Specialised agents in every module. One governed AI core.
Vigil8 is built around AI that does real GRC work — drafting, preparing, testing, mapping, reading — under the controls a regulated organisation needs.
- specialised agents
- 0+
- models trained on your data
- 0
- red-team attacks
- Weekly
- AI actions on the ledger
- 0%
specialised agents
one job each
models trained on your data
red-team attacks
against every agent
AI actions on the ledger
One gateway for every agent
Agents do not call models directly. They go through Vigil8's AI core, which applies the same rules to all of them.
Governed AI core
Guardrails in and out
Personal data and secrets redacted before the model sees them; injected instructions caught in documents and evidence; your own rules added in configuration.
Grounded and cited
Agents answer from your records and cite them. Anything they cannot ground is dropped before you see it.
The AI ledger
Every model call, tool use and decision — encrypted, append-only and readable by your auditors.
Kill switch and budgets
Stop all AI, one agent or one model at once. Budgets per team, agent and model, with alerts.
Scorecards
Per agent and month: runs, failures, test pass rate, citations kept and how often people corrected it.
Models in your region
Models deployed for you in your hosting region — or your own model, validated before it serves.
Autonomy you switch on, never assume
Suggestions under each field
Writers suggest field by field. Accept, edit or dismiss — no copy and paste, nothing applied without a click.
Findings, not decisions
Workflow agents say what they found — clear, exceptions or unsure — and people take it from there.
"Same as before", only by choice
An organisation may let a person sign off together findings identical to periods confirmed one by one. Anything new is looked at alone.
Agents learn your style — without retraining on your data
No reinforcement learning on customer data. Instead, what your people accept or correct can become an agent's examples — chosen by your administrators, removable at any time.
Examples, not training
Accepted and edited outputs are shown to the agent as examples of your preferred style.
Gated by tests
A change to an agent's examples takes effect only if its tests pass no worse with it.
Your data stays yours
Examples stay in your environment, encrypted, like everything else.
One job each, done well
Policy writer · suggestion
Cited: Access control policy v3, section 4.3