I want a way back when AI gets it wrong.
I like being able to hand a task over to AI. I also want to understand what it has done & have a way to correct it. That becomes especially important when it can change something other people rely on.
Imagine using an AI assistant to update a customer database after a meeting. You add your notes, it finds the account & suggests the changes. I can see why that would be useful. Keeping records up to date takes time, especially when the information is spread across calls & notes. But imagine the notes mention two companies with similar names. The assistant picks the wrong one & adds a pricing discussion to an unrelated account. I’d want to know how easily I could notice that mistake. & if I only found it tomorrow, what I could still do about it. Before saving, I’d want to see the account name, the current information & what the assistant proposes to change. Having the relevant part of the meeting notes beside the suggestion would help me check it. If one detail is wrong, I’d want to edit or reject that part while keeping the useful changes. When two accounts could be a match, being asked to choose would make sense to me. A confidence percentage alone wouldn’t tell me which company the conversation was actually about.
The amount of checking I’d want depends on what happens next. A draft note gives me room to read & change it. Updating a deal value affects information someone else might use. Sending a message to a customer goes further again. I’d want to check the recipient & the wording before it leaves. Working through these differences early would help me decide where a quick interaction is enough & where the interface needs to give someone more time.
I want to know what changed & what I can still undo.
“Undo” is a reassuring word when it does what I expect. In this example, it might mean restoring the previous value in a field. A message someone has already received is harder to take back. Stopping an assistant halfway through a task also leaves a question: what has it already done? I’d want the interface to show me which changes were saved & which are still waiting. If I need someone’s help to fix the result, I’d want to know how to reach them. There’s another detail I’d want to work through with an engineer. What happens if a teammate edits the same account after the AI update? Restoring an earlier version could remove their work, too. The product might need to show both changes & let us decide what to keep. That’s something I’d want to understand while designing the recovery flow.
To try the design, I’d start with the situations that make the task less straightforward: similar account names, contradictory notes, a connection that drops halfway through saving. I’d run them more than once & ask someone unfamiliar with the feature to use it. I’d watch whether they could spot a mistake, understand what had happened & find their way through the correction. I’d also include that checking & correction time when judging how useful the feature is. Generating the update quickly is one part of the task. I still need to leave the account in a state I can rely on. Being able to review a change, put something right & carry on would make me much more comfortable using the assistant again.
