Your AI projects are stalling in review, and usually for the same reason. Someone asks what happens when an agent does something expensive and wrong, and the honest answer is that it happens, and afterwards nobody can say who allowed it.
What it changes for the business
- Agents get into production. The actions legal and risk will not sign off on unattended become actions a named person approves, so the project ships instead of waiting for a policy that never arrives.
- You can answer "who authorised this?" with a person, a device and a timestamp, rather than a service account and a shrug.
- The blast radius of a manipulated agent is capped. An agent that is talked into doing something it should not still cannot complete the actions you put behind a person.
- Only the expensive actions slow down. Approval is applied per destination, so routine traffic is untouched and the queue stays small enough that people actually answer it.
How it works, briefly
The gateway pauses the action and asks a person on their phone. They see what is actually being requested, in plain language, and allow or deny. The answer is signed on their device and tied to that specific action, so it cannot be reused for a different one. Nothing in your application changes.
What this is worth against the alternatives
| How this is handled elsewhere | How Squidder does it |
|---|---|
| Log the action and review it later | Ask before it happens, not after |
| Block it by policy, with no way to say yes | Allow the exception, with a named person on record |
| Approve a sign-in, then trust the session | Approve the action itself, each time it matters |
| Notify a shared channel and hope someone reads it | Reach one accountable person on their own device |
| Afterwards: "the service account did it" | Afterwards: this person approved this action |
No LLM gateway or AI firewall in this category offers any of this. It requires a companion app and a device identity, which is a product to build rather than a feature to add.
Where it fits, and where it does not
Approval is for actions that are rare and consequential. Put it on payments, production changes and data export, not on a summariser that runs a thousand times an hour. A machine with no person behind it cannot be asked, so unattended traffic needs a policy decision instead, which is a different control.
Where to start
Pick the one action your security review keeps stopping. Put approval in front of that destination only, leave everything else alone, and see whether the objection survives the demo.