Our AI agents write code, draft proposals, read our accounts, watch our servers and write social posts. Not one of them can send an email, publish a post or spend real money without a person saying yes. That is deliberate, and it is the first thing we set up for any business adopting AI.
The charter
Every agent works under the same short charter. Never invent a fact. Record honest limits. Check numbers before using them. Stage anything that faces a customer for approval. It sounds obvious, and it is the difference between an assistant and a liability.
How it plays out
- The accounting agent reads live figures from Xero and writes a daily brief. It can draft an invoice. It cannot approve or send one.
- The proposals agent writes quotes from a library of what we have actually built and charged before. Tim sends them.
- The social agent collects real facts from our work and drafts posts, and blocks filler phrases before Tim even sees them. It never publishes.
- The studio agent must ask before any job over US$50 or any day over US$100.
- The building agents each work in their own copy of the code, and changes reach live systems through one reviewed path.
Where it went wrong, and what changed
Being honest about this matters more than the successes. Two agents once picked up the same task and one overwrote the other's tests. A summary report once overstated a figure several times over before a check caught it. An agent killed processes it should not have touched on a shared server. Each incident became a rule, and some became hard technical blocks: the server now refuses that kind of command outright.
What this means for your business
AI is most useful when it can act, and most dangerous for the same reason. The answer is not to keep it away from your systems. It is to give each agent only the access it needs, keep a record of what it did, and keep a person on every decision that matters. That is how we set up AI for clients through our AI consulting service.