What happens when an AI system makes a mistake?

Every action that goes out on your behalf or moves money is approved by a person. That is not a setting you can switch on, it is how the route is built.

The judgement stays with you. What the system takes over is the volume: reading, looking up, copying over, queueing.

On most routes that means looking through a list of thirty lines once and clicking thirty times, rather than doing the same work over thirty times. That gate also makes a mistake cheap. Clicking away a wrong draft costs seconds, a wrong mail at a client costs a conversation and sometimes the client.

If the model is unsure about what it sees, it says so instead of returning a guess. The case then goes to the pile someone checks, with the reason attached.

What a model returns we treat as a proposal and not as a fact. Before the answer lands anywhere, the code checks whether it has the shape that was agreed: the fields that belong in it, an amount that is an amount, a date that exists. If the shape is wrong, the route stops.

A model that does not know sounds exactly like a model that does. Confidence in the tone is not a signal, so the doubt sits in the route on purpose.

Per run the log records what came in, which steps the system took and what went out. That is what an audit, an administrator and a sceptical colleague all three need.

Because the system carries its own identity, your own logs show a name next to every action: the system or a colleague.

Systems are built with fallbacks, so an outage degrades neatly instead of stopping your operation. An outage should be a message saying that today it has to be done by hand.

We keep the code running, offer ongoing monitoring and maintenance, and reply within one working day. We fix bugs without a bill.