1Password’s agent launch puts permission at the center of the AI business case.

A demonstration can make an AI assistant look like an employee. The business owner still has to decide what it is allowed to do on Monday morning.

1Password’s October 22 announcement addresses one part of that decision. The company describes a way for an agent to request a credential, receive the user’s approval, and sign in without seeing or storing the credential itself. Its launch post distinguishes integrations rolling out from others still coming. These are the company’s descriptions. This column is not an independent test of the platform or a recommendation to buy it.

The interesting commercial idea is the separation of useful access from permanent possession of a secret. A business ought to be able to delegate a task without treating its entire account as the price of admission. That principle could make automation more practical for the owner who does not have an internal security department.

Consider a hypothetical catering company. An agent might be useful for organizing incoming event requests and preparing draft proposals. Permission to read inquiries would not logically require permission to change the bank account receiving customer deposits. The valuable product would help the owner distinguish those jobs, rather than require trust in one giant bundle.

That distinction also changes how I would judge the economics. The right comparison is not the subscription price against an imaginary employee who works for free. It is the total cost of completing a defined task before and after delegation. The owner’s time approving access, reviewing output, and correcting mistakes belongs in that calculation. This is a proposed way to evaluate a tool, not a claim that a particular percentage of labor will disappear.

There is a trap on the other side. A system that interrupts a person for every harmless step can become an expensive way to do the same work twice. Excessive friction is a real business cost. The opportunity for developers is to make the boundary intelligible: approvals where authority changes, useful visibility while work proceeds, and an understandable way to stop.

Protecting the password does not settle the question of behavior after login. An agent could be authorized to enter an account and still misunderstand an instruction. The company announcement should therefore be read as a statement about credential handling, not proof that every possible task is safe or every action reversible.

For an entrepreneur, that is a reason to narrow the experiment rather than abandon it. Choose a task with a clear finish, limited authority, and an output a person can assess. Then decide whether the finished work justifies the full cost. There is no need to declare an entire company automated before one recurring job works well.

The strongest AI sales pitch would make that judgment easier. Tell the customer what access is required, what remains outside the assignment, and which integrations can actually be used now. A tool earns a place in the business when its usefulness can be inspected. Permission is part of the product, and a credible business case should price it in.