The case for human-in-the-loop AI
Autonomy is exciting until an agent sends the wrong email. Approval gates are what make AI safe to actually use at work.
There's a race to make AI agents fully autonomous. For real work, that's the wrong finish line.
Reads are safe. Writes are not.
Reading your calendar, searching the web, pulling CRM records — these can run freely. But anything that sends, writes, or changes state carries risk: a wrong recipient, a bad number, an irreversible action.
The approval gate
The safe pattern is simple: the agent proposes the write, you approve it, then it runs. If a job proposes several actions, you approve or skip each one. With Medlar's default policy, nothing leaves your workspace without an explicit click — and even when you let ordinary actions run on their own in apps you chose, high-risk ones still stop for you.
Why this wins
- Trust. You can hand over more work because you control what actually happens.
- Speed with a brake. You get the agent's velocity without the runaway risk.
- Accountability. Paired with a full log of every step, you always know what was done and why.
Full autonomy sounds impressive in a demo. Human-in-the-loop is what ships to production.