CMP Agent#

The CMP Agent is a conversational AI assistant built into Reply CMP. Ask it questions in plain language — it queries your cloud data, runs live provider metrics and log queries, manages alert rules, and can produce formatted HTML reports — all without leaving the platform.

Opening the CMP Agent#

Click the CMP Agent icon in the top navigation bar. The agent panel opens on the right side of the screen without navigating away from your current view.

The agent is always accessible, on every page. You can work in the main view (switching tabs, checking dashboards) while the conversation persists in the side panel.

Conversations and Feedback#

CMP Agent conversations are saved so you can return to previous work, resume context, or review an answer later. You can create a fresh conversation for a new topic and delete old conversations when they are no longer useful.

Use the message feedback controls to mark an answer as helpful, unhelpful, or neutral. Feedback helps improve answer quality over time, but it does not change your permissions, expose additional data, or make the Agent bypass the same access rules used by the UI.

What the Agent Can Do#

The CMP Agent covers multiple capability domains. It only works with the areas your role allows you to access, so a user who can read Discovery but not FinOps will not get cost answers through the Agent.

At a high level, the agent works like this:

Phase

What happens

Planning

The agent reads your goal and chooses which data sources are needed

Data lookup

It reads the relevant Reply CMP data under your tenant and permissions

Evaluation

It checks whether the data answers the question or whether more context is needed

Validation

Before replying, it verifies that the answer is supported by the data it collected

Final response

The agent writes the answer in the language of your request and may attach an HTML report

The agent avoids showing internal identifiers unless you explicitly ask for technical details. It translates raw fields into plain-language answers and suggests follow-up questions when useful.

1. Resource Discovery Queries#

Query your resource inventory — counts, filters, distributions, properties, and relationships. The agent uses the same rationalised Discovery model shown in the UI, including support level, category, type-specific metadata, and resource graph data.

Examples:

  • “How many Azure VMs do I have?”

  • “List all untagged GCP resources”

  • “Which resources are in the production group?”

  • “Show distinct resource types across all connections”

  • “Show the upstream and downstream relationships for this VM”

  • “Which resources are CostOnly rather than Full Discovery resources?”

  • “Show type-specific properties for the production load balancer”

2. Cloud Cost Analysis#

Query cost data with filters, group-bys, and time ranges — the same data powering the FinOps Analyze view.

Examples:

  • “What were our top 5 most expensive resources last month?”

  • “Compare this month’s Azure costs to last month”

  • “Show costs by connection for Q3”

  • “Which groups are over budget?”

3. FinOps Allocation Management#

Inspect and manage the allocation structures used by FinOps reporting: Organisation groups, Environments, Projects, owners, rules, and budgets.

Examples:

  • “Show me the current allocation hierarchy”

  • “Create an Environment called Production”

  • “Create a Project called Data Platform and add a rule for tag project = data-platform”

  • “Create a Platform Engineering group under Engineering and assign Alice as owner”

  • “Add a monthly budget of 5000 EUR to the Mobile team”

  • “Which resources are still unallocated?”

  • “Remove the obsolete Sandbox project”

Note

Creating, editing, or removing allocation groups requires FinOps.Allocation / Write. Budget changes also require FinOps.Budget / Write. The Agent asks for confirmation before applying these changes.

4. Live Cloud Monitoring Queries#

Run live metric, log, and alert queries directly against your cloud providers — the same queries that power Monitoring dashboards.

Metrics:

  • “Show average CPU utilisation for my Azure VMs over the last 24 hours”

  • “What is the Lambda error rate for the checkout function in the last hour?”

  • “Show GCP Cloud Run request count for the last 7 days”

Logs:

  • “Show the last 50 errors from the AppServiceLogs table in Log Analytics”

  • “Query Lambda errors from /aws/lambda/checkout in the last 6 hours”

  • “Find all HTTP 5xx entries in Log Analytics from today”

Alerts / Alarms:

  • “Are there any active Azure alerts with severity Sev1 or higher?”

  • “Which AWS CloudWatch alarms are currently in ALARM state?”

  • “Show active GCP alert policies for Cloud Run”

Available monitoring data:

  • “What metric namespaces are available for my Azure subscription?”

  • “List available metrics for Azure Kubernetes Service”

  • “Which Log Analytics workspaces does the agent have access to?”

5. Alert Rule Management#

Inspect and manage the Alert Rules configured in Monitoring.

Examples:

  • “List all alert rules — which ones are currently Firing?”

  • “Show the full configuration for the prod-cpu-high alert rule”

  • “What is the state history for the database-latency rule?”

  • “Disable the staging-error-rate alert rule”

  • “Run an immediate evaluation of the prod-memory alert rule”

Note

Toggling (enable/disable) an alert rule requires the Monitoring.AlertRule / Write permission. Read-only operations require Monitoring.AlertRule / Read.

6. Savings Recommendations#

Ask about cost savings opportunities and cloud reservations. The agent uses advisory recommendations from tenant connections and reservation cost tracking from FinOps.

Examples:

  • “What reservations do we have expiring soon?”

  • “Summarise our top savings recommendations”

  • “What would we save if we acted on all High confidence recommendations?”

  • “Show ongoing Azure Reserved Instance costs”

7. Automation Policies#

Inspect automation policy definitions and active policy instances.

Examples:

  • “What automation policies are available?”

  • “List all active policy instances — which ones are enabled?”

  • “Show the last 10 execution runs for the stop-dev-vms policy”

  • “Which policy instances ran in the last 7 days and used the most credits?”

8. Provisioning#

Query Provisioning Deployments and the approved Modules available to your tenant.

Examples:

  • “List all active provisioning deployments”

  • “Which approved modules are available for Azure?”

  • “Show deployments that target the production connection”

9. Tenant and Platform Information#

Query platform metadata — connections, users, identity.

Examples:

  • “How many connections are configured?”

  • “List all users and their roles”

  • “What cloud provider is the eu-prod connection?”

  • “Are we close to any tenant quota?”

  • “Show recent discovery or cost-refresh events”

10. Trigger Discovery#

Ask the agent to start a discovery sync for a connection. The agent will confirm with you before triggering.

Example:

  • “Trigger a discovery sync for the Azure production connection”

  • The agent responds: “Are you sure you want to trigger discovery for [connection name]?” → Type yes to confirm.

11. Report Generation#

Ask the agent to generate a full HTML report. The agent collects data from the relevant areas and produces a styled, downloadable HTML report.

Examples:

  • “Generate a cloud cost report for last quarter”

  • “Create a monitoring health report for this month”

  • “Generate an advisory report with savings recommendations”

The report is delivered as a download attachment directly in the chat.

Data the Agent Can Use#

Depending on your role, the CMP Agent can answer questions using:

Area

Examples

Discovery

Resource inventory, support levels, provider metadata, tags, lifecycle, and relationships

FinOps

Costs, trends, forecasts, allocation groups, budgets, recommendations, and reservations

Monitoring

Live metrics, logs, cloud alerts, Alert Rules, dashboard definitions, and Azure Resource Graph queries

Automation

Available automation policies, active policy instances, execution history, and credits used

Provisioning

Deployments and available catalog templates

Administration

Connections, refresh history, tenant quotas, users, roles, and current user identity

You do not need to know where each data point comes from. Ask in plain language and the agent chooses the right source.

Actions and Confirmations#

Most Agent capabilities are read-only. The agent asks for explicit confirmation before operations that change platform state or start background jobs.

Action

What changes

Permission

Trigger Discovery

Starts a background discovery scan for one or more connections

Discovery.Resource / Write

Toggle alert rule

Enables or disables an existing alert rule

Monitoring.AlertRule / Write

Evaluate alert rule now

Runs an immediate alert evaluation

Monitoring.AlertRule / Write

Create/edit/remove allocation groups

Changes FinOps allocation hierarchy, owners, rules, or budgets

FinOps.Allocation / Write, and FinOps.Budget / Write when budgets are involved

When confirmation is required, the chat shows a confirmation card. If you decline, the action is not executed.

Report Generation#

When you request an on-demand report from the CMP Agent, the agent first gathers the relevant data. If there is enough data to build a report, a separate report generation flow runs:

Stage

What happens

1. Data analysis

Collected data is combined into a structured analysis

2. Executive summary

A dedicated AI pass writes business-level findings

3. Detailed findings

Tables, calculations, and chart descriptions are built from the collected data

4. Insights and recommendations

A consultant-style pass extracts actions, risks, and follow-ups supported by data

5. Report assembly

A final pass creates a polished HTML report with Reply CMP styling and charts

6. Validation

The service checks that the report is complete, uses real data, and contains no placeholders

Note

Report generation takes approximately 30–60 seconds. The agent will display a “Generating html report…” status indicator while processing.

Language auto-detection: The report is written in the language of your query — the same language you used when requesting the report.

This on-demand chat report is different from the reusable FinOps Report Service used by FinOps → Reports. See FinOps Report Service for the scheduled report designer.

Inline visualisations: For non-report answers that contain numeric data, the agent renders Mermaid diagrams (charts, pie charts, flow diagrams) directly in the chat response without generating a file.

Example Conversation#

Sample chat conversation

Limitations#

Limitation

Detail

T-1 cost data

Cost data is refreshed up to the previous day. Today’s cost data is not yet available.

Discovery freshness

Inventory and relationship graph data reflect the last successful Discovery run for each connection. Trigger a manual sync when you need the latest topology.

Live monitoring data

Monitoring queries call provider APIs in real time and return current data.

Cloud resources are read-only

The Agent cannot create, edit, or delete cloud resources. Write actions affect Reply CMP state only: triggering Discovery, toggling/evaluating alert rules, or changing FinOps allocation data, each with permissions and confirmation where required.

Report generation time

30–60 seconds per report. Complex queries or large tenants may take longer.

Very large results

Very large answers may be summarized or narrowed so the chat stays usable.

Repeated requests

The agent avoids repeating the same lookup indefinitely.

Context window

Long conversations may lose context from very early messages. Start a new conversation for a different topic.

Language

The Agent understands queries in multiple languages and always responds — including generated reports — in the language of your query.

Permissions Summary#

Domain

Required permission

Resource Discovery

Discovery.Resource / Read

Trigger Discovery

Discovery.Resource / Write

Cloud Cost Analysis

FinOps.Cost / Read

Allocation read

FinOps.Allocation / Read; budgets require FinOps.Budget / Read

Allocation changes

FinOps.Allocation / Write; budget changes also require FinOps.Budget / Write

Live Monitoring Queries

Monitoring.MonitoringDashboard / Read

Alert Rule read

Monitoring.AlertRule / Read

Alert Rule toggle/evaluate

Monitoring.AlertRule / Write

Automation Policies

Automation.PolicyInstance / Read

Provisioning

Provisioning.Deployment / Read

Catalog items

Provisioning.Catalog / Read

Connections

Onboarding.Connection / Read

Recommendations

Onboarding.Tenant / Read

Users

Onboarding.User / Read

Quotas

Onboarding.Tenant / Read

Identity

No additional permission

Users with the built-in Owner, Contributor, or FinOps Contributor roles have access to all Agent domains. Users with Monitoring Reader can ask live monitoring questions and read alert rules. See Roles & Permissions for the full reference.

Tips for Best Results#

  • Be specific about time ranges: “last 30 days” or “October 2025” works better than “recently”

  • Reference your group names: “Show costs for the Frontend group” — use your actual allocation group names

  • Ask for allocation setup help: the Agent can create groups, environments, projects, owners, rules, and budgets when you have write permissions

  • Ask follow-up questions: the Agent maintains conversation context, so “Which of those are in prod?” works after a list response

  • Use Discovery terms: ask for “Full resources”, “CostOnly resources”, “type-specific properties”, or “resource graph” when you need the current Discovery model

  • Use metric discovery first: for monitoring queries, ask “What metric namespaces are available?” or “List metrics for Azure VMs” before querying a specific metric

  • Specify provider for monitoring: “Show AWS alarms” or “Query Azure metrics” to avoid ambiguity when you have multi-cloud connections