When does AI work become automation?

Short answer. AI work starts as help with ad hoc work: your team asks, and your AI agents help. While they work, your agents write down how they did it: where they found the information, what they decided and what they did. Work that repeats and is well defined is then automated: it runs on its own, and it is monitored. A few kinds of decision always go to a person.

This guide describes the three stages, when a task is ready to be automated, and what automated means in practice.

Three stages, from help to automation

Each piece of your AI agents’ work is in one of three stages. The stage says what your team does.

• You ask, the agent helps. Ad hoc work: a question, a one-off, a small task. Someone on your team asks, the agent does the work, and the person uses the result.
• The agent prepares, you approve. Work that repeats and is written down. The agent follows the written description and prepares the result. The person who approves checks it, corrects it or rejects it before anything happens.
• It runs on its own, and you check the results. Work that repeats, is well defined and has been automated. It runs without anyone starting it, every run is logged, and your team checks the results.

A job is rarely in one stage. Reading an invoice can run on its own while the posting still waits for approval.

What your agents write down

While they work, your agents write down where they found the information, what they learned, what they decided and what they did. That record lets them do the job again the same way.

When someone on your team corrects the agent, the correction goes into the agent’s rules and the company dictionary, a written list of what the company’s words and numbers mean. The model itself does not change, so what the agent learns stays with the company.

When is a task ready to be automated?

Condition	The question to ask	What it looks like when it is not ready
It repeats	Does the task come back, with the same kind of input in the same form each time?	A one-off, new suppliers, changing formats, documents from many sources
It is written down	Are its sources, decisions and steps written down, so it can be done the same way again?	Steps that rely on someone’s judgement no one has written down
Its exceptions are known	Do we know the ways the task goes wrong, and does each one have a path?	Surprises in the last month’s corrections
Someone owns the result	Is one named person responsible for the result?	“The team” owns it, or no one does

All four must hold. Your team decides, task by task. There is no fixed number of clean runs: a task that posts small amounts can be automated sooner than one that changes customer data.

Evals help with the decision. Evals are repeatable tests that run the task on known cases and check the result. They show whether the task works before it is automated, and that a change of model or rules is safe before it goes live. Nothing is automated by itself: a person decides, and the decision is written to the logbook.

What always goes to a person

Four kinds of decision stay with a person, however long the work has run without a mistake:

• Money leaves the company: a payment, a refund, a credit.
• Something goes outside the company: an email to a customer, a file to an authority, an order to a supplier.
• A commitment is made: a price, a delivery date, a contract term.
• A rule is unclear: the agent cannot tell which rule applies, or the rules disagree.

A person must be able to answer for each of these decisions, so they stay with a person even when the agent has never made a mistake.

What automated means here

Automated work is still watched:

• Monitored. Each run is watched, and exceptions go to a person.
• Logged. Each run is written to the logbook: what came in, what was done and with what result.
• Results checked. Your team checks the results, not each step.

If something changes, it comes back to you for approval: a new supplier, a new system, a new rule, or a correction no one expected. The agent then prepares and you approve, until your team decides the task can run on its own again.

An example: posting a supplier invoice

Step	Stage	Why
Find out why a supplier says an invoice is unpaid	You ask, the agent helps	A one-off question; a person answers the supplier
Read the invoice and pull out the lines	It runs on its own, and you check the results	It repeats, with stable input from known suppliers, checked against the order
Propose the posting (account, VAT, cost centre)	The agent prepares, you approve; later it runs on its own for known suppliers	Corrections go into the written description until the proposals stop needing them
Post an invoice from a new supplier	The agent prepares, you approve	The input is new, so the task does not yet repeat
Approve the payment	Always a person	Money leaves the company

The example is general. Every company’s tasks and stages are its own.

You stay accountable

When work runs on its own, the company is still responsible for it. A machine cannot be held accountable; a person can. Four things carry that responsibility: access the company controls, guardrails that hold, people who understand the system, and a logbook of what was done. Read more in Who is accountable when an AI agent makes an entry?