How to automate a startup, piece by piece

A ten-person company has perhaps forty small repeated tasks. You can automate maybe six of them well in a year.

Picking which six is the whole exercise. Here is the order, and it is close to the reverse of where most founders start.

The test each piece has to pass#

Before ranking anything, apply four questions. They take ten minutes each and they remove most of the list.

Is the input already written down? In one place, in a consistent shape, without a person assembling it first. If somebody has to gather the inputs before the automation can run, you have not automated the task, you have moved it.

Does it have one right answer? Filing a receipt has one right answer. Deciding whether a customer is unhappy does not. Start with the first kind.

Is being wrong cheap? A misfiled receipt costs a minute. A wrong message to a customer costs trust you do not get back at this size.

Does it happen often enough to matter? Weekly is worth automating. Twice a year is worth a written checklist.

A task that fails any of these goes on the second list. That list is longer than the first and writing it down is what stops you rebuilding the same bad idea in six months.

The order#

First: the moving of things#

Data between two systems. A signup into the CRM. An invoice into the accounting tool. A support ticket into the right queue.

This is dull, and it is where the hours are. It passes all four tests, it needs no model most of the time, and it teaches you how your own systems actually behave, which every later piece depends on.

Start here even though it is the least interesting thing on the list. Especially then.

Second: the chasing#

Following up on a document somebody has not sent. Reminding a customer their card failed. Nudging an invoice. Asking for the missing field on a form.

Chasing is repetitive, has a clear trigger, and a person doing it is doing nothing else. It is also where a small company loses money quietly, because nobody has time to chase consistently.

Third: the reading and filing#

Pulling the three figures off an invoice. Reading a CV into a structured record. Summarising a call into a note somebody will actually read.

Here a model earns its place, because the input is unstructured. Being wrong is still cheap: a wrong figure gets corrected by the person who was going to type it anyway.

Fourth: the drafting#

First-pass replies, first-pass copy, first-pass anything. A person edits and sends.

Note the words first pass. The automation produces a draft and a human is in the loop by design rather than as a safety net. That distinction is what keeps this in fourth place rather than eighth.

Fifth, and not before: anything a customer sees unedited#

Support answers, onboarding sequences, anything that acts on an account.

By the time you get here you will have learned how your automations fail, what your data is actually like, and how much correcting costs. Those three things are what make a customer-facing automation survivable, and none of them can be learned in advance.

Most founders start at fifth because it is the visible one. That is why the first attempt is often also the last.

What this looks like in practice#

Six pieces in a year, roughly.

That is deliberately unambitious. A company that finishes six pieces and can prove each one worked is in a far better position than one that started fifteen and abandoned eleven.

The two mistakes that cost the most#

Automating before instrumenting. If you cannot say how long the task takes now, and how often it happens, you will not be able to say afterwards whether the automation helped. Count it for a week first. A tally on a desk is enough.

Buying a tool per task. Six tools, six subscriptions, six logins, and nobody remembers which one does what. Prefer one place where automations live, even if it is slightly worse at each individual job.

What to do this week#

Write down every repeated task in the company. Not the important ones. The repeated ones. Aim for thirty and you will find forty.

Then mark each with a yes or no against one question only: is the input already written down? Everything that gets a no is a different project, and usually a cheaper one.

The list that survives is your order for the year. The four tests are the same ones that predict whether any AI project reaches production, and if you want a number against each candidate rather than a hunch, the way to get one is here. If you would rather somebody ran that fortnight for you, that is what the Identify step is.