AI should ask before it acts
The AI features people keep using are the ones they can take back. Trust is a design decision.
Every few weeks a product announces that its AI will now handle something for you. Your calendar, your inbox, your groceries. The demo is always impressive, and a month later the forums fill with the same complaint: it moved my meeting, it replied to my client, it ordered the wrong thing.
The feature did exactly what it was designed to do, and that was the trouble.
What makes people trust an AI feature?
The same thing that makes them trust a new colleague. It shows its work, it checks before doing anything that matters, and when it gets something wrong the mistake is small and easy to undo.
Software has had a version of this for decades. Undo is the most reassuring button ever made, because it makes trying something free. An AI feature without a clear way back asks people to bet on it every time, and most people decline the bet.
Where should the line sit?
I use a simple rule in everything I build now. The AI may suggest anything. Nothing is created, moved, finished or deleted until a person taps once to agree, and one more tap reverses it.
In Manekatti, AI reads a plumber’s quote and points out that the taps are missing. It changes nothing in the budget. The owner decides what to do with the warning. In the to-do app I am building, AI notices that a call has happened and asks whether the task can close. The answer is one tap either way.
People move faster with it, because they stop double-checking. A suggestion they can wave away costs them a second. An action they have to audit costs them their confidence in the whole product.
Does that limit what AI can do?
It limits what AI does unattended, which is the part most worth limiting. Everything else gets better. The model can be bolder in what it proposes, because a wrong proposal is harmless. The person stays the owner of their own day, their own money, their own client.
The companies that win with AI over the next few years will be the ones whose customers feel more in control after the feature ships. That is a design brief, and a very specific one. Before any AI feature goes in, I ask what the person sees when it is wrong, and how quickly they can make it right.