Evidence-grounded explanations
Structure assistance around what was observed, its impact, likely cause, confidence, evidence, contradictions, missing signals and the next validation.
Assistance grounded in your evidence
Ask better questions. Keep control of the answer.
Ask questions about your environment and follow the answer to its evidence. AmrudAi Copilot summarizes incidents, retrieves relevant context, compares explanations and recommends the next useful check while your team controls data access and privileged actions.
The application experience
Investigation Workbench
See the complete workspace. Enlarge any image to inspect its panels, navigation and evidence at actual size.
Evidence-Led RCA Workbench — illustrative product view with sample data. Illustrative product view · Sample data
Bring hypotheses, contradictions, configuration changes and remediation checks into one investigation. Shift Handover & Operational Memory — illustrative product view with sample data. Illustrative product view · Sample data
Carry verified facts, ownership and next checks into the incoming shift. The capability in detail
Keep assistance connected to evidence and control.
Bounded evidence retrieval
Reduce large telemetry sets into relevant structured, graph and keyword context before requesting a language-model explanation.
Suggested next tests
Rank permitted diagnostics by usefulness, safety, latency and cost. A suggestion still follows the normal authorization and execution path.
Similar incidents and runbooks
Explain why historical incidents or knowledge references may help, while keeping previous outcomes separate from the current diagnosis.
Post-incident assistance
Help summarize an evidence timeline, accepted cause, workaround, corrective action and review notes without inventing a test result.
Local, private and hosted choices
Use a governed provider abstraction for approved local, private-hosted or hosted inference, plus operation with language models disabled.
An independent monitoring core
Keep collection, alerts, topology, incident evidence, safe manual workflows and deterministic RCA independent of AI availability.
Resource-aware inference
Bound CPU, memory, context, concurrency and queue time. Defer or unload optional inference before it can disrupt protected monitoring.
No model runtime on endpoints
Keep routine endpoint logic focused on deterministic triggers and evidence collection. Normal Agents do not need an LLM or GNN.
A deterministic tool firewall
Validate proposed tools, parameters, resource scope, credentials, preconditions, approvals and audit. Generated text cannot execute privileged actions directly.
Controlled data and model lifecycle
Apply masking, provider/region and retention policy; signed offline artifacts; model provenance; compatibility; evaluation and rollback. No silent cross-tenant learning.
Connected predictive intelligence
Use cohort and seasonal baselines, anomaly detection, forecasts and learned graph context to focus investigations. Drift monitoring and reviewed model updates keep analytical behavior visible.
From information to action
A connected operational workflow.
- 01
Retrieve authorized evidence
Resolve the incident, resource and permitted context before any model request.
- 02
Reduce and explain
Build a bounded evidence summary and separate observations from hypotheses.
- 03
Validate the proposal
Check grounding, contradictions, permissions and action limits through deterministic controls.
- 04
Review and improve
Preserve model, prompt, evidence and operator-feedback provenance; evaluate changes before promotion.
Choose how intelligence operates
Fit the way your team operates.
Use GNN-assisted candidate ranking, learned correlation, local language models and Copilot alongside deterministic RCA, with evidence and operator authority preserved.
- Choose local, private or approved hosted inference through the AI Gateway. Configure model access, permitted context, data masking and egress by customer policy.
- Bound inference with CPU, memory, time and concurrency budgets. Model work yields to protected collection and monitoring when resources are constrained.
- Keep customer evidence and feedback tenant-scoped. AI proposals pass typed tools, role checks, target restrictions and approval policy before any privileged action.
What can the AI Copilot access and do?
Is a language model required for monitoring or RCA?
No. Deterministic collection, qualification, incidents and RCA remain independent of a language model. AmrudAi supports all AI assistance being disabled.
Will AmrudAi place an always-on AI model on each monitored host?
No. Normal endpoint Agents use deterministic capture eligibility and bounded evidence collection. Optional model inference belongs behind the Core AI Gateway or an approved private inference boundary.
Can Copilot automatically change network configuration?
Not directly. Any action must pass deterministic authorization, target and parameter checks, approval policy, impact limits, audit and post-action validation.
Can customer data be sent to an external provider?
Only within the configured, approved data-egress policy. Provider destination, region, data classes, masking, retention and opt-out must be governed. Choose local, private or no-LLM operation to match your deployment policy.
Are predictive results guaranteed?
No. Forecast or anomaly outputs require relevant backtesting and calibration. They remain supporting analysis and must disclose uncertainty, comparison population and missing evidence.
See the workflow in your environment.
Explore the technology, service dependencies and operating requirements that matter to your team.