I think that in the coming years a product’s value will be determined not by how nicely it answers, but by whether you can safely entrust it with an action.
Asking a model for an email idea is convenient. Entrusting it with sorting the inbox, updating the CRM and preparing invoices is a different class of product. Here you need not only intelligence but clear control: what exactly was done, on what data, what can be undone and where the agent stopped.
The argument against: people don’t want to dig into logs and settings. And that’s true. But control doesn’t have to be complicated. Like in a banking app: you don’t study the payment system, but you see the amount and the recipient and can confirm the transfer.
The practical conclusion follows. If you’re building a product with AI, design more than the prompt and the chat. Think through action previews, permission boundaries, decision history and undoing changes from the start.
If you use AI, choose tools that not only “do it for you” but also let you check the result. That’s how trust in agents will be built.