Start with what your software already does #
Every business already runs automation and mostly does not call it that. The accounting package raises a recurring invoice on the first of the month. The booking system sends a reminder the day before. The card machine reconciles overnight. None of that needed a model, because each one is a rule: when this happens, do that.
Rules are excellent. They are predictable, cheap, and when they break they break loudly. If a job can be written as a rule, it should be, and a consultancy that sells you a model for it is selling you the expensive version of something you could already have had.
What a model adds, and it is one thing #
The thing rules have never been able to do is deal with input that arrives in no fixed shape. A supplier who emails "can you push Thursday's drop to Monday" has not filled in a form. A customer who leaves a voicemail has not selected from a dropdown. A delivery note photographed on a phone in a van is not a spreadsheet. For decades the only way to get that into a system was a person retyping it.
A language model can read those things and produce structured output from them: a date, a reference number, a category, a draft reply. That is the addition. It is narrow, and it is genuinely new, and almost every useful business automation built since is some arrangement of that one capability wired into the rules you already had.
Which is why the outputs are guesses #
Because the model is inferring rather than looking something up, it can be wrong in a way a rule cannot. The Information Commissioner's Office, in its guidance on AI and data protection, is unusually blunt about this: it advises organisations to treat AI outputs as statistically informed guesses rather than facts.
It also notes that an AI system does not need to be one hundred per cent statistically accurate to comply with the accuracy principle in data protection law.
That is not a reason to avoid the technology. It is a design constraint, and it is the reason a sensible build puts the model where being wrong is cheap and recoverable.
Three kinds of work in your week #
It helps to stop sorting work by department and start sorting it by what it actually requires.
| Kind of work | What it looks like | What it needs |
|---|---|---|
| Rule shaped | Same trigger, same steps, same output, every time | Ordinary automation. No model involved |
| Messy in, rule shaped out | Someone reads an email, a form or a document, then does the same thing with it | This is where a model earns its place |
| Judgement shaped | The answer depends on context, relationship or a decision somebody owns | A person. Automation can prepare it, not make it |
Most jobs that feel automatable are in the middle row, and most jobs that feel hopeless are in the middle row too, because the person doing them describes the whole thing as one task. Splitting it is the useful move, and it is the subject of how to tell which jobs are worth automating.
The word AI is doing too much work #
"AI" now covers a chatbot on a website, a model summarising a call, a spam filter, a forecasting tool in a spreadsheet and a piece of software that books appointments. They have almost nothing in common operationally. In practice, three questions tell you more about a proposed automation than the label ever will:
-
What does it read?
Your systems, your inbox, your documents, a customer typing at it.
-
What is it allowed to do?
Draft, or send. Suggest, or spend. Flag, or delete.
-
Who sees it before the outside world does?
Somebody, or nobody.
Any supplier should answer all three in a sentence each. If the answers are vague, the build is vague. There is a fuller list in the questions to ask an AI supplier.
What it does not mean #
It does not mean replacing a system that works. The most common pattern in a business is not a new platform, it is a small piece of connective tissue between things that already exist, so that the same information stops being typed in twice. That is unglamorous and it is where most of the recoverable hours actually sit.
It also does not mean the whole job goes away. A job that took forty minutes and now takes six still takes six, and somebody still owns it. Counting the difference honestly is the only way to know whether any of this is worth doing, which is why our cost of admin time calculator counts the hours you spend now and refuses to predict the hours you might save.
Where to go from here #
If you are trying to work out whether there is anything here for your business, do not start with the technology. Start with the list of things your business repeats every week, sorted by how many hours they eat. If that list is short, or the hours are small, the honest answer is that you have nothing worth automating yet, and no amount of good software changes that.