Chatbot, automation, agent: the difference that matters
These three get used interchangeably and they behave nothing alike, which is why firms end up disappointed by the wrong thing.
- A chatbot responds to what you type. It has no memory of your files unless you paste them in, and it does not act on anything. Useful, bounded, and entirely dependent on you knowing what to ask.
- Traditional automation follows a fixed script. If this happens, do that. It is reliable and it is brittle: a renamed column or an unexpected document breaks it, and it cannot handle anything the author did not anticipate.
- An agent is given an objective and works out the steps, using tools it has been granted, adapting when the situation is not what it expected. That adaptability is exactly why it is powerful and exactly why it needs a review boundary.
In accounting terms: a chatbot explains a code section, automation moves a file when it lands in a folder, and an agent takes a folder of client documents and hands back an organized, tied-out workpaper with the open questions flagged.
The review boundary is the whole design
The single most important decision in an agent-native practice is where the human sits. Get that wrong and nothing else matters, because either the agent is doing things nobody checked, or a person is re-doing work the agent already did and the whole exercise was pointless.
The rule I work to: an agent may prepare, gather, organize, calculate, and propose. A person decides, approves, and signs. Anything that leaves the practice, and anything that becomes a position on a return, passes a human first. That is not a technology constraint. It follows from who is professionally responsible for the work, and no tool changes that.
Two practical consequences. First, an agent must be able to say it does not know: flagging the ambiguous K-1 is a feature, and silently guessing is a defect. Second, the proposal has to be reviewable, which means showing the work and citing where each figure came from. A number with no traceable source costs more time to verify than it saved.
Where agents genuinely earn their place
The work that suits an agent is repetitive, document-heavy, deadline-bound, and verifiable. Those four together are the tell.
- Document intake and extraction. Reading source documents into structured data. The most mature use in a practice today, and it still needs a person on the messy ones.
- Workpaper assembly. Sorting, indexing, cross-referencing, and tying out into a first-pass file. Hours of setup, nearly no judgment.
- Transaction categorization and reconciliation. The repetitive pass, with anomalies surfaced for review instead of every line.
- Reconciling against a second source. Statements against the ledger, a report against a return. Machines are better at this than people are, and they do not get bored in the third hour.
- Drafting the routine document. The engagement letter, the request list, the status update, all from facts already on file.
And the work that does not suit one: choosing a position on an unsettled question, deciding what a client should do, weighing a risk the client has to live with, or anything where being confidently wrong is expensive and hard to detect.
What this requires of the practice
Agents are unforgiving about ambiguity, which turns out to be a benefit: they surface the places a firm never actually decided how it works.
- Written conventions. If two people on your staff would file the same document differently, an agent cannot do it right either.
- Explicit permissions. An agent should reach exactly the data it needs for the task and nothing else. Folders that must never be read should be walled off in configuration, not by good intentions.
- An audit trail. What ran, on what, and what it changed. This is ordinary professional documentation, and it is the difference between an agent you can defend and one you merely trust.
- Tool tiering by data sensitivity. Which data class may reach which tool, decided in advance and written down.
How this relates to being AI-native
They are the same direction, one step apart. An AI-native firm is designed so AI handles the mechanics. An agent-native practice is what that looks like once the AI is not waiting to be prompted at each step but running the sequence and reporting back.
Most firms should get the first one right before reaching for the second. The structure, the written context, and the guardrails that make an AI-native firm work are the same things that make an agent safe to point at real client files. Skipping them does not save time, it just moves the cost somewhere less visible.