The Ultimate Guide to Automation

Automation used to mean connecting one app to another. When this thing happens, do that thing. You wrote the rules, and the machine followed them exactly — including when they were wrong.

That is no longer what most of the conversation is about.

Tech Feed Watch has covered 50 videos on automation, from Futurepedia, Y Combinator, WorldofAI and a range of practitioners. 36 of the 50 are about AI agents. Only six are about the workflow tools — Zapier, Make, n8n — that defined the field for a decade.

And 31 of the 50 discuss failure, limits or risk. Nearly two thirds. That ratio is the guide in one line: the field moved from tools that follow instructions to systems that make decisions, and most of the serious coverage is about what that costs.

The two kinds of automation

Almost every disagreement in this field comes from mixing these up, so it is worth separating them clearly.

Rule-based automation. You describe the steps. When an order comes in, copy these fields into that spreadsheet and send this email. It does exactly what you said, every time, for years. When the situation changes and your rules do not, it keeps doing the wrong thing at full speed — but it does it predictably, and you can read the rules to find out why.

Agent-based automation. You describe the goal. The system works out the steps: search, read a result, call a tool, check whether it is done, try again. An AI agent handles situations you never anticipated, which is the whole appeal. It also means you cannot read the rules afterwards, because there were none.

The trade is not subtle. Rules are limited and predictable. Agents are flexible and are not. Most of the mess in the archive comes from people choosing the second when the first would have done the job.

A useful test: if you can write the steps down, write them down. Reach for an agent when the number of possible situations is genuinely too large to enumerate — not when writing them down would just be tedious.

What people actually automate

Across the 50 videos, four areas come up repeatedly.

Customer support — 19 videos. The routine questions, the ones answered forty times a week in nearly the same words. This is the most common starting point, and the one with the clearest before-and-after.

Business operations — 22 videos. Invoices, quotes, data moved between systems that were never designed to talk to each other. Unglamorous, and where most of the hours actually go.

Marketing and sales — 12 videos. Follow-ups, lists, campaign tagging. Worth noting that the failures here are more visible than elsewhere, because the output goes straight to a customer.

Development work — writing scripts, wiring up an API, handling data. The one area where the person automating usually understands exactly what is happening.

The 31 videos about it going wrong

More than half the archive. Four patterns come up again and again, and they are not the ones people expect.

Automating something nobody measured. The most common story by a wide margin. A process gets automated, and afterwards nobody can say whether it helped, because nobody wrote down how long it took before. You end up defending a feeling.

Silent failure. A rule-based flow breaks loudly — something errors, and you find out. An agent that misunderstands does not error. It does something plausible and wrong, confidently, and keeps going. The failures in the archive are rarely dramatic: a field filled with the wrong value, a summary of a document that said something else, an email that was slightly untrue.

The 90 percent trap. A flow that handles nine cases out of ten sounds excellent until you meet the tenth. If nobody planned for the exception, it either sits unnoticed or lands on a person who now has to reconstruct what the machine was thinking. Handling the last ten percent is usually harder than the first ninety.

Counting the licence and not the setup. The subscription is the cheap part. The expensive part is the weeks of building it, the testing, and the work that stopped while people learned. A tool that pays for itself in three months usually has not counted any of that.

Keeping an agent on a leash

Since 36 of the 50 videos are about agents, this deserves its own section.

An agent decides what to do next. That is the feature. It is also why the sensible ones are wrapped in limits that sit outside the model, not inside it — because a system that is confidently wrong cannot be the thing that checks whether it is wrong. The industry term is guardrails, and in practice it means four things.

A cap on steps. Without one, a confused agent will loop, and each loop costs money and time.

A budget. Both in money and in how long it may run before someone looks.

Approval before anything irreversible. Sending, publishing, paying, deleting. Reading is cheap to get wrong; sending is not.

A log you can read afterwards. When it does something odd, you need to see what it tried. This is the one people skip, and the one they miss most.

About jobs

Twenty of the videos touch what this does to work, and they do not agree. Worth saying plainly: several of the people making confident predictions are selling something that depends on which prediction comes true.

The pattern the archive supports best is narrower than either headline. Work that was already written down as a procedure goes first — because a procedure is exactly what a machine can follow. The parts that resist are the judgement calls nobody wrote down, because they depend on context, on knowing which rule to break, and on noticing that this case is unusual.

That is neither reassuring nor alarming. It is a useful way to look at your own week: which parts could you hand to someone new with a written instruction and expect a decent result? Those are the parts under pressure first.

Starting from nothing

If you have automated nothing and want to start sensibly, the archive points at the same sequence every time.

Pick something you already do badly or slowly. Not the most impressive thing. The most annoying one.

Time it first. How long does it take now, how many times a week? Two numbers on a piece of paper. Without them you will never know whether it worked, and every later discussion becomes an argument about impressions.

Write the steps down. If you cannot describe it in writing, no tool will manage it either. This step alone often reveals that the process is not what anyone thought it was.

Automate the boring middle, not the ends. Keep a person deciding what goes in and checking what comes out, at least at first.

Run both for a while. The old way and the new way in parallel, long enough to see the exceptions. They always show up, and they show up late.

Then compare against your two numbers. Not against how it feels.

Small tools that save real time

Not everything needs an agent. A fair share of what people automate is one tedious step that a browser can handle in a second.

Cleaning up a list of addresses or entries with duplicates in it: the duplicate line remover. Changing one recurring phrase across a long block of text: find and replace. Reading a machine-generated file that arrived as one unbroken line: the JSON formatter. Tagging campaign links so your analytics can actually tell them apart: the UTM builder.

All of them run inside your browser — nothing you paste is uploaded anywhere. None of them are impressive. Together they remove more manual minutes from a normal week than most agent projects do in their first month.

The short version

Automation stopped being about connecting apps and became about systems that decide — 36 of 50 videos are about agents, six are about workflow tools. That makes it far more capable and much harder to predict, which is why 31 of the 50 are about limits rather than possibilities.

If you can write the steps down, write them down and use rules. Save agents for the cases where you genuinely cannot. Measure before you change anything, keep a person on the irreversible actions, and expect the last ten percent to take longer than the first ninety.

Every article behind this guide links to its original video, with the creator credited. The Automation tag has all 50.

---

This guide draws on 50 videos covered on Tech Feed Watch, from channels including Futurepedia, Y Combinator and WorldofAI. The counts quoted — 36 on agents, 31 on failure and limits, 6 on workflow tools — come from the archive itself, not from an industry report. Written and maintained by Jacob S. Olsen. If something here is wrong, the corrections policy explains how to tell me.

The numbers in this guide are a snapshot of the archive as of July 2026; the archive itself keeps growing.