AI governance
Govern AI across your whole organisation
Set the guardrails once and they apply to every agent and every surface. Cap spend before it happens. Keep an audit trail nobody can edit.
Guardrails
Presets you can make mandatory
Three guardrail presets are named after PSPF security classifications. Each masks tax file numbers and Medicare numbers. Opt-in filter policies add driver licence numbers for every state and territory, passport numbers, and checksum-validated ABNs and ACNs. Set a preset as mandatory and it overrides every agent in your organisation.
Baseline content safety and prompt-attack detection. Masks tax file numbers, Medicare numbers, card numbers, emails, and phone numbers. Blocks hate, insults, sexual, violent, and misconduct content.
Adds word and topic policies, and blocks emails, phone numbers, names, and IP addresses. Grounding checks apply when you supply sources.
Adds expanded PII, classification markers, and sensitive government topics, with a stricter grounding threshold.
Spend
Cap it before it happens
Spend limits are checked before a request runs, not reported afterwards. Set a monthly and daily ceiling for the organisation, then narrow it per user, per API token, and per interface.
- Monthly and daily caps for the organisation
Per-user caps, with named overrides for the ones who need more
Per-token caps, so an integration cannot spend the team’s budget
Per-interface caps for Slack, autonomous jobs, streaming, and the API
A warning email to your governance managers when spend crosses the threshold you set
Model policy
Choose which models your organisation may use. Run an allowlist when you want a short approved set, or a blocklist when you want everything except a few. Unrestricted is also available.
Model choice is enforced at the request, so a request for a model outside the policy is refused.
Filter policies
Built-in policies cover the common cases, such as detecting secrets in prompts. Add up to 50 of your own with rules that match your organisation’s risks.
Mark policies mandatory and they apply to every chat, completion, and agent request regardless of how an individual agent is configured. Enable and disable them per organisation without deleting the definition.
Audit
A record nobody can edit
Every chat, agent, completion, embedding, and AI Search request writes an audit record to object storage under Object Lock in compliance mode, with a seven-year retention period. Records cannot be modified or deleted during that period — not by your administrators, and not by us.
Each record carries the user, the API token, the session, the model, token usage, and hashes of the prompt and the response. Export a date range from the Governance page as JSON Lines, then filter by user or by token.
When someone asks who used what and when, you answer from the record rather than from memory.
What this makes possible
Governance is what turns a pilot into something you can put in front of a risk committee.
Pass a security review
Answer the control questions with configuration rather than policy documents. Guardrails, model policy, and audit are all demonstrable.
- Demonstrable controls
- Exportable evidence
- No custom work
Run a pilot without a blank cheque
Cap the organisation, the team, and the integration separately. A runaway agent hits its ceiling instead of your invoice.
- Pre-request checks
- Per-interface caps
- Warning thresholds
Delegate without losing control
Let teams build their own agents while mandatory guardrails and filter policies stay enforced above them.
- Mandatory presets
- Org-wide policies
- Per-user attribution
Put controls around AI
Talk to our team about governance for your organisation.