Governance
You see everything your AI does, and you stop it in real time
By default, everything passes. Your rules say what must not, for your teams and for external agents alike. You see first, then you block. Every decision is traced, and with Enterprise the log is sealed: facing an auditor, you show the record instead of declaring.
No credit card
Decide
4 possible decisions
A rule does more than refuse. It lets through, stops the request, keeps a doubtful answer to itself, or lets through while flagging. You dose control without slowing your teams.
- Allow
The request proceeds. This is the default decision when no rule applies.
- Deny
The request is stopped outright: the question, action or hand-off it targets does not happen. For what must never be attempted.
- Withhold
The agent does its work, then its answer is checked before delivery. Not reliable, for instance claims your documents do not support? Contains data that must not leave? The user receives your message instead. To judge on the evidence, not on the question.
- Flag
The request goes through and the event is marked in the log, so someone takes a look.
Enforce
8 moments where a rule can act
You choose when the rule steps in, from the incoming question to the delivered answer. Block a question, deny access to a document, forbid an action, bound a hand-off, withhold an answer: every moment has its rule.
- 01 When the question arrives Turn entry
- 02 When a previous answer is reused Response cache
- 03 When the agent searches your documents Retrieval
- 04 When the agent acts in one of your tools Tool call
- 05 When the agent hands off to another agent Agent delegation
- 06 When the answer goes out to the user Turn exit
- 07 When an administrator changes a rule or a setting Administration
- 08 When someone opens a cited document Document opening
Mask
10 data classes, 6 placements
A detector recognizes sensitive data and replaces it before it leaves. In the question, the history, the context sent to the model, tool arguments, delegation and the answer.
What is detected
- Phone
- IBAN
- Social security number
- Payment card
- Postal address
- Postal code
- IP address
- API key
- Date of birth
Where masking applies
- Query
- History
- Model context
- Tool arguments
- Delegation
- Response
Prove
Every action leaves a trace
An auditor, a customer or the AI Act asks what your AI did? The log answers, line by line. Nobody can alter it afterward: it is append-only. With Premium and Enterprise, every entry is chained to the previous one and you can export it; with Enterprise, it is also sealed in segments and timestamped by a third-party authority.
What the log tells
- Every question evaluated, and the decision taken
- Every refusal: document access, action in a tool, hand-off to another agent
- Every answer withheld, every piece of data masked, every hidden text flagged
- Every rule created, changed or deleted, every export, every verification of the log
The 14 recorded actions
- Turn evaluated
- Query blocked
- Document access denied
- Tool call denied
- Delegation denied
- Response withheld
- Content masked
- Hidden text flagged
- Engine failure
- Rule created
- Rule updated
- Rule deleted
- Chain verified
- Export generated
Every decision is traced
The audit log keeps a record of every request and the decision taken. With Enterprise, it is sealed, timestamped by a third party and can be exported for an audit.
Audit log
app.guardweaver.ai
| Time | Who | Request | Decision |
|---|---|---|---|
| 09:42:18 | HR assistant | Question about the collective agreement | Allowed |
| 09:41:55 | External agent | Export of the customer file | Blocked |
| 09:41:02 | Payroll assistant | Social security number in a request | Masked |
| 09:40:37 | Sales agent | CRM record update | Allowed |
| 09:39:12 | Legal assistant | Contract sent to an unauthorized model | Blocked |
Every line is written append-only: it can be neither edited nor deleted.
Governance, plan by plan
Rules, sensitive data and the audit log in each plan, as of November 16, 2026. The full detail is on the pricing page. Offers →
| Governance | Free | Premium | Enterprise |
|---|---|---|---|
| AI usage rules What AI is allowed to do: flagged in observation, stopped in blocking. | 3 rules in observation | Unlimited in observation, 5 in blocking | Unlimited, in observation and blocking |
| Sensitive data Emails, IBANs, card numbers… masked before they reach the AI. | Not included | Automatic masking | Automatic masking |
| Audit log | 3 days, lightweight version | 1 year | Contract term, sealed and timestamped export |
| AI spending caps You cap spending with your AI provider, per person, agent or model. | Not included | Per user, agent and model | Per user, agent and model |
| Roles and permissions Owner, administrator or member, per person or per team, with a role per workspace and extra rights. | Included | Included | Included |
| Shadow AI Detects the AI tools used without your approval. coming soon | Not included | Included at launch | Included at launch |
| Dashboard Usage, spending, questions, satisfaction and channels. | Included | Included | Included |
| External agent management (A2A) Agents from other tools go through your rules too. | Not included | Included | Included |
| Audit and compliance | Free | Premium | Enterprise |
|---|---|---|---|
| Immutable log No line can be edited or deleted. | Included | Included | Included |
| Log integrity check You check at any time that nothing was altered. | Not included | Included | Included |
| Log sealed and dated by a third party Evidence you can show during an audit. | Not included | Not included | Included |
| Data export | Not included | Included | Included |
Blog
Guides and articles
Our practical guides on this topic, to decide and deploy.
Governance questions
- Does governance slow users down?
- Not by default: with no rule, everything passes. Rules apply at the precise moment they target, and only the requests concerned are refused, withheld or flagged.
- Are external agents covered?
- Yes. An external agent going through your MCP or A2A servers crosses the same enforcement points as your teams: tool call, delegation, output. Its decision is traced like the others. See the Orchestration page.
- How does an auditor verify the log?
- Every line is chained, with Premium and Enterprise, and the chain verifies end to end. With Enterprise, segments are sealed and timestamped by a third-party authority, and the export delivers the lines with their integrity proof, verifiable offline. Retention depends on the plan: 3 days on Free, 1 year on Premium, the contract term on Enterprise. The access log is described on the Security page.
- Which plan do I need for governance?
- It is in all three plans, with different limits. Free observes with 3 rules and a light audit log. Premium observes without limit, blocks up to 5 uses and keeps the audit log for 1 year, at €12 excl. tax per user per month billed annually. Enterprise blocks without limit and seals and timestamps the log, with custom pricing. Compare the plans on the offers page.