One AI Agent per Department: Mirroring the Org Chart

One practical way to give a busy owner an AI assistant is to mirror the org chart. Create one agent per department, give each only the systems that department uses, name it after the person who runs that department, and put one top-level agent above them that you talk to. Each department agent stays focused, its access stays narrow, and you get one place to ask "how are we doing on X?"

Why split into several agents instead of one?

One agent connected to everything gets a huge context and broad permissions. Splitting by department gives you three things:

• Narrow access. The finance agent sees accounting, not your inbox.
• Clear questions. "Ask the expansion agent" is easier to reason about than one assistant that knows everything.
• A familiar mental model. Naming each agent after the real head of finance or marketing makes it obvious what it should know and what it shouldn't.

What should each agent connect to?

Take a franchise business opening new branches as an example:

• Finance agent: the accounting system plus monthly finance reports.
• Expansion agent: the tracker where construction and opening progress of every new branch lives.
• Marketing agent: campaign data, plus an aggregation of Google reviews.
• Main agent: the one the owner talks to, by text or by voice from a phone. It asks the department agents and combines their answers.

Several cloud agent products can run this today without you hosting anything, and some support agents that talk to each other.

What does it actually do day to day?

Scheduled reports. For example, every two days: the status of each branch in the opening pipeline. Nobody has to ask, and nothing falls through the cracks when eight openings run in parallel.

Proactive alerts. It watches the mailbox and flags what matters, such as a franchisee giving notice or a key supplier changing terms.

Voice inbox triage. Over morning coffee: unsubscribe from noisy tool notifications, mark as read, delete, dictate a reply. An assistant can run the same routine on the owner's mailbox.

Meeting prep. Before a meeting with a prospective partner, it summarises who is coming and researches their company. With enough history it also learns your unwritten rules. For instance, it can notice that you always require a signed NDA before sharing numbers, and check that the signed copy has arrived.

How do you build trust with an agent like this?

Earlier agents forgot what they'd learned within a day, so skepticism is reasonable. Build trust in a low-risk order:

1. Start where mistakes are cheap. Email triage: a wrongly archived newsletter costs nothing.
2. Watch the error rate for a couple of weeks before giving it more.
3. Test it with sloppy instructions on purpose, to see whether it still does the right thing.
4. Only then connect business systems, one department at a time.

And a pragmatic argument for doing it at all: if you don't connect company data to AI deliberately, your colleagues will do it anyway, without rules.

What should you be careful about?

• Read access first, write access later. Let department agents report before they act.
• Know where the data goes. A cloud agent sends your mail and accounting data to a vendor. Check the data processing terms before connecting finance.
• Limits and cost. Entry plans are cheap, around $20 a month, but heavy daily use of any agent product hits weekly caps eventually.

Where should you start?

Pick the one department whose status you ask about most often, and give an agent read-only access to that one system plus a scheduled summary. If you check it every time it arrives, add the next department.

If you want help wiring agents to systems like accounting or Jira with sensible permissions, that's work I do.