The banks are right about AI agents. Most businesses have the same login problem.

Four South African banks have drawn a line on AI agents. FNB, Discovery Bank and Capitec told TechCentral they do not allow third-party AI agents into customer accounts, and Absa warned that customers who hand their credentials to one may be held liable for any losses. I agree with the banks, and I think most business owners should read that story as a warning about their own systems first.

Why sharing a login with an AI agent is the real risk

Sharing a password with an AI agent makes the agent indistinguishable from you. Absa's Jochem Melis said it plainly: "The primary concern is not technology itself, but the sharing of credentials." An agent with your banking login signs in as you. To the bank, the payment looks like yours, and Absa says a customer who shared the login may carry the loss.

Now look at how a typical SME runs. The owner, the bookkeeper and whoever covers in December all use one login for the accounting package. The supplier portal password lives on a sticky note. And now an AI tool gets someone's email or CRM password so it can draft replies. That is the exact pattern the banks refuse. With one shared login, you cannot see who did what, and you cannot limit what a person or a tool is able to touch.

A Visa study found that only 23% of South African consumers would trust an AI agent to complete a purchase on their behalf. I think that caution is healthy.

What role based access control looks like in a real business system

Role based access control means every person or tool gets its own identity, and the system decides on each action what that identity is allowed to do. FNB describes the same idea for the framework it is building, with "explicit consent for every action" and every interaction "monitored and recorded".

We scoped a platform for a freight brokerage that coordinates road and rail providers, and the proposal was built on this principle from the first phase. Cargo owners, providers and the brokerage's own team would each get a separate portal and login. The user's organisation and role would be checked on every request, never inferred from what the screen shows. A provider would see only its own leg of a job and its own rates. Every status change would be written to a history with the old status, the new one, the time and the user who made it. No provider could be allocated work until someone at the brokerage had approved them by hand. There was no AI in that design. It is still the shape you want in place before any AI tool touches a system, because a tool with its own login and a narrow role can be traced, and switched off, without locking anyone else out. It is the same reason I argue for keeping the approval step for AI output inside your own business.

Questions to ask before you connect a tool or a new hire

Before an AI tool or a new staff member gets access to any system you run, ask three questions. Does it get its own login, or is it using mine? Can I limit it to only what it needs? Is every action recorded with the name of whoever did it?

If any answer is no, fix that before adding the tool. Start with the software you already pay for, and check whether it supports separate users and roles before you buy anything new. Where your process does not fit a package, permissions and an audit trail are the part of custom software development worth paying for properly.

The banks are not against AI. FNB is building a way for approved assistants to connect, and sums up its position as "no autonomy without control". Hold your own systems to that before the first AI tool asks for your password.