AI agent vs automation vs chatbot: which one your process actually needs

Every week someone asks us for “an AI agent” when what they need is a form with three rules, and someone else asks for “a simple automation” for a task that needs judgement at every step. The words matter because they change the cost, the risk and the amount of supervision the thing will need for the rest of its life. Here is how we sort it out, in the order we ask the questions.

The three things, defined by what they are allowed to do

Question 1: can you write the rule down?

If you can describe what should happen as a rule — when a form is submitted, create a record, if the country is X route to Y — you want an automation. It is the cheapest thing on this page to build, the cheapest to run and the easiest to trust, because it does exactly what it did yesterday. Most “AI” requests turn out to be three or four of these once the process is mapped.

The tell: the person doing the job today does not have to think. They copy, paste, check a field, send. A rule can do that.

Question 2: is the hard part reading, or deciding?

When the input is unstructured — an email, a PDF, a voice note — the difficult step is often just reading it: extracting the invoice number, classifying the enquiry, summarising the thread. That is a single AI step inside an otherwise ordinary workflow, not an agent. The model reads, the rules route, a person approves anything that matters. This is the pattern behind most of our AI workflow automation work and it is where the reliable wins are.

Question 3: does the work happen inside a conversation?

If the value is in answering people — customers, tenants, patients, staff — you want an assistant. The important design decisions are not about the model: they are about what it is allowed to answer from (your documents, your live systems, nothing else), what it is allowed to do (book, quote, update) and how it hands over to a person when it is unsure. An assistant that can act is worth ten that can only talk.

Question 4: does the task need many steps and judgement between them?

Only now do we reach agents. An agent is given a goal — research this lead and draft a first email; reconcile these invoices; onboard this client — and works out the steps, calling tools, checking its own results and asking for help when the risk is real. It is the most capable option and the one most often deployed badly, because people skip the boring parts: a scope (which tools and records it may touch), a budget (spend per task), checkpoints (a person approves anything irreversible) and an evaluation set (real cases with the expected outcome).

An agent without a budget is not autonomous. It is unsupervised.

A worked example

A firm asked for “an AI agent for client intake”. The mapped process had nine steps. Six were rules (create the record, send the acknowledgement, file the attachment). Two were reading (extract the parties and matter type from a messy email; check them against the conflicts list). One needed judgement (draft an engagement letter with the right fee scale). The build was one workflow with two AI steps and one human approval — a third of the cost of the agent they had budgeted for, and easier to audit. When they later wanted the follow-up chasing to run on its own across several systems, that became a small, bounded agent.

How to decide in one minute

• Write the rule down. If you can, it is an automation.
• If the hard part is reading unstructured input, it is a workflow with an AI step.
• If the value is in conversations, it is an assistant — with actions and a hand-off.
• If it needs many steps and judgement between them, it is an agent — with scope, budget, checkpoints and an evaluation set.
• Whatever it is: build it in your own accounts, with logs and alerts, and make sure you own it.