Why one agent is the wrong shape

Everybody asks for a single AI assistant that knows the whole business. That one would need every permission any job ever needed. Why a small team of narrow agents is safer, easier to fix, and the only version you can actually explain.

Almost everybody asks for the same thing, in almost the same words: an AI assistant for the business. One of them. Something you can ask anything, that knows how the place works and gets on with it.

It is a completely reasonable thing to want, and it is the wrong shape. Not because the technology cannot do it, but because of what you would have to let it do.

Start with the permissions, not the capability

Write down every job you would want this assistant to handle. Answer customer enquiries. Chase overdue invoices. Reorder stock when it runs low. Draft the social posts. Check whether a price is still right. Put a quote together.

Now write down what each job needs access to. Email that goes out in your name. The CRM, with write access. The accounting system. Supplier ordering. Your price list. Your marketing tools.

One assistant doing all of that needs all of that at once. It is a single account holding the union of every permission any job ever needed, available at every moment, including the moments it is doing something completely unrelated.

The question you suddenly cannot answer

Sooner or later somebody asks what the thing is allowed to do. It might be your accountant, your insurer, a customer who got an email they did not expect, or you at two in the morning.

With one assistant, the honest answer is a shrug and the word "most". With a team of narrow agents, the answer is a sentence each, and you can say them out loud without flinching. That difference is not paperwork. It is whether the thing gets used at all, because systems nobody can explain quietly stop being trusted and then quietly stop being used.

When it goes wrong, you want the damage contained

It will get something wrong. Everything does. The question is what that costs.

A purchasing agent that starts drafting nonsense orders is a bad afternoon. You switch it off, the other five keep working, and the enquiries still get answered while you sort it out. One assistant that starts behaving oddly is an outage across your whole business, and the only lever you have is off.

Logs are only useful if the jobs are separate

Everything an agent does should be recorded. But a single log of one assistant doing nine unrelated jobs is close to unreadable, and the thing you actually need to know, which is why a particular customer got a particular email, is buried in it.

Separate agents produce separate records. When somebody asks what happened with an account, you look in one place and it is short.

What a team actually looks like

Each one gets its own login in each system it needs, never a shared account and never a person's. You can see what it did, and you can switch one off without switching off the rest.

You do not start with six

This is the part people miss when the list above looks like a big programme. Nobody buys six agents. Most businesses start with the front door and one other, usually whichever job is currently eating somebody's week, and add a role when there is clearly a job for it.

The team is the shape, not the starting order. What matters is that the second agent is a second agent, with its own job and its own logins, rather than a widening of the first one's permissions.

The short version

One assistant that can do everything is one account that can do everything. Split the work by job, give each job only the access it needs, keep a person on anything that spends money or goes out in your name, and you end up with something you can explain, contain and switch off a piece at a time. It is also, as it happens, easier to build.