Our blog

Practical notes on putting AI to work without losing control: who approves what an agent does, why we let AI write code instead of improvising, and what a voice agent should never say.

Three approval tiers rising with risk: run freely, run and report, and stop and ask, where an action waits for a person to approve it.01 · AI GOVERNANCERUN FREELYdraft · read · searchRUN & REPORTupdate · book · fileSTOP & ASKsend · pay · deleteAPPROVELOW RISKHIGH RISK →AGENTS MOVE FAST — PEOPLE OWN THE RISKY CALLS
AI GovernanceJul 2026

Who approves what your AI agent does?

A simple model for approval points: agents move fast on low-risk work and stop for a human on the rest.

AI first versus code first: improvised AI runs wander to different results, while code written once by AI takes the same path to the same result every run.02 · ENGINEERINGAI FIRSTCODE FIRSTAI WRITES IT ONCEdef run():erp.sync()audit.log()return okDIFFERENT EVERY RUNSAME RESULT · TESTED · LOGGED
EngineeringJul 2026

From AI first to code first

Why we now use AI to write the code that does the work, instead of letting a model improvise every run. Cheaper, faster, and easy to test and audit.

A voice agent's speech bubble: a promise, card details and legal advice are struck out by guardrails, and handing over to a colleague is allowed.03 · VOICE AITONO · LIVE CALLI promise you'll get a refund.NO PROMISESYour full card number is…NO CARD DATALegally, you should…NO LEGAL ADVICELet me connect you to a colleague.HAND OVER ✓GUARDRAILS GO IN BEFORE THE FIRST REAL CALL
Voice AIJun 2026

What a voice agent should never say

The guardrails we put on every production voice agent before it takes its first real call.