AI Savings Estimation#

When Reply CMP surfaces cost optimisation recommendations (from Azure Advisor, AWS Trusted Advisor, and GCP Recommender), it enriches them with AI-computed savings estimates. These estimates go beyond native recommendation figures by computing what you would actually save based on your real cost history.

How Estimation Works — the ReAct Agent#

Savings estimates are produced by a ReAct-pattern AI agent that runs at recommendation refresh time. The agent uses a 3-step reasoning loop:

Step 1 — get_cost_data#

The agent queries the cost database for the affected resources (identified in the recommendation) over the past 30 days of actual billing data. For each resource, it retrieves actual daily costs.

Step 2 — validate_estimate#

Using the retrieved cost data, the agent calculates a realistic savings figure based on:

  • What the recommendation is (e.g., rightsize a VM, delete unused disk, switch to Reserved Instance)

  • The actual cost of the affected resources

  • The expected cost reduction the recommendation would produce

The agent runs a validation check: does the proposed estimate exceed the actual cost of the resource? If yes, it is capped — you cannot save more than the resource costs.

Step 3 — finalize#

The agent produces the final estimate and assigns a confidence rating based on how much real cost data was available.

The loop runs up to 3 iterations, refining the estimate if the initial validation reveals inconsistencies.

Confidence Ratings Explained#

Rating

Condition

What it means for you

High

Cost data available for 100% of affected resources

Use for business cases, ROI calculations, and executive reporting

Medium

Cost data available for < 50% of affected resources

Directional estimate — good for prioritisation, not precise forecasting

Low

No cost data available for any affected resource

Heuristic estimate based on recommendation type only. Treat as “there’s something here”, not a precise figure

Note

Low confidence does not mean the recommendation is wrong — it means the savings estimate is uncertain. The underlying recommendation from Azure Advisor, AWS Trusted Advisor, or GCP Recommender may still be valid.

The Cap Guarantee#

Savings estimates are always capped at the actual cost of the affected resources. The AI cannot produce an estimate that exceeds what the resource actually costs — this prevents misleading “save more than you spend” figures.

Example: If a recommendation says “switch this VM to a smaller SKU” and the VM costs €50/month, the maximum estimate the system will produce is €50/month (less the cost of the new SKU).

Plain-Language Explanations#

Each recommendation in the Optimize tab includes an AI-generated plain-language explanation of the savings opportunity. This explanation is:

  • Written for a non-technical audience

  • Does not include specific dollar amounts (to avoid misleading figures from Low-confidence estimates)

  • Does not reference internal resource IDs

  • Suitable for copying into a stakeholder presentation or email

Dev/Test Savings — Separate Calculation#

The Dev/Test Savings figure shown on the Optimize Overview card uses a separate, deterministic calculation — not the ReAct agent:

  • Identifies resources in groups tagged as Dev/Test environments

  • Identifies resources with automatable types (VMs, VMSS, etc.)

  • Calculates 50% of their last-30-day cost as the potential savings from automated shutdown

This is a heuristic assumption (50% saving from ~12h/day shutdown ≈ 50% compute cost), not an AI estimate. No confidence rating applies.

Refreshing Estimates#

Click Refresh on the Optimize tab to trigger a new estimation pass. The ReAct agent re-runs for all active recommendations using the latest cost data. Refresh takes 30–120 seconds depending on the number of recommendations and connections.

Tip

Refresh after the first week of each month when fresh billing data has settled. Estimates improve as more monthly cost data accumulates.

Comparison to Native Recommendation Tools#

Aspect

Azure Advisor / AWS Trusted Advisor / GCP Recommender

Reply CMP AI Savings Estimation

Source

Native cloud provider

Reply CMP AI (based on your actual bill)

Savings figure

Estimated by provider algorithms

Based on your real historical costs

Cross-provider

No (per-provider only)

Yes — unified view, single currency

Confidence rating

No

Yes (High / Medium / Low)

Plain-language explanation

Technical / abbreviated

Non-technical, stakeholder-ready

Capped to actual cost

Not always

Always