Who gets to build live tools, and who builds concepts only
Everyone can build a concept. A few named people can build something that reads live data. Fewer still can build something that writes to a system or sends anything on its own. The line isn't seniority or skill – it's what the thing can do without a person watching.

The question that decides the tier
Compliance and data integrity aren't threatened by people building things. They're threatened by things that act unattended – an agent that updates a record, sends an email, places an order or changes a price while nobody's looking. So the useful question about any build isn't "who made it?" but "what can it do on its own?". Answer that and the tier follows.
| Tier | What it may do | Who may build it | What it needs |
|---|---|---|---|
| Concept | Work with sample data, public data or the builder's own files. Nothing live. | Anyone | Nothing. Build freely. |
| Assist | Read live data and show a person something: a draft, a summary, a ranked list. The person acts on it. | A named group of 2 to 5 | A read-only connection granted by the systems owner. A named owner. |
| Act | Write to a system, send a message, or run on a schedule without a person in the loop for each case. | Builders in the named group, after a review | A review by the AI lead and the systems owner, a named owner, a log, a stop point, a sample check. |
Where both extremes end up
Letting everyone build Act-tier tools is how you end up with an agent emailing customers from a mailbox nobody checks, built by someone who left in March. Letting only IT build anything is how you end up with a 6-week queue for a summariser, and a business that goes back to doing it by hand – or worse, does it with personal accounts the company can't see.
In our experience 2 to 5 named builders in a company of up to a couple of hundred people is about right. Enough that each department has someone to go to, few enough that the systems owner knows every connection by name.
The three controls that actually protect you
The tiers only work with three plain controls behind them:
- Connections get granted, not copied. Nobody pastes an API key into a tool. Access to the CRM, the mailboxes or the accounts system is granted by the systems owner to a named tool, read or write, and can be taken away in one place. What AI should be allowed to see covers the data side.
- Anything that writes has an owner and a log. If it changed a record or sent a message, there's a line saying what, when and why, and a person whose name is on it.
- Anything that reaches a customer gets sampled. A person reads a share of what went out, weekly, for as long as it takes to trust it.
Finding the agents already running
Most businesses asking this question already have unattended automations they don't know about – a scheduled export someone built, an assistant with a mailbox connection, a form that triggers a message. Before you write the policy, make the list. Ask every department head: what runs without anyone pressing a button, and who built it? Put each one in a tier and give it an owner, or switch it off. The list is usually longer than anyone expected, and we think it's the single most useful afternoon in this whole subject.