What is AI automation? A plain-English guide

AI automation is software that carries out a repeating job on its own, and uses artificial intelligence for the parts of that job which need judgement rather than a rule. The job runs the same way every time. Something starts it, then a few steps happen in order, and at the steps where a rule would not be enough, the software works out what to do instead.

That is the whole idea. The rest of this is what it looks like in practice, where it goes wrong, and how to tell whether a particular job in your business is worth handing over.

Automation, before the AI part

Ordinary automation has been around for decades and it is completely literal. You write a rule that says when this happens, do exactly that, and it follows that rule perfectly, forever, including the times when following it is obviously the wrong thing to do.

That works for jobs with no judgement in them. Copy the form submission into the spreadsheet. Post the alert into the channel. Create the folder when the project starts. Nothing in those needs an opinion, so nothing is lost by handling them literally.

It falls apart the moment a step needs a decision you cannot write down. Write a follow-up that sounds like the person who normally sends it. Work out whether this email is a complaint, a question or a nudge. Pull the figure out of a supplier invoice when every supplier lays theirs out differently. You can feel the rule you would have to write, and you can feel that it would not survive contact with the second example.

What the AI part actually does

This is the bit most explanations skip, and it is the only bit that matters.

AI does not take over the automation. The shape is the same as it always was: something starts it, some steps happen in order, it stops. What changes is that one or two of those steps can now be a judgement call instead of a rule.

Usually one of four:

• Writing something. A reply, a summary, a first draft of a report, in a particular voice and about this specific client, rather than a template with the name swapped in.
• Sorting something. Deciding which of a handful of categories a message, a lead or a request belongs to, when the wording varies every time.
• Pulling something out. Getting the total, the date and the reference off a document, where every sender formats it differently.
• Deciding whether. Whether this is worth escalating, whether that counts as a reply, whether this one is different enough to need a person.

Everything around those steps is still ordinary, boring, literal automation, and that is a good thing. The reliable parts stay reliable. The AI is used where a rule was never going to work, and nowhere else.

What it looks like in practice

Three examples, none of which need any particular software.

Chasing a quote. When a proposal has been out for three working days with no reply, draft a follow-up that refers to what was actually quoted and sounds like the person whose name is on it, and send it, unless the client has already replied or the deal has moved on. The chasing is the literal part. Writing something that does not read as a mail merge is the judgement part.

Sorting the inbox everybody shares. When something arrives in the general enquiries address, work out whether it is a new enquiry, an existing client with a problem, a supplier, or nothing, and route it accordingly. The routing is trivial. Deciding which of those four a two-line email is has never been expressible as a rule.

The monthly report nobody wants to build. Collect the numbers from wherever they live, put them in the same layout as last month, and write the two paragraphs at the top explaining what changed. Collecting is mechanical. The two paragraphs are the reason it still takes somebody half a day.

Notice the pattern. In all three, most of the work was always automatable and the whole job was blocked on one step in the middle that was not. That single step is what has changed, and it is why jobs that were not worth automating in 2020 are worth automating now.

What AI automation is not

Three things get called the same name and behave completely differently.

It is not a chatbot. A chatbot answers whoever is in front of it and stops when they go away. An automation runs whether or not anybody is there, and it acts on your systems rather than on a conversation. Some things are both, such as a support bot that also opens the ticket and emails the client, but the two halves fail in different ways and are worth thinking about separately.

It is not an AI agent. An agent is something you hand a task to now and watch. You ask, it works, you look at what came back. An automation runs in the background on things you are not watching, possibly for months. That difference sounds academic and is not: nobody is reading an automation’s output the way they read an agent’s, so an automation that has started doing the wrong thing can carry on doing it for a very long time. The distinction is worth more space than it usually gets.

It is not the same as putting AI in your product. This is about the running of the business, the admin and the chasing and the reporting, rather than about what you sell.

What goes wrong

Every guide to this stops at the definition and the benefits. The failures are more useful, and there are three, in rising order of cost.

1. It breaks

The systems an automation sits on top of change without telling anybody. A field is renamed, a permission is tightened, an app changes how it returns something. The automation stops.

This is the failure everybody worries about and it is the cheap one, because you find out. Something errors, somebody notices, it gets fixed. Annoying, recoverable, and the good outcome. That last part is worth internalising.

2. It keeps working and is quietly wrong

The expensive one. The automation runs perfectly and does the wrong thing, and because every run succeeded, nothing alerts anybody.

A rule that was slightly too broad chases clients who already paid. A step that reads the wrong field puts last month’s figure in this month’s report. A condition that was right for the pipeline you had in March is wrong for the one you have in June, and nobody changed the automation because nobody thought to. It runs green for six weeks and you find out from a client.

This is the risk that scales with how good the tools get. When building something took a fortnight, you understood it. When it takes a sentence, you have a great deal more of it and considerably less understanding of any of it. Ask any tool you are evaluating what it does about this specific problem, and treat a vague answer as an answer.

3. It makes a bad process faster

The one nobody counts as a failure. If the report nobody reads takes half a day to build, automating it means nobody reads it instantly and for free. The half day comes back and the underlying waste is now permanent, because it has stopped being annoying enough for anyone to question.

Automation makes a process cheaper to run, which makes it much harder to kill. Before handing a job over, it is worth asking whether the job should exist.

How to tell whether a job is worth automating

A job is a good candidate when most of these are true:

• It repeats on something you can name. Every Tuesday, every new client, every time a deal reaches a certain stage. If you cannot say what starts it, it is not ready.
• It has the same shape every time. The details differ, the sequence does not.
• You could describe it to a new starter in a few sentences. If it takes twenty minutes and three exceptions, those exceptions are the job and they are where it will go wrong.
• Getting one wrong is recoverable. An awkward email you can apologise for, not a payment you cannot take back.
• Nobody is paying you for that part. Clients pay for judgement and creative work. They do not pay for chasing, formatting or copying between systems, which is exactly what should go first.

And the strongest signal of all: several people independently complain about the same thing. One person saying a job is painful is a preference. Most of a team saying it is a cost you can count.

Where to start

Not with the tool. Nearly everybody starts by opening one, looking at the empty editor, and building whichever job they happened to think of first, which is very rarely the one costing the most.

Start instead by finding out where the hours actually go.

Ask the people doing the work, not just whoever runs the place. Ask about last week specifically rather than in general, because people are poor at generalising and good at remembering. Ask separately, so the second person is not agreeing with the first. Then count. The job to automate first is the one several people named without being prompted, that satisfies the test above.

Then do one. Watch it for a month before you build the next. The agencies that end up with forty half-working automations all got there by building the second one before they understood the first.

Where we come into it

Everything above is true whichever tool you use, including a spreadsheet and a reminder. This part is about ours, and it is the only part that is.

Nrth Star is built for marketing agencies and it is organised around the two problems in this post that tools normally leave with you. It starts with the audit rather than an empty editor, so the first thing you get is a ranked list of where your hours go rather than a blank page. And before anything runs, it replays the automation against your last 30 days with everything switched off and shows you what it would have done, client by client, which is aimed squarely at failure number two.

If you already know what you want to build and are happy maintaining it, a general-purpose builder is the better answer and there are several good ones. The audit is free either way, and it will tell you where your hours are going even if you go and automate them somewhere else.