to catch drift, and with them they survive policy changes.
• Cost: RPA is cheap per run and expensive to maintain. Assistants are cheap and rarely move a KPI. Agents cost more to run and are the only ones you can sign against a KPI.
• Governance: RPA is auditable by design. Assistants often are not. Agents need guardrails, evals and an audit trail, which is why a platform exists.

When to use which

Use RPA when the input is structured, the rule is fixed and the volume is high: batch postings, reconciliations, migrations. Use an assistant when the outcome is knowledge, not action: internal helpdesk, policy questions, drafting. Use an agent when a person is currently reading, deciding and acting on a system, and you can measure the result: tier 1 support, collections, invoice capture, driver check calls, screening interviews.

Where the money is going

Gartner estimates $234B of enterprise SaaS spend at risk from agents by 2030 (Jul 2026), and McKinsey reports that 32% of companies decided not to buy at least one piece of software because of AI (2026). That is not happening because of better assistants. It is happening because agents do the work the software was sold to organize.

What this means for the assistant you already have

Keep it. An assistant that answers policy questions well is a good first layer, and most of our agents inherit that layer as one of their tools. The upgrade is not replacing the conversation. It is connecting it to the system of record, giving it a policy for when to act and when to ask, and measuring the result against a KPI. That is a 8 to 12 week pilot, not a rewrite.

A short test for vendors

Ask three questions. Which system does it write to? Show me the log of the last hundred actions. What does it do when it is not sure? If the answers are none, there is no log, and it guesses, you are looking at an assistant with an agent's name. Ask for the maturity label too: production with a client, pilot, or reference architecture. We publish ours on every agent page for the same reason.