How you stay in control

Every action an AI tool takes through Keystile is checked before it happens. Here is what that gives you.

Approvals {#approvals}

Anything that matters waits for a person: sending to clients, moving money, opening sensitive files. You see what the AI tool wants to do and why it was held, and you approve exactly that one action — or reject it. AI tools can never approve their own.

Your rules, in plain English {#rules}

Write rules the way you’d say them — “Payroll files must not be shared by AI tools” — and Keystile Intelligence applies them to every action by what it means, not by keywords. Each rule either stops an action or holds it for a person. Changes to rules need a second person’s approval.

Limits and access {#limits}

Firm limits that give the same answer every time — payment caps, who messages may go to, actions that always need a person — and each AI tool gets only the apps and actions its job needs.

Budgets {#budgets}

Monthly and yearly money limits for paid tools such as HeyGen, ElevenLabs or Claude, checked before every call so they never go over. Your AI tool can propose a budget from your plans; it applies only when a person approves it. Spend is recorded only for apps you put a budget on.

Tasks {#tasks}

Give an AI tool one job — a goal, what it may use, and for how long. Anything outside the job waits for you, even when the action itself is allowed.

Three levels of protection {#protection-levels}

Keystile Core	Firm limits, permissions, approvals, budgets and sensitive-data checks — the same answer every time
Keystile Insight	Keystile Intelligence — what each action means against your rules, unusual behaviour, hidden instructions
Keystile Fortress	Both together. The highest protection, and the default

What comes back is checked too {#results}

What apps return to an AI tool is screened as well: hidden instructions are caught, and client identifiers are masked before the AI tool sees them.

A record you can hand over {#activity}

Every decision — who asked, what, and why it was allowed, held or stopped — goes into a signed record. Any later change to it is detectable, and you can export it for an auditor or examiner.

Measure before you switch it on {#measure}

Run Keystile on a few weeks of your past agent activity, or beside one workflow without blocking anything, and see exactly what it would have done before you turn it on.

---

Want to see the details? Every one of these is shown live, with your own use cases, in a demo.