Which tasks to hand an agent first — a five-question test

What an AI agent gets to do first is usually not decided at all. Somebody shows a demo, somebody else says "let's try it on quotes", and a week later the company has an agent on quotes.

A shame, because it is the one moment when everything else is still open. The first task sets the sentence the team will repeat about AI all year, and those sentences are short and durable: "it works", or "it guesses".

This is not the same choice as picking a process#

Two decisions get muddled here. The first is where to start with AI as a whole: which area of work to take, who answers for it, how you will know it worked. That happens in a meeting, and somebody writes it down.

The second is narrower, comes later, and is what this piece is about. Inside that process sit many separate activities, and the agent gets one or two. Almost nobody picks those deliberately — they fall out of the configuration, decided by whoever is clicking.

The five questions#

None of them disqualifies a task on its own. What counts is the balance: the more "no" answers, the further down the queue it belongs.

Can the effect be undone — and by whom. Not "is there a way in theory", but: will whoever spots the mistake reverse it themselves, in a minute, without asking anyone. A job status change, yes. A message already with a customer, no, whatever the "unsend" button promises.

How often does it happen. A daily task gives you a verdict inside the same week. With a monthly one, nobody remembers what went wrong last time by the time it comes round again. Frequency matters more than the size of the saving: at the start you are buying knowledge of where the agent is good and where it invents, not time.

Can the output be checked in fifteen seconds. It drops off the list most often and costs the most. Checking is part of the task while nobody trusts the agent, and at the start nobody does. If verifying takes as long as doing the job by hand, there is no gain — only a new chore.

Is the data the task needs already in order. Only the narrow slice: the two or three sets the agent will read. If the same fact lives in three places and nobody knows which is current, the agent will not choose better than a person — only faster, in the same confident tone. Prepare company data for AI in that one place first.

How many people have to change a habit. A task where one person starts working differently will be live on Monday. One that needs a whole team to fill in a new field will not go live at all, and that is not the agent's fault. Pick something one person does anyway.

Tasks that pass#

Common thread: the agent reads a lot, writes little or nothing, and the result is visible at a glance. Several look like a plain chatbot, which is not a criticism — same agent, same access, writing tap turned off.

Answering questions about the current state of things. How many open jobs this customer has, when the last visit was, what changed since Friday — or where the clause covering it sits in a procedure. Checking means opening one record, the cost of a miss is nil, and one habit changes: that of the person who stops searching by hand.

A daily or weekly summary for one person. The service manager gets a morning list of what is stuck and why. A mistake costs a minute of reading, and gaps show immediately to a reader who knows the subject.

A draft that goes through a human anyway. A job note, a reply to a routine enquiry, a fault description built from a phone call and photos. Reversible by definition: until somebody clicks send, nothing has happened.

Tasks that fail#

Sending anything outside the company. Emails, quotes, appointment confirmations. The first question is a straight no, and where the effect cannot be undone, human approval of AI actions stays permanently. As a first task it buys the most work for the least learning.

Bulk tidying of historical records. Tempting, because it looks like a one-off win. It fails twice: it happens once, so teaches nothing, and the result cannot be checked short of reviewing every change.

Qualifying or prioritising for a whole team. Here the fifth question bites. Everyone has to trust an ordering they did not set — and that trust gets built earlier, on tasks where it is cheap.

Anything somebody can write down as steps on paper. If the route is known and identical every time, this is not agent work but work for a plain automation — cheaper, predictable and far easier to debug.

The order that works#

Three stages. Not a schedule but a dependency: each produces the evidence the next needs.

Reading and summarising first. The agent writes nothing. This stage answers whether your data is in order, and the answer tends to be more interesting than anyone expected. A bad answer here costs you surprise, not cleanup.

Then proposals for approval. The agent prepares a write, a human clicks. The interesting part is not how many proposals go through but what you correct in the rejected ones: those corrections add up to a list of what the agent did not know and somebody in the company carries in their head. Usually sharper than the pre-launch conversations managed.

Acting alone last, and only where it is cheap. The condition for moving on is not elapsed time but boredom: the two earlier stages stopped producing surprises on this kind of action. And it applies per action, not to the whole task — one step can run alone while the next still goes through a person.

Summary#

The first task is chosen once, rarely on purpose, and it decides whether there is a second. Five questions — reversibility, frequency, checking time, the state of the data, and how many habits change — take a quarter of an hour and separate showy tasks from the ones that leave something.

The rest is order: reading, then proposals for approval, then autonomy where a mistake is cheap. The wider context — what agents do, where they fail, and which controls they need — is in the guide on AI agents for business.