Skip to content
All insights

Agentic AI

Chatbots answer. RPA repeats. Agents decide.

Buyers are asked to choose between a chatbot, a robotic process automation program and an AI agent as though they were competing products. They are not. They fail in different places, and knowing where saves a year.

Published December 16, 20257 min read

What this covers

  • A chatbot produces language. Its output still has to be acted on by a person, which is why it rarely removes a step from an operational process.
  • RPA reproduces a fixed sequence of clicks. It is excellent where inputs are identical every time and brittle the moment they vary.
  • An agent handles variation by deciding, which is why it suits work arriving as email and documents rather than as structured records.
  • The three overlap less than the market implies. Most operations have work that suits each, and the mistake is picking one tool for all of it.

Three things get pitched into the same budget line. A chatbot on the website or in the service desk. A robotic process automation program moving data between two screens. An AI agent taking over a workflow. The pitch decks look similar and the demonstrations all end with something appearing on a screen, so the natural question is which one to buy.

It is the wrong question, and it leads to a predictable outcome: a tool bought for work it structurally cannot finish, and eighteen months spent concluding that the technology does not work. A better question is narrower. For a specific piece of work, what does each approach leave for a person to do?

The three, described by what they produce

ChatbotRPAAI agent
What it producesLanguage: an answer, a summary, a draftA repeated sequence of actions in an interfaceA completed transaction in a system of record
What it needs to startA question in a chat windowAn input that is identical every timeA request in whatever form it arrived
How it handles a case it has not seenAnswers anyway, plausiblyFails, or worse, proceeds incorrectlyEvaluates it, or stops and hands it over
Who acts on the outputA personNobody: the action was the outputThe agent, within its limits
What breaks itAnything requiring your live dataA layout change, a new field, an exceptionMaster data nobody can resolve, or an undefined rule

Why chatbots rarely reduce operational work

A chatbot is a good product for a bad fit. It is genuinely useful where the answer already exists in written form and the user simply cannot find it: policy questions, documentation, first-line support on a known product. In that shape it deflects work that was only ever going to be a lookup.

Operational work is a different shape. When a customer asks whether their order will ship on Thursday, the answer is not in a document. It is in the order record, the stock position and the carrier booking, and it changes hourly. A chatbot without access to those systems will produce a fluent sentence that nobody can rely on. A chatbot with access to them has quietly become an agent, and the interesting questions are now about permissions and stopping points rather than about conversation design.

The chat window is a user interface, not a capability. What matters is what sits behind it and what it is allowed to touch.

There is a second, quieter problem. A chat interface requires a person to show up and ask. Most operational work does not arrive that way. It arrives as an email with an attachment, at 17:40, addressed to a shared mailbox. Nobody is going to copy it into a chat window first.

Why RPA works until it does not

Robotic process automation earned its place. Where a process is genuinely identical every time, a program that reproduces the clicks is cheap, fast and completely predictable. Moving a fixed export between two systems every night is work RPA should be doing, and an agent would be a worse answer.

The limitation is structural rather than technical. RPA follows a sequence that somebody wrote down. Every case the sequence did not anticipate is an exception, and exceptions in operational work are not rare events. They are a standing share of volume: the customer who sends a revised quantity in a follow-up email, the invoice with a freight line the purchase order did not carry, the order with a delivery address that is not the billing address.

Teams respond by writing more branches. Each branch is more maintenance, and the maintenance grows faster than the coverage. The familiar end state is an automation that handles seventy percent of cases and a person who handles the thirty percent plus the automation.

What the agent adds, and what it costs

An agent handles the exception rather than failing on it, because it is not following a script. It reads the request, resolves it against live master data, applies the rules and reaches a conclusion. A revised quantity in a follow-up email is not a new branch anybody has to write. It is a request, read the way a person would read it.

That flexibility has a price, and it is worth being explicit about it. A scripted automation is deterministic: you can read the script and know what it will do. An agent reasons, so its behavior is governed by rules and boundaries rather than by a transcript of steps. That is why the parts of a deployment that look least exciting matter most: what it may write, what it must hold, what threshold routes a case to a person, and a log complete enough to reconstruct any single transaction.

Done properly this is not a weaker guarantee than RPA's, it is a different one. RPA guarantees the steps. An agent guarantees the boundary: here is what it is permitted to commit, here is where it stops, here is the record of both. Where an agent stops covers how that boundary is designed.

Matching the approach to the work

Three questions sort most workflows correctly.

  1. 01Does the input vary? If every input is structurally identical, RPA is cheaper and more predictable. If inputs arrive as email, PDFs and spreadsheets written by other companies, only an agent will cope.
  2. 02Does the output have to land in a system? If the work ends with a record created or updated, a chatbot cannot finish it. If the work ends with a person better informed, a chatbot may be exactly right.
  3. 03Can somebody state the rules? If correctness can be defined, an agent can check it and act. If correctness is a judgment call that experienced people disagree on, the honest answer is that this step stays with a person, whatever you buy.
WorkFitsWhy
Nightly file transfer between two systemsRPAIdentical every run, no interpretation required.
Policy and documentation questions from staffChatbotThe answer exists in writing and does not change hourly.
Inbound RFQs with specification attachmentsAgentEvery request is written differently and has to be resolved against the catalog and contract pricing.
Supplier invoices matched to receiptsAgentLayouts vary by supplier; tolerances and exceptions are rules, not steps.
Customer asking where a shipment isAgentThe answer is in live data, and the reply has to be written and sent.

Where the three coexist

The most common healthy end state is not one technology. It is a scheduled transfer still running as a script, a knowledge assistant answering internal questions, and agents handling the workflows where documents arrive and transactions have to be created. Each is doing what it is structurally good at.

What changes is the center of gravity. The work that consumes operational headcount in a distributor, a manufacturer or a logistics provider is almost entirely the third kind: documents in, decisions made, records updated. That is the work our solutions are built around, and it is the work neither of the other two approaches can finish on its own.

Questions

Questions we get about this.

We already have an RPA program. Does an agent replace it?

Usually not. Scripted automations that run reliably on identical inputs should keep running. The agent takes the work those scripts could never cover: the requests that arrive as email and documents, and the cases that currently fall out of the script as exceptions.

Could we not just connect our chatbot to our ERP?

You can, and at that point the design questions change entirely. Once something can write to a system of record, what matters is permissions, validation rules, approval thresholds and an audit log, not conversational quality. That is an agent project with a chat interface on the front.

Is an agent less predictable than a scripted automation?

Its path is less predictable; its boundary should not be. A scripted automation guarantees the sequence of steps. A well-built agent guarantees what it is permitted to commit, what it must hold for a person, and that every action is logged. For work with variable inputs, the second guarantee is the useful one.

Which is cheaper to run?

For a narrow, stable, high-volume transfer, a script is cheaper and should stay. For work with variable inputs, the comparison is misleading, because the script cannot do the job at all without a person covering the exceptions. Compare total handling cost including the exception queue, not license against license.

How do we decide for a specific workflow?

Ask whether the input varies, whether the output has to land in a system, and whether somebody can state the rule for a correct outcome. Those three answers place almost every operational workflow correctly, and a Workflow Review walks a real one through them.

See what this would look like in your operation.

A Workflow Review traces one of your real workflows end to end and names the first step an agent could take over. No platform to evaluate first.

Start with one workflow

Fixed scope for the first workflow, agreed before we start. Expansion is your decision once it proves value.

How it works

Security, stated plainly

Where the agents run, what they can and cannot do, and exactly what we do and do not claim.

Security