Authors: Han Che, Luffy Chen

Across Foundry Agent Service and the broader ecosystem of agent frameworks and runtimes, it is becoming easier to build, deploy, evaluate, and optimize production agents. You can instrument your agent, collect traces, run evaluations, and use the agent optimizer to improve behavior over time.

But quality is only part of the production story.

A customer support agent may answer more questions correctly, but is it reducing support cost? A sales agent may handle more conversations, but is it creating more qualified opportunities? A claims agent may complete tasks faster, but is it improving business outcomes enough to justify the model, tool, and operational cost?

Customers today can answer whether an agent is working technically. They have traces, evaluation scores, latency, token usage, and logs. What is harder to answer is whether the agent is creating measurable business value.

Today we are introducing ROI for Agents in Foundry, available in private preview. ROI for Agents in Foundry helps customers connect agent quality, usage, and cost to business impact, so they can understand whether production agents are delivering value and where to focus next.

What is ROI for Agents in Foundry?

ROI for Agents in Foundry helps customers measure the business value, cost, and efficiency of their production agents based on customer-defined value models, configured assumptions, and observed system costs. It gives customers a way to move from operational and evaluation signals to a business outcome view.

At the center is a simple starting point:

ROI = (business value gained – total cost) / total cost

This gives teams a common way to ask:

• How much value did the agent create?


• How much did the agent cost to run?


• Is the agent improving business outcomes over time?


• Which workflows or tasks are driving the most value?


• Where is cost growing faster than value?

Every organization measures value differently. ROI in Foundry starts with a standard ROI model and helps teams connect business value assumptions, cost signals, and agent performance in one place.

Why ROI matters for production agents

Evaluations tell you whether an agent is behaving correctly. Traces tell you what happened. Metrics tell you how much the agent was used and how much it cost.

ROI in Foundry connects those signals to business impact.

For example:

• A support agent resolves a billing issue without escalation.


• A travel agent completes a booking change that would otherwise require human support.


• A field service agent reduces time spent on manual troubleshooting.


• A sales agent helps qualify leads faster.


• A claims agent improves completion rate while reducing review time.

Each of these scenarios has both a value side and a cost side. ROI in Foundry helps teams bring those together in one place.

How it works

ROI for agents is available in the Monitor tab within the Agents experience in Microsoft Foundry — the same place you already go to track agent traces and performance signals. The Monitor tab now surfaces not just how your agent is behaving, but whether it’s worth it — combining agent traces, business-value evaluation, and estimated operating cost into a unified view. That means customers can move beyond operational health metrics and evaluate whether an agent is generating more value than it costs to run. Let’s take a closer look.

A typical flow looks like this:

1. Connect traces. ROI in Foundry uses Application Insights trace data from the agent. Traces provide the conversation, evaluation, token, tool, cost, and version signals used for ROI calculation.


2. Choose a business-value evaluator. Select a built-in or custom evaluator that represents what “value” means for the agent, such as customer satisfaction, case deflection, or task completion.


3. Estimate business value. Conversations that pass the selected business-value evaluator are mapped to configured business value. Conversations that do not pass can still incur cost.

4. Measure operating cost. ROI in Foundry uses default cost estimates when available. Users can override cost settings when the default estimate is missing or not appropriate for the preview scenario.

5. Review ROI metrics. The Agent Monitor experience shows ROI KPIs such as Net Value, Business Value, Total Cost, and Current ROI, along with trend charts and version comparisons.

Note: Values shown are illustrative preview examples; Actual ROI depends on customer-defined value assumptions and actual operating costs.

6. Investigate low-ROI behavior. When ROI is low or negative, users can drill into traces to understand whether the issue is low business value, high model cost, expensive tool calls, or a specific agent version.

The goal is not to replace evaluations or observability. ROI in Foundry builds them. It adds a business lens to the post-deploy journey so customers can connect quality, cost, and business impact.

Built for real-world value models

ROI in Foundry is designed for real production scenarios where business value can vary by workflow, team, and industry.

In private preview, we are working with customers to understand how they want to model:

• How teams define business value gained for different agent scenarios.


• Which costs should be included in total cost, such as model tokens, tool calls, infrastructure, human review, or operational overhead.


• Whether ROI should be calculated per task, workflow, customer interaction, time period, or aggregate agent.


• Scenario-specific ROI models or templates.


• How ROI should connect to evaluation scores, task completion, customer satisfaction, and business process metrics.


• ROI workflows across UI, SDK/API, automation, export, and reporting.

These questions matter because every agent scenario is different. A support deflection agent, a sales assistant, a claims agent, and an internal operations agent may all create value in different ways.

Get started

ROI in Foundry is currently in private preview. We are looking for customers who have production or near-production agents and want to understand the business value, cost, and efficiency of those agents.

The best-fit preview participants are teams that can share:

• A production or near-production agent scenario.


• Traces or a plan to enable tracing.


• Evaluation signals or quality criteria.


• A business value model or an initial hypothesis for how the agent creates value.


• Feedback on UI and code-first ROI workflows.

If you are interested in joining the private preview, please complete the Private Preview Sign-up form. Our team will review all submissions and reach out to companies that meet the criteria and are a strong fit for the program. 

To see ROI for agents in action, watch our Microsoft Build session for an end-to-end observability walkthrough and live demos, including ROI in Foundry: From observability to ROI for AI agents on any framework.