Every AI system we build follows one rule: the AI prepares, a person approves. This is how we design it, and why.
Drafts, not decisions
Language models are excellent at summarising an inquiry, drafting a reply or outlining a project. They can also be confidently wrong. So our agents produce drafts that are checked against a strict format, and they have no direct access to email, payments or your systems.
Three levels of automation
- Suggest only: the agent writes a draft and nothing else happens.
- Execute after approval: the action waits in an approval queue until you approve it.
- Limited automation: only low-risk internal actions, which you choose, run automatically.
Some actions always need a person: sending proposals, confirming prices or deadlines, contracts, refunds, spending money, publishing content and deleting records.
Numbers come from code
Scores, estimates and reports are calculated by ordinary code from real data. The AI writes the explanation around those numbers, and if something is unknown, the report says it is unknown.
We use this approach in our own operations dashboard, and we build it into every client system.