Business Automation with AI

The first automation is not the one you think

Almost everyone starts with the most annoying task they have. That is usually the wrong one, because the most annoying task tends to be annoying precisely because it is irregular, full of exceptions, and needs a human to decide something. Start instead with the most boring task — the one that happens the same way every single time, several times a week, and that nobody enjoys owning.

A good first candidate has three properties: it triggers on something you can detect (a form submission, an incoming email, a row added to a sheet), it follows the same steps every time, and getting it wrong is cheap to undo. Copying form responses into a spreadsheet and notifying the right person qualifies. Deciding which invoices to pay does not.

Where the time actually goes

The build is the short part. Wiring a five-step workflow in n8n, Make or Zapier is an afternoon. What takes the rest of the week is everything around it: getting API access from a tool whose admin is on leave, discovering that the CRM's "date" field is a string in three different formats, and handling the case where the customer's name has an apostrophe in it.

This matters when estimating. A workflow that takes two hours to build and a day to make reliable is still worth it if it runs fifty times a month — and clearly is not if it runs twice. Count the runs before you count the steps. The honest threshold for most small automations is somewhere around an hour of manual work saved per week; below that, the maintenance will outlast the benefit.

Where AI belongs in a workflow, and where it does not

The useful mental split is between moving things and judging things. Moving a file, formatting a date, posting to a channel, writing a row — that is ordinary automation, and it should be deterministic. It either works or it visibly fails.

Judging is where a model earns its place: classifying an incoming message, summarising a long thread, pulling structured fields out of an unstructured document, drafting a reply for someone to approve. These are tasks with no single right answer, where "roughly right, reviewed by a person" beats "not done at all".

The failure mode worth naming: putting a model in the middle of a deterministic chain. If a step has one correct output, a model is a slower, more expensive and less reliable way to produce it. Parsing a date does not need intelligence — it needs a date parser.

Three things that break in production

• Silent failure. The worst automation is not the one that crashes — it is the one that quietly stops running while everyone assumes it is fine. Anything that matters needs to tell you when it did not run, not only when it did. A weekly "still alive" message is cruder than proper monitoring and far better than nothing.
• Duplicate runs. Retries, double-submitted forms and webhook redelivery all produce the same event twice. If your workflow sends an email or charges a card, it needs to recognise an event it has already handled — usually by storing the incoming id and checking before acting.
• Rate limits and quiet quotas. A workflow tested on five records behaves differently on five hundred. Most APIs will throttle you, some will simply drop requests, and the AI steps bill per call. Test at the volume you actually expect, not at the volume that is convenient to test at.

Choosing a platform, briefly

The platform question gets more attention than it deserves, because for a first workflow all the mainstream options work. The differences show up later, and they are mostly about who owns the thing: hosted tools are faster to start and priced per task or per operation, while a self-hosted tool costs a server and some maintenance in exchange for no per-run bill and data that never leaves your infrastructure.

That last point decides more real cases than any feature comparison. If the documents flowing through the workflow cannot legally leave the organisation, the hosted options are out regardless of how good they are, and the choice narrows to what you can run yourself. The full comparison is in n8n vs Make vs Zapier.

How to read the clusters above

• Quick start — for the first working workflow. Read one guide, build the boring task, and stop there for a week.
• Comparisons — worth reading once you have built something and can tell which limits you are actually hitting. Before that, comparisons are abstract.
• Agents and advanced building — the step up from "if this, then that" to something that decides its own next action. Considerably more powerful, considerably harder to keep predictable, and not where anyone should start.
• Templates and scripts — the shortcut. Useful for seeing how a finished workflow is shaped, which is often faster than reading about it.

There is no ranking of "the best automation tool" here, and no claim that any of this replaces a person. The consistent pattern in the work that survives is narrower than that: the machine handles the repetitive middle, and a human still opens the thing that matters.