The short answer
AI automation means getting software to do a repeating job for you, where at least one step involves an AI model making sense of messy input rather than following a rigid rule. Ordinary automation needs you to spell out every condition in advance. AI automation lets one step handle an email, a form or a scanned invoice that nobody could reasonably have written a rule for.
It is genuinely useful for sorting, routing, summarising, notifying and drafting. It is genuinely risky for anything where being confidently wrong is expensive. Most of the skill is knowing which of those two situations you are in.
So what is it, actually?
Strip away the branding and every automation, AI or otherwise, is three things: a trigger, some ordered steps, and a result that lands somewhere.
Ordinary automation has been around for decades and it is wonderfully dull. "When a form is submitted and the country field says Germany, add a row to the Germany sheet and notify Priya." It works perfectly, forever, as long as reality keeps arriving in the shape you predicted. The moment someone types "germany " with a trailing space, or writes "Deutschland", your rule quietly does nothing.
AI automation adds a step where you never had a rule to begin with. "Read this email and decide whether it is a complaint, a sales enquiry, a supplier invoice or a recruiter." You cannot write that as a keyword match, or rather you can, and it will be wrong in a hundred creative ways. A model reads it the way a slightly distracted colleague would, and returns an answer in a format the rest of your automation can use.
That is the whole trick. The rigid steps stay rigid, because rigid is reliable. One or two steps in the middle handle the mess.
In plain English
- Trigger:
- The thing that starts the automation. A new email, a form submission, a file appearing in a folder, or simply nine o'clock on a Monday.
- Step:
- One action in the sequence. Most steps are ordinary software doing ordinary things: move this, copy that, send a message.
- Model:
- The AI part. It reads messy input and returns an interpretation. It is the only step in your automation that can be confidently, fluently wrong.
- Human in the loop:
- A deliberate pause where a person approves, edits or rejects before the automation continues. The most important design decision you will make.
What does one look like, start to finish?
Imagine a shared enquiries inbox at a small business. Forty messages a day, half of them nonsense, and somebody spends the first hour of every morning triaging them. Here is a real-shaped automation for that.
TRIGGER A new email arrives in the shared inbox.
STEP 1 Plain rule, no AI. If the sender is on the newsletter
list, file it and stop. Cheap, exact, no reason to
involve a model.
STEP 2 Model step. Read the subject and body. Return:
- a category, from this exact list of five
- urgency: low, medium or high
- a two-line summary
- a suggested reply, polite, under 120 words
If the message does not clearly fit a category,
return UNSURE rather than picking the closest one.
STEP 3 Plain rule again. File into the folder that matches
the category. Anything marked UNSURE goes to a
review folder, not to a guess.
STEP 4 Post the summary and the draft reply into the team
chat channel, with a link back to the original email.
STEP 5 A human reads it, edits the draft if needed, and
clicks send. Nothing leaves the building without
this step.
STEP 6 Append a row to a log: date, sender, category the
model chose, category the human kept, whether the
draft was edited.
Notice how little of that is AI. One step out of six. The model is doing the single job a machine could not previously do, which is reading something written by a human being who did not know they were talking to software.
Step six is the one everybody skips and the one that saves you. If you find yourself overriding the category on a third of the messages, the automation is not working and the log is how you find out. Without it you would just have a vague feeling that the inbox is still annoying.
Notice also that step five is a person. The automation does not answer customers. It sorts, summarises and drafts, then hands you something that takes ten seconds to approve instead of five minutes to write. That is a genuinely good deal, and it does not require you to trust the model with anything irreversible.
What is it genuinely good at?
The pattern is consistent. AI automation earns its keep when a step requires judgement about language, the output is cheap for a human to check, and a mistake costs a small correction rather than an apology.
Routing and sorting. Which team should see this? Which folder does it belong in? Is this urgent?
Summarising. Turning a forty-message thread, a long form response or a meeting transcript into something you can read in fifteen seconds.
Notifying with context. Not "a new ticket arrived", which you will learn to ignore by Thursday, but "a new ticket arrived, here is what it says and why it might matter."
Drafting. First versions of replies, descriptions, summaries and updates. Never final versions.
Extracting from documents. Pulling the date, total and supplier name out of an invoice that arrives as a photograph. This used to require either a person or an expensive, brittle template.
If you want to build one of these end to end without writing code, that is exactly what the no-code automation course covers, and AI That Does the Work goes further into designing the steps themselves.
What is it bad at?
Anything with legal or financial consequence. Approving an invoice for payment, changing a customer's contract, issuing a refund, sending anything that could be read as advice. The model has no idea what a mistake costs, and it will make one in the same calm, competent tone it uses for everything else.
Anything where a mistake is invisible. If a wrongly categorised enquiry goes into a folder nobody opens, the error does not announce itself. It just accumulates.
Anything that needs accountability. When somebody asks who decided this, "the automation did" is not an answer that survives contact with a regulator, a customer or your own manager.
Anything requiring exact recall or arithmetic. Models are language machines. They are excellent at reading a number and hopeless at guaranteeing they carried it correctly. Let ordinary software do the sums.
โ Weak prompt
Prompt
Read this email and tell me what it is about.
Output
A friendly paragraph describing the email, in slightly different wording every single time.
Your next step needs a category it can act on, and you have given it prose. Anything downstream that tries to match on the text will break the first week.
โ Good prompt
Prompt
Categorise this email as exactly one of: complaint, sales enquiry, invoice, recruiter, other. If it does not clearly fit one, answer UNSURE. Then give urgency as low, medium or high, and a two-line summary. Do not add anything else.
Output
complaint / high / Customer says their order arrived damaged for the second time and is asking for a refund.
A fixed list, a permitted way to say I do not know, and no room for improvisation. Now the rest of the automation has something it can reliably act on.
What is silent failure, and why does it matter most?
An automation that breaks loudly is not a problem. Something errors, you get an alert, you fix it, you carry on mildly irritated.
The dangerous failure is the quiet one. Your supplier changes their invoice layout, or a new kind of enquiry starts arriving, or someone renames a folder. The automation does not stop. It keeps running happily on input it no longer understands, producing confident, plausible, wrong output, week after week. Nothing turns red. You find out in month four when somebody asks why an entire category of enquiries went unanswered since spring.
Silent failure is the real risk in AI automation, not dramatic breakage. Build in two defences from day one. First, make it announce what it did: a daily or weekly digest listing what it processed and how it categorised things, sent somewhere a human will actually look. Second, sample the output every week. Pick five items at random, check them properly, and write down what you found. Ten minutes a week is the entire cost of not being blindsided.
Our AI safety and verification course is largely about building that habit, and it applies to anything you leave running unattended.
What does it cost?
Speaking generally, because pricing changes constantly and anything specific here would be out of date by the time you read it: most automation platforms and most AI models charge by usage. You pay for how many times things run and how much text the model reads and writes, sometimes on top of a monthly subscription.
The practical consequence is that cost scales with volume, not with how many automations you have built. Ten automations that each run twice a day cost very little. One automation on a busy trigger can cost a surprising amount, because a loop that fires on every change to a spreadsheet will happily fire several thousand times while you are at lunch.
So: test on a handful of items before you connect the real trigger, set a spending cap if the platform offers one, and check the usage figures after the first week rather than after the first bill. Then check them again after you change anything.
Where does a human have to stay in the loop?
Three places, and they are not negotiable.
Before anything leaves your organisation. Emails, replies, published content, anything a customer sees.
Before anything irreversible. Money moving, records being deleted, statuses that cannot be set back.
Before anything becomes the official version. A summary that gets filed as the minutes, a category that drives a report, a figure that ends up in a ledger.
Everywhere else, let it run. Sorting, drafting, notifying and preparing are exactly the jobs you want automated, and the ten seconds you spend clicking approve is what keeps the whole thing honest.
If you are ready to build one, start with the no-code automation course. If you want to understand what happens when these things start making their own decisions about which steps to take, Intro to AI Agents is the natural next step. And if you are not sure where to begin at all, start here and we will point you at the right one.