Before you try it.
The five things teams ask last — the ones that decide whether this is worth an afternoon.
Why doesn't approving every AI agent action work?
Because at volume, approving stops being a decision. Anthropic measured that Claude Code users approve 93% of permission prompts, and shipped classifiers specifically to cut permission fatigue (Anthropic Engineering, 25 March 2026). A separate 40,000-run study by Scale X found people missed 1 in 3 genuine threats while clicking through, and 7% approved every prompt they saw. That leaves two bad escapes — approve everything by hand and get nothing done, or skip permissions and hope. Policy-based gating is the third option: your policy answers in under 120ms on the common path, and a named person is pulled in only when your rules require one, so the approvals someone does see are rare enough to deserve real attention.
Who gave Decionis the authority to stop an action?
You did. Decionis holds no policy of its own. It evaluates your organisation's policy — authored, approved, and versioned by your team, with the author barred from being their own approver — against each intended action at runtime. When your rules require a person, a named human in your organisation approves. Decionis has no opinion about what should be allowed; it applies the decision you already made, and proves afterwards which version of your policy was in force.
How is this different from logic my own engineers could write?
Any team can write a check inside one service. What is expensive is everything around it: policy authored and versioned outside the application, evaluated identically across every surface an action can leave from, maker-checker approval, a named human in the loop when the rules demand one, and a signed record anyone can verify afterwards. Decionis is the permission step as infrastructure, not business logic living inside a single app.
Do I have to replace my existing workflows?
No. Decionis sits in front of the action you already take — your workflow, your systems, your approvers. It answers one question before execution, should this run, and hands the verdict back to whatever asked. Nothing is migrated into Decionis.
Can I start without changing production?
Yes. Shadow Mode evaluates your real traffic and records what it would have allowed, held, or blocked — while stopping nothing. When you turn on enforcement, what happens on a timeout or engine error is decided by consequence: gates in front of agents, payments, and margin floors fail closed, while retail checkout with a present buyer fails open so an outage never stops a sale. Either way the degraded decision is recorded, and no signed authority is ever issued for an action policy did not evaluate.
Do my policies stay mine?
Yes. Policies are versioned inside your own workspace and are never shared between customers. If policy cannot leave your network at all, the whole control plane runs in your own VPC with one helm install.
How long does integration actually take?
Minutes for the API, an SDK, or MCP — one call in front of the action you want gated. Native platform connectors take longer because they follow each marketplace's install and review path, and an enterprise rollout across several systems is measured in weeks, not minutes.
Can I verify a decision myself?
Yes. Every verdict produces an Ed25519-signed Decision Dossier, and you can check the signature against our published JWKS offline — no account, and no need to trust our code for the result to mean something.
