AI Trust Center 2.0
AI You Can Review, Control and Audit
How Prakura designs AI agents for enterprise control: what they may access, when people approve, how quality is measured and how every action is recorded.
Eight commitments
How Prakura AI is governed
Human Oversight
AI suggests; people decide. High-impact actions always need a named human approver.
- • Approval rules per action
- • Override and feedback on every output
- • Escalation to a human at any time
Data Boundaries
Agents read only the sources their role allows — and never more than the person using them may see.
- • Authorized sources only
- • Role-inherited access
- • Sensitive fields excluded by default
Explainability
Every recommendation can answer "Why am I seeing this?" with the inputs it used.
- • Reasons shown in plain language
- • Sources cited
- • Confidence labels
Model Evaluation
Quality, groundedness and safety are evaluated before and after release.
- • Test sets per agent
- • Regression checks on change
- • Human review samples
Agent Permissions
Each agent has an explicit list of what it may read, write or request.
- • Least privilege
- • Approval-required actions
- • Denied by default
AI Safety
Agents refuse out-of-scope or high-risk requests and say why.
- • Scope limits
- • Sensitive-topic handling
- • No automated hiring, credit or clinical decisions
Prompt Security
Instructions found inside documents or web pages are treated as data, not commands.
- • Prompt-injection testing
- • Tool-call validation
- • Output filtering
Auditability
Who asked, what the agent did, who approved — recorded for review.
- • Event log per workflow
- • Approver identity
- • Exportable audit trail (future)
Agent Permission Visualizer
What each agent may — and may not — do
Agent: HR Assistant
| Action | Permission |
|---|---|
| Read HR policy | Allowed |
| Read own leave balance | Allowed |
| Read employee salary | Not allowed |
| Modify payroll | Not allowed |
| Send email | Requires approval |
| Create leave request | Requires approval |
Future architecture concept. Permissions are denied by default and an agent can never exceed the access of the person using it.
AI Human Approval Framework
AI suggests. People decide.
This pattern appears in every high-impact demo across the site.
- 1AI Suggests
- 2Employee / Manager ReviewsHuman
- 3Approval
- 4Action
AI Audit Trail
Every step recorded
Workflow W-303 · Placement drive bootcamp
10:32 · Placement Agent (AI)
Generated recommendation: 42 students for the Acme drive bootcamp.
10:33 · System (System)
Approval required — action affects student records.
10:35 · Placement Manager (Human)
Reviewed recommendation; removed 3 students already placed.
10:37 · Placement Manager (Human)
Recommendation approved (39 students).
10:38 · System (System)
Invitations queued for sending (demo — nothing sent).
AI Evaluation Dashboard
Quality measured before and after release
Simulated metrics showing what Prakura evaluates. These are not production measurements.
Response Quality
86%
Rubric score on demo test set
Groundedness
92%
Answers supported by cited sources
Safety
98%
Out-of-scope requests handled correctly
Tool Success
89%
Tool calls that completed as intended
User Feedback
81%
Thumbs-up share (demo)
Human Overrides
7%
Share of outputs changed by reviewers
AI Governance Explorer
Where is your AI governance today?
Rate each area; we suggest the next step. Nothing is stored or sent.