Building an AI feature you can switch off

One of the first questions we ask when designing an AI feature is not how clever it can be. It is what happens when we turn it off. If the honest answer is that the business grinds to a halt, or that nobody actually knows how to run the work without it any more, we have built something fragile, however impressive it looks on a good day.

This sounds cautious, almost pessimistic, coming from a team that builds these features for a living. It is not. It is the opposite. Designing for the off switch is what lets a business say yes to AI in the first place, because it removes the quiet fear underneath every adoption decision: what if this thing does something strange and we cannot stop it.

So we treat the off switch as a real feature, planned from the start, not a hypothetical we wave away. In practice that means a few things. The AI sits on top of a workflow that still makes sense without it. The model does a job a person could step back into, perhaps more slowly, but without the wheels coming off. When you flip the switch, the work routes to a human queue, or falls back to the simpler rule it was speeding up, and the operation keeps running. Degraded, maybe. Stopped, never.

Why this matters more with AI

You could argue this is just good engineering, and you would be right, but it matters more with AI than with ordinary software. AI behaviour can shift in ways traditional code does not. A model gets updated. The kind of input changes in a way nobody anticipated. Something that worked beautifully for months starts producing subtly worse results, and because the output is fluent and confident, it can take a while to notice. When that happens, you want a calm, obvious way to pull the AI out of the loop while you investigate, without taking the whole operation down with it.

The off switch also changes the conversation when we are deciding whether to build at all. If a feature can be cleanly switched off, we can be braver about trying it, because the downside is contained. We can turn it on for a slice of the work, watch how it behaves on real cases, and expand only once it has earned trust. If it cannot be switched off, every decision becomes heavier and slower, because the stakes of being wrong are so much higher.

There is a cultural payoff too. A team that knows it can turn the AI off tends to trust it more and use it better. The switch they rarely touch is precisely what makes them comfortable leaning on the feature the rest of the time. Control and confidence turn out to be the same thing.

So when we hand over an AI feature, we hand over the off switch with it, and we make sure someone knows how to use it. A feature you can stop is a feature you can actually trust.