For a decade, RPA sold enterprises a promise: teach a bot the clicks, and the process runs itself. It worked — for the 60–80% of transactions that follow the happy path. The other 20–40% became a second career for operations teams, forwarding exceptions to the bot's inbox and resolving them by hand.
Agentic automation — an LLM with tools, memory, and policies — closes that gap. Where RPA is rules, agents are reasoning. Where RPA breaks on novelty, agents classify and act. But agents are also stochastic, and they fail in ways RPA never did.
When to reach for an agent (and when not to)
A useful heuristic: if the process can be described as a decision tree a junior analyst walks through, and the decisions depend on reading unstructured input, an agent earns its keep. If the process is a fixed sequence of deterministic steps, RPA remains cheaper, faster, and safer.
- Reach for an agent: AP coding against policy, FNOL triage, policy admin letters, L1 IT support.
- Keep RPA: month-end close bookings, bank reconciliations, system-of-record updates, report generation.
- Combine them: an agent decides; RPA executes the clicks. This is where the best architectures live.
The five things every agent in production needs
- Bounded tools. Agents get a tight toolbox — read this queue, call that API, post to that ERP — not a browser and a prayer.
- Retrieval. Every decision is grounded in your policies, your contracts, your product catalogue. No agent reasons from training data alone.
- Human-in-the-loop. Anything above a threshold — a payment, a refund, a credit change — stops for a human.
- Observability. Every prompt, tool call, and output is logged with operator identity and cost. If you can't replay a decision, you can't operate an agent.
- A kill switch. One flag in config pauses the agent. Every SabanaTech agent ships with it.
A reference architecture
Our reference looks like this: a planner LLM routes work; a retrieval layer grounds it in your knowledge; a tool layer exposes safe actions; RPA handles clicks the tools can't reach; a guardrails layer enforces policy; an HITL layer catches the risky moves; and an observability layer records everything.
Two patterns matter more than any particular stack. First, state lives outside the agent — in a queue, in a database, in your system of record — never only in a model's context window. Second, every action the agent takes is idempotent and reversible, or it is quarantined behind a human.
The commercial logic
Agents are not cheaper than RPA per transaction today. They become cheaper when you count the exception tail RPA leaves behind. Our AP-coding agent runs at under forty cents per invoice fully loaded — half the cost of the human rework it replaced — and ships the last 15% of straight-through that pure RPA could not.
If the model is cheap and the action is reversible, agents beat humans on cost. If either assumption breaks, you need a human in the loop.
Where to start next week
Pick one process with high exception volume, clear policies, and reversible actions. Instrument the current state for three weeks. Ship a scoped agent behind an HITL review for one more. You will learn more in five weeks than in any readiness workshop.