The CloudWise Agent

CloudWise is the agent for your AWS bill. You talk to it in plain English in the Workspace; it answers, renders the right view inline, and hands you the fix. This guide covers how the agent works and the kinds of questions it answers.

"AI Copilot" is now "the agent"

Earlier versions of CloudWise called this an "AI Copilot" that lived in a panel in the corner of the screen. It's now the agent at the center of the Workspace—the conversation is the spine of the product, not a bolt-on. The Workspace still labels it "Copilot" (for example the "Copilot … today" question meter). It replaces the old corner panel, and some of that panel's features (saved conversation history, forecasts, the per-answer analytics line) did not carry over.

🤖 What does the agent do?

The agent understands your AWS cost and waste data and answers naturally—no query language, no hunting across dashboards. Ask a question and it replies in plain English and renders the matching structured view (a table, a chart, a finding card with its dollar impact). For waste, it proposes the fix; on the Agentic plan it can execute the fix after you approve it.

Key Capabilities

• 📊 Natural Language Queries - Ask questions like "Why did my costs spike last week?" instead of writing complex queries
• 🧩 Generative, structured views - The agent renders tables, charts, and finding cards inline—not chat-only
• 🔍 Your own data - Answers trend, alert and waste-finding questions from your cost data and scan findings (anomaly alerts are on the Shield and Agentic plans)
• 🗑️ Waste Detection Integration - Automatically surfaces idle resources, unused EBS volumes, and optimization opportunities
• 🔧 Propose-and-approve remediation - Shows the fix as AWS CLI commands; on Agentic it can run an approved fix and roll it back where the action can be undone
• 💡 Follow-up suggestions - Suggests follow-up questions after an answer
• 📈 Trends & budgets - Answers questions about spend over time, budget status and anomaly alerts

✨ The verb ladder—what each plan unlocks

The agent is on every plan. What changes as you climb is how much it can do for you—from Ask to Fix.

Ask — Free ($0 forever)

5 questions per day

• ✅ Talk to the agent about your cost and waste data
• ✅ Cost breakdowns, top services, trends
• ✅ Waste findings with propose-only fixes (ready-to-run commands)
• ✅ 1 AWS account

Perfect for: Individual developers, personal projects, getting started with the agent

Watch — Shield

$19/mo or $149/year

50 questions per day

• ✅ Everything in Ask
• ✅ Longer cost history than Free
• ✅ Cost anomaly alerts the agent can explain
• ✅ Cost anomaly detection + Slack/email alerts
• ✅ Up to 5 AWS accounts

Perfect for: CTOs, Engineering Managers, and Lead Developers who own the AWS bill but don't have time to become FinOps experts. Ask "Why did my bill spike?" and get a plain English answer.

Fix — Agentic AI

$49/mo or $399/year

100 questions per day

• ✅ Everything in Watch
• ✅ The agent executes approved remediations in your account (with one-click rollback where the action can be reversed)
• ✅ Remediation queue with propose-and-approve controls
• ✅ Monthly remediation credits

Perfect for: Teams that want the agent to close the loop—not just find the waste, but clear it.

Govern — Compliance

$99/mo or $799/year

500 questions per day

• ✅ Everything in Watch
• ✅ Air-Gapped Mode with 365-day retention
• ✅ Upload history (12 uploads), QoQ variance, PDF audit reports
• ✅ Persistent anonymization salt, QoQ variance by tag

No fixes run by CloudWise, by design: Compliance is for regulated teams that cannot connect AWS accounts to third-party tools. Data arrives as anonymized offline uploads, so CloudWise makes no plans and runs no fixes on this plan (those are Agentic only). Findings still show a recommendation and Action steps you apply yourself.

Perfect for: Regulated and security-focused teams who need air-gapped AWS cost analysis with audit-ready reporting.

🚀 How to Use the Agent

Getting Started

1. Open the Workspace - The agent is the center of the Workspace; just start typing
2. Ask Your Question - Type naturally, as if talking to a colleague
3. Review the Response - The agent answers in plain English and renders the matching view (table, chart, or finding card)
4. Act on a finding - Use the proposed command, or (on Agentic) approve the agent to execute the fix
5. Follow Up - Ask clarifying questions or dive deeper; the agent keeps context across the conversation

What to Ask

High-level questions (every plan)

• "What are my top 5 services?"
• "How much am I spending this month?"
• "Which region costs the most?"
• "Show me my cost breakdown"

Deeper analysis (longer history on Shield and above)

• "Why did spend spike on the 10th?"
• "Compare last month vs this month"
• "What are my top services over the last 90 days?"

💬 Example Queries

🗑️ Waste Detection & Optimization

"What waste do I have in my accounts?"

• Surfaces idle EC2 instances, unattached EBS volumes, and unused resources
• Shows potential monthly savings with confidence levels
• Provides actionable CLI commands to clean up

"Show me idle resources I can clean up"

• Lists resources by waste category (compute, storage, networking)
• Prioritizes by savings potential (high confidence items first)
• Includes step-by-step remediation actions

"How can I reduce my AWS spending?"

• Combines cost analysis with waste detection findings
• Recommends specific resources to terminate or rightsize
• Shows total monthly savings opportunity with breakdown

"What are my top optimization opportunities?"

• Ranks opportunities by potential savings
• Explains each finding with context
• Provides both quick wins and longer-term recommendations

"Are my Glue jobs optimized?"

• Detects oversized jobs wasting DPU-hours
• Flags missing timeouts that risk runaway billing
• Identifies failing jobs burning money on retries
• Warns about Data Catalog growth exceeding the free tier

"Are any Transfer Family servers idle?"

• Identifies servers with no file transfer activity in 30+ days
• Flags unused protocols costing $216/month each
• Detects idle web apps costing $360/month per unit
• Provides stop/delete CLI commands with savings estimates

"Are any Global Accelerators idle or disabled?"

• Detects accelerators with zero traffic for 30+ days ($25.55/month each)
• Flags disabled accelerators still incurring fixed hourly charges
• Identifies accelerators with no endpoints configured
• Provides delete CLI commands with static IP release warnings

"Are my DocumentDB clusters right-sized?"

• Detects clusters with CPU < 20% and low connection counts
• Identifies old manual snapshots accumulating storage costs
• Recommends instance downsizing with savings estimates
• Advisory-only guidance for overprovisioned clusters

"Are any of my FSx filesystems oversized or overprovisioned?"

• Detects filesystems with storage utilization below 40% of provisioned capacity
• Identifies throughput overprovisioning for Windows/ONTAP/OpenZFS filesystems
• Flags old manual FSx backups past retention threshold
• Advisory-only guidance for storage/throughput rightsizing; executor support for old backup cleanup

"Are my Step Functions workflows efficient?"

• Detects retry storms: a high share of state transitions are retries, a meaningful share of executions fail, and there are enough executions to judge
• Identifies high transition density (>50 transitions per success) wasting state transition costs
• Flags Express workflows with P95 duration above 30 seconds (better suited for Standard)
• All findings are advisory-only with actionable optimization guidance

"Are my OpenSearch domains right-sized?"

• Detects oversized domains with avg CPU < 20% and 95% of hours under 40% (one spike does not block it) — recommends stepping down instance type
• Flags EBS volume overprovisioning when free storage exceeds 60% with low growth rate
• Identifies Reserved Instance purchase opportunities from AWS Cost Explorer RI recommendations
• Advisory-only guidance for domain rightsizing; RI purchases require manual action

"Are my ElastiCache clusters optimized?"

• Detects idle clusters with zero connections for 7+ days
• Flags oversized nodes with CPU below 10%
• Identifies non-production clusters running unnecessary replicas (replication waste)
• Recommends Valkey migration for Redis OSS and Memcached clusters, priced from AWS's Valkey rate for the same node type (the saving varies by node type; Memcached needs application changes)
• Flags spiky workloads better suited for ElastiCache Serverless
• Identifies large R6g clusters eligible for R6gd data tiering (up to 52% savings)
• Checks engine versions for Extended Support surcharges

"Can I save by migrating from Redis to Valkey?"

• Checks which clusters run Redis OSS or Memcached
• Calculates the saving per cluster from AWS's Valkey vs Redis OSS node pricing for the same node type
• Notes Redis 7.x API compatibility and existing RI applicability

"Do any dev/staging caches have replicas?"

• Uses environment tags to identify non-production clusters
• Flags replicas that cost the same as the primary node
• Calculates per-replica savings across all shards

"Is my Neptune cluster over-provisioned?"

• Analyzes CloudWatch CPUUtilization over 14 days
• Flags instances with <20% average CPU for downsizing
• Calculates savings from moving to the next smaller instance class

"Should this Neptune cluster be Serverless?"

• Compares provisioned instance cost against Neptune Serverless ($0.1608/NCU-hr), estimated per instance at 2 NCU or more
• Only recommends Serverless when provisioned cost exceeds that estimate
• Warns if SPARQL is used (not supported on Serverless)

"Are there old Neptune snapshots I should clean up?"

• Finds manual cluster snapshots older than 90 days
• Calculates storage cost at $0.021/GB-month
• Flags snapshots from migrations, upgrades, and pre-delete backups

�️ WorkSpaces Optimization

"Which WorkSpaces should switch from AlwaysOn to AutoStop?"

• Identifies AlwaysOn WorkSpaces disconnected for 7+ days
• Calculates break-even hours and monthly savings from switching to AutoStop
• Only flags when savings exceed $10/month to avoid false positives

"Do I have oversized WorkSpaces?"

• Analyzes CloudWatch UserSessionsCount over 14 days
• Recommends conservative single-step bundle downgrades (e.g., PowerPro → Power)
• Shows estimated monthly savings per WorkSpace

"Are my WorkSpaces Pools overprovisioned?"

• Detects pools with utilization below 75%
• Calculates excess capacity cost per slot by running mode: $36.50/month for an always-on pool (the lowest running pool rate in any region), $18.25/month for an auto-stop pool. These are the lowest rates in any region, so most pools cost more
• Only flags pools with 2+ excess slots and savings above $25/month

"Can I reduce WorkSpaces licensing costs with BYOL?"

• Groups Windows license-included desktops across bundles
• Estimates $4/desktop/month savings from Bring Your Own License
• Only flags when 5+ eligible desktops detected
• Includes licensing compliance disclaimer

�🔍 Discovery & Exploration

"What are my biggest cost drivers?"

• Returns top services/regions with cost breakdown
• Suggests areas for optimization

"Show me my cost trends over time"

• Displays spending patterns
• Highlights increases or decreases

🎯 Targeted Analysis

"Why did my costs spike last week?"

• Reads your daily cost trend
• Checks your cost anomaly alerts (anomaly alerts are on the Shield and Agentic plans)
• Ranks your top services for the period

"How much am I spending on RDS?"

• Ranks your costs by service, region or account
• Lists your waste findings and their savings

🤖 SageMaker Optimization

"Are any of my SageMaker notebooks idle?"

• Identifies InService notebooks with no recent activity
• Shows instance-type-based pricing, reconciled against your actual bill where your CUR allows it (never a generic estimate)
• Provides stop/delete CLI commands

"Which SageMaker endpoints can I right-size?"

• Analyzes CPU and memory utilization via CloudWatch
• Flags oversized endpoints (low CPU + low memory)
• Detects previous-generation instance types (ml.m4, ml.c4) where the current generation is cheaper
• Recommends current-gen replacements with savings estimates

� CloudWatch Logs Optimization

"Do any of my log groups have excessive retention?"

• Identifies log groups with 365+ day retention and significant stored data
• Calculates savings from reducing to 30-day retention ($0.03/GB/month)
• Shows reduction factor (e.g., 365→30 days saves 92% of storage cost)

"Show me empty or orphaned log groups"

• Finds log groups with 0 bytes stored that are 30+ days old
• Typically left behind by deleted Lambda functions or ECS services
• Helps reclaim log group quota (see the CloudWatch Logs quotas)

📈 Trends & Budgets

"Show my spend trend over the last 90 days"

• Charts your total spend by day

"Am I on track with my budget?"

• Compares current spending to your budget alerts
• Alerts to potential overages
• Suggests cost-saving measures

🔬 Root Cause Analysis

"What caused the Cost Explorer increase on July 17?" (Shield)

• Day-by-day analysis of which services drove an increase
• Anomaly alerts (Shield and Agentic plans) show the service, baseline, current cost and percent change

"Why are my Lambda costs higher than usual?"

• Shows Lambda spend over time
• Lists any Lambda waste findings (unused functions, over-provisioned memory, idle Provisioned Concurrency, deprecated runtimes)

🗑️ Waste Detection + the Copilot

How It Works

CloudWise combines automated waste detection with the agent. The agent decides which of its read-only tools to call (waste summary, findings, spend, commitments, budgets, alerts, remediation activity), reads the results from your latest scan and cost data, and writes a short answer with cards. It can run several tool rounds for one question.

Responses include:

• Total monthly savings potential
• Top waste items by priority
• AWS CLI commands for cleanup
• Step-by-step remediation guidance

What Gets Detected

Category	Examples	Detection Method
Compute	Idle EC2 instances, unused Elastic IPs, idle Lambda Provisioned Concurrency, excessive Lambda timeouts, ARM64 migration opportunities, idle Lightsail instances, unattached Lightsail static IPs/disks, old Lightsail snapshots, idle Lightsail load balancers/databases, idle/oversized WorkSpaces, WorkSpaces AutoStop opportunities, overprovisioned WorkSpaces Pools, WorkSpaces BYOL licensing, idle Elastic Beanstalk environments, Beanstalk idle traffic, unnecessary ALB on single-instance Beanstalk, Beanstalk previous-gen instances, over-provisioned Beanstalk, orphaned Beanstalk RDS	CPU utilization, network activity, PC utilization, invocation metrics, attachment status, snapshot age, connection count, WorkSpaces connection/session metrics, pool utilization, environment health, RequestCount metrics, ALB config, instance type generation
Storage	Unattached EBS volumes, old snapshots, orphaned snapshots (source volume deleted or AMI deregistered), S3 rapid growth, S3 wrong storage class, S3 empty buckets, old backups, redundant backups, backup lifecycle tiering, stale backup plan assignments, oversized FSx filesystems, FSx throughput overprovisioned, old FSx backups	Attachment status, age analysis, volume/AMI cross-reference, CloudWatch BucketSizeBytes metrics, Backup vault enumeration, FSx FreeStorageCapacity/throughput metrics
Database	Idle RDS instances, over-provisioned Aurora, Aurora I/O optimization, Aurora Extended Support surcharges, Aurora Serverless v2 opportunities, ElastiCache (7 detectors: idle, oversized, extended support, replication waste, Valkey migration, Serverless optimization, data tiering), Neptune Serverless opportunities, over-provisioned Neptune, old Neptune/DocumentDB snapshots, overprovisioned DocumentDB clusters	Connection count, CPU/request metrics, I/O metrics, engine version, instance class, snapshot age, replication topology, environment tags, connection variance
Networking	Unused NAT gateways, orphaned load balancers, low-traffic ALBs, high LCU cost ALBs, Classic LB migration	Traffic analysis, target health, LCU cost analysis
Containers	Idle ECS services, oversized ECS tasks, missing auto-scaling, unnecessary Container Insights, over-provisioned memory, unused ECR images	Task count, CPU/memory utilization, auto-scaling config, image pull dates
Analytics	Idle EMR clusters, long-running EMR clusters, over-provisioned EMR, EMR auto-termination gaps, previous-gen EMR instances, EMR Spot opportunities, idle/over-provisioned Kinesis streams, On-Demand downgrade, extended retention waste, idle fan-out consumers, idle Firehose, oversized MSK clusters, idle/oversized MQ brokers, idle/oversized Glue jobs, failing Glue jobs, unused crawlers, Data Catalog bloat	API status, run history, CloudWatch metrics, YARN utilisation, instance groups, catalog object counts
ML/AI	Idle SageMaker notebooks, idle endpoints, oversized endpoints, stopped notebook storage, previous-gen instances	Activity, invocations, CPU/memory utilization, instance type
Serverless	Unused Lambda functions, deprecated Lambda runtimes, unused API Gateways, unused AppSync APIs, idle AppSync caches, idle AppSync subscriptions, Step Functions retry storms, high transition density, Express duration waste	Invocation history, runtime version, CloudWatch execution/transition metrics, cache hit/miss ratios, WebSocket connection metrics
Management	No-retention log groups, stale log groups, excessive retention log groups, empty orphaned log groups	Retention policy, stored bytes, age analysis
Commitments	Unused Savings Plans, expiring RIs/SPs, convertible RI exchanges, SP coverage gaps	SP utilization, expiry dates, pricing analysis

EC2 Reserved Instance & Savings Plan Intelligence

The Copilot understands the full commitment lifecycle and can help with:

• Buy analysis: "Which instance types should I commit to?" → Powered by ri_opportunity_* and sp_opportunity_* detectors

• Utilization review: "Show me any Savings Plans I'm not fully using." → Powered by the unused_savings_plan detector

• Expiry alerts: "What commitments are expiring in the next 60 days?" → Powered by expiring_reserved_instance and expiring_savings_plan detectors

• Exchange opportunities: "Can I upgrade any of my Convertible RIs for free?" → Powered by convertible_ri_exchange_opportunity detector

Sample AI Response

When you ask "What idle resources can I clean up?", the agent responds with a finding card like this (amounts shown as placeholders—your real numbers come from your scan):

💰 Total Monthly Savings Potential: $X,XXX

I found N waste items across your accounts. Here are the top opportunities:

🔴 High Priority (High Confidence)

1. Unattached EBS Volume - vol-0abc123def456
   • Account: Production (123456789012) | Region: us-east-1
   • Monthly Waste: $XX.XX
   • Action: aws ec2 delete-volume --volume-id vol-0abc123def456
2. Idle EC2 Instance - i-0def789abc012
   • Account: Development (987654321098) | Region: eu-west-1
   • Monthly Waste: $XX.XX
   • CPU utilization below 2% for 14 days
   • Action: aws ec2 stop-instances --instance-ids i-0def789abc012

💡 Recommended Next Steps:

• Review each finding in the Workspace
• Start with high-confidence items first
• Consider stopping instances before terminating (recoverable)

Waste questions to try

• "What waste do I have across all accounts?"
• "Show me my biggest savings opportunities"
• "What idle resources can I clean up?"
• "Which unused resources are costing me the most?"
• "Which Lambda functions have idle Provisioned Concurrency?"
• "Do I have any Lambda functions on deprecated runtimes?"
• "Which log groups have excessive retention I can reduce?"
• "Show me empty or orphaned CloudWatch log groups"
• "Are any of my MQ brokers over-provisioned?"
• "Show me waste findings with the CLI commands to fix them"

🧠 How the Agent Works

Waste Detection Integration

When you ask about waste or savings, the agent calls its findings tools and ranks results by savings and confidence, with the commands to fix them.

How the agent reads your data

The agent answers by calling read-only tools that return digests of your latest scan and cost data, such as a ranked cost breakdown or your top findings. It does not send the model a pre-filtered copy of your bill.

Memory

The agent can remember a small number of short preference notes you give it (for example "propose only, never execute") across visits. Ask it to forget them any time.

The model behind the agent

The agent runs on Amazon Bedrock (Anthropic Claude), server-side. The model context stays in CloudWise's infrastructure and Amazon Bedrock, and the agent builds its context from your real scan data via typed tools, never by sending your raw bill to a third party. The model selection is managed for you; you don't configure it.

What the agent checked

Each answer shows a short line listing what the agent checked (for example "checked your remediation activity"). It does not show the model name, token counts or a question classification.

🎓 Tips for Best Results

Ask Clear, Specific Questions

✅ Good: "Why did RDS costs increase 50% in October?"
❌ Too Vague: "Why are costs high?"

Provide Context When Needed

✅ Good: "Compare my Lambda costs in us-east-1 vs eu-west-1"
❌ Missing Context: "Compare Lambda"

Use Follow-Up Questions

The agent remembers the conversation for your current visit:

1. "Show me Cost Explorer costs"
2. "Why did it spike on August 15?" (AI knows you mean Cost Explorer)
3. "How can I reduce this?" (AI provides Cost Explorer optimization tips)

Build on earlier answers

Ask a broad question, then drill in with follow-ups in the same conversation:

1. "Show me my costs grouped by service"
2. "What are the main trends here?"
3. The agent can use your earlier question and answer as context

🔒 Privacy & Security

Your Data Protection

• ✅ Encrypted in transit - All AI queries use HTTPS/TLS
• ✅ No training on your data - Your cost data is never used to train AI models
• ✅ Secure storage - Conversations are not kept as a history: each question is sent to Amazon Bedrock to answer. Short notes you ask the agent to remember, and your daily question count, are stored in encrypted DynamoDB
• ✅ User isolation - Your data is never shared with other users

What the agent does NOT do

• ❌ Train on your data - Your cost data is never used to train models
• ❌ Share data with advertisers or use it for training - To answer, the Copilot sends your messages and the data its tools look up to Amazon Bedrock (an AWS service) in the United States
• ❌ Execute changes without approval - Remediation is propose-and-approve. Only the Agentic plan can execute fixes, and only after you explicitly approve each one (with rollback). Free and Shield are propose-only, and Compliance shows steps too: on all three you run the steps yourself.
• ❌ Touch IAM, security groups or your application data - The remediation role cannot change IAM, security groups or network ACLs, cannot create or delete VPCs or subnets, and cannot read your application data (S3 objects, database contents). Within its allow-list, and only after you approve, it can delete specific idle resources such as NAT gateways, VPC endpoints, volumes and snapshots (see the AI Remediation Guide for the full list)

📊 Performance & Efficiency

Speed

The agent reuses cached scan and cost data and prompt caching on the server.

Daily questions

Each question uses one of your daily questions regardless of its size. A turn that fails, or that the agent cannot finish because it is busy, is not counted.

🆘 Troubleshooting

Agent Not Responding

1. Check your daily limit - View remaining questions in the Workspace
2. Check your data - Ensure you have cost data for the period you're asking about
3. Try a simpler question - Start with basics like "What are my costs?"

Unexpected Responses

1. Review the disclosure line - The agent shows what it checked ("checked your remediation activity")
2. Rephrase your question - Be more specific or provide more context
3. Start fresh - Use + New if the context seems wrong

Daily Question Limit Reached

Free (5/day): Upgrade to Shield for 50 questions/day
Shield (50/day) · Agentic (100/day) · Compliance (500/day): Limit resets daily at midnight UTC

🎯 Best Practices

For Developers

• Use the agent to understand cost patterns during development
• Ask "What changed since my last deployment?" after releases
• Monitor specific services you're actively working on
• Ask the agent to scope a view, then dig into the trends

For FinOps Teams

• Start your day with "Show me yesterday's cost changes"
• Plan budgets with trends: "Show my spend trend over the last 90 days"
• Investigate anomalies: "Why did costs spike on [date]?"
• Track optimization impact: "Did costs decrease after rightsizing?"

For Business Leaders

• Monthly reviews: "Compare this month to last month"
• Budget planning: "What are our cost trends?"
• Executive summaries: "What are my top 3 cost concerns?"
• ROI analysis: "How much are we saving from optimizations?"

💡 Need Help?

Resources

• 📖 Quick Start Guide - Get started with CloudWise
• 💰 Optimization Guide - Cost-saving strategies
• ❓ FAQ - Common questions answered
• 🔧 Troubleshooting - Fix issues quickly

Contact Support

• ✉️ Email: support@cloudcostwise.io
• 💬 In the app, open Help in the navigation for FAQs and contact options, or ask the agent a product question

Upgrade Options

Ready to unlock more AI power?

• 🚀 View Pricing
• 📊 Compare Plans
• 🎁 Get Started for Free

---

🤖 Pro Tip: The more specific your questions, the better the agent can help. Instead of "Why are costs high?", try "Why did Lambda costs in us-east-1 jump between September and October?"

⚡ Quick Start: Open the Workspace now and ask: "What are my biggest cost opportunities?"