Back to NeuralWorks
Layer 3 · Governance & Control

Governance for agents that act.

Most AI governance frameworks address what a model may say. Ours addresses what an agent may do — to your data, your systems and your customers.

Risk gates

Not every action deserves the same scrutiny.

Reading a report is not deleting a table. NeuralWorks classifies every agent action on four axes — action type, scope, reversibility, novelty — and routes it to one of four gates.

Tier 1
Proceed
Executes. Logged.
Tier 2
Notify and proceed
Executes, notifies, stays cancellable for a defined window.
Tier 3
Wait
Queued for human approval. Timeout cancels — never proceeds by default.
Tier 4
Block
Refused. The agent re-plans. Logged as a planning failure.
Hard scope guards

Above all four sit hard scope guards — maximum records per action, a tool allowlist per initiative, rate ceilings on even the lowest tier. Guards run before classification. If an action trips one, it is blocked regardless of how harmless its tier would have been.

Privacy

The action stays on the record.
The identity goes.

Personal data lives in a separate erasable store, referenced by internal ID. Where personal data must appear in a log, it is written as a token.

An erasure request deletes both mappings in a single operation across every store. The audit trail survives intact; the person is no longer in it.

Data residency

Cloud-agnostic, by design and by necessity.

Containment, identity and audit are architectural properties of NeuralWorks, not features of a particular cloud. The same framework deploys wherever your obligations require — a specific region, your own infrastructure, or fully in-country — without re-engineering the control model.

RegionRegimeDeployment
European UnionGDPREU-resident cloud, EU-only inference endpoints where required.
United KingdomUK GDPRUK-resident cloud, adequacy-based transfers where permitted.
United Arab EmiratesTDRA / UAE PDPLIn-country deployment; sovereign region where mandated.
Saudi ArabiaCITC / SDAIAIn-Kingdom deployment; approved model providers only.
IndiaDPDP ActIndia-resident cloud; consent architecture in the identity layer.
SingaporePDPASingapore-resident cloud; regional endpoints for inference.
AustraliaPrivacy ActAU-resident cloud; agent inference in-region.
Start here

Start with one thing that works.

A 60-minute working session with two of our engineers. You bring a delivery problem. You leave with a written assessment of where AI would help, where it would not, and what it would take to find out.

No pitch deck. No obligation.