Every few years a technology arrives that promises to do everything for everyone, and for a while it is applied to everything, whether or not it fits. Automation and artificial intelligence are in that phase now. The useful question is not whether to use them, but where they earn their place.
A filter for what to automate
Before building anything automatic, we find it helpful to ask four plain questions.
- Is the task repetitive and well understood? Automation is excellent at doing the same thing reliably. It is poor at doing something nobody has fully defined.
- What happens when it is wrong? If a mistake is cheap and visible, automate with confidence. If it is expensive or silent, keep a person in the loop.
- Will people still understand it? Automation that nobody can explain becomes a hazard the day it misbehaves.
- Does it free people for better work? The best automation removes drudgery. The worst removes judgement.
Where AI fits, and where it does not
Intelligent tools are most useful where inputs are messy and a good-enough answer is valuable: summarising, sorting, drafting, finding patterns in large amounts of text or data. They are least useful where precision and accountability matter more than speed, such as anything touching money, legal commitments or someone's safety, unless they are clearly supervised.
A sensible pattern is assist, then verify. Let the tool do the first pass, and design the product so that a person can review, correct and override it easily. That keeps the benefits and bounds the risks.
Resist the demonstration
It is easy to build an impressive demo. It is much harder to build something that remains useful in daily work, behaves predictably, and degrades gracefully when it is wrong. Judge any intelligent feature by its worst day, not its best.
Technology should make a product calmer to use. If adding automation makes a product louder, more confusing or less trustworthy, it has not earned its place, however modern it looks.