From promise to impact: How companies can measure—and realize—the full value of AI

Do you know how your AI program is performing?

A surprisingly large number of companies can’t answer that seemingly simple question with much confidence. Follow-on questions often lead to more awkward silences: Are our users adopting the solutions as intended? Is this paying off in terms of customer experience and bottom-line impact? Are we making progress toward a more competitive business model?

Being able to answer those questions is critical as companies turn to AI for productivity, growth, and competitive advantage. McKinsey’s latest Global Survey on AI found that nearly eight in ten organizations are using gen AI in at least one business function, and 62 percent are experimenting with agentic AI. Yet 60 percent of respondents still have not seen enterprise-wide EBIT impact from their AI programs. The gap between AI activity and AI impact only seems to be widening.

A key reason: Many organizations are deploying AI in ways that are more visible than valuable. Horizontal tools (for example, chatbots, copilots, and summarizers) improve the employee experience but are quickly becoming table stakes—solutions that help people work faster but rarely change a P&L.

By contrast, a smaller group of leaders is using AI to automate end-to-end workflows within specific domains (for example, claims processing, customer service, and demand planning). These deployments do more than accelerate work—they reshape it. Yet even in these higher-value plays, organizations struggle to prove impact. Teams disagree on what to measure and how to attribute improvements, leading to stalled scaling, budget skepticism, and business cases that are reopened instead of reinforced.

Our view is straightforward: AI impact is fully measurable, but it must be measured with the same rigor as any other capital investment. If leaders expect AI to materially alter their cost position or revenue trajectory, they need a system that ties technical performance to business outcomes with clear accountability and recurring proof.

Organizations that break out of the “pilot trap”—releasing endless AI pilots without ever scaling—tend to do three things differently:

• Define value up front and link relevant metrics, from technical performance, user adoption, operational change, through to financial impact.
• Build measurement and attribution into rollout (via A/B testing or staggered deployment) so results stand up to scrutiny.
• Run AI as a managed investment with a fixed review cadence, clear stage gates, and a single evidence pack that tracks both benefits and total cost of ownership—so only use cases that prove value advance to scale.

This article lays out that system: a five-layer framework that creates an auditable line from model performance to financial impact. It works across industries and applies to both gen AI and traditional machine learning and analytical AI. We then show how to operationalize the framework through a simple management cadence and a set of project phases, ensuring that only use cases demonstrating real, defensible impact advance to scale.

The five-layer framework: From model health to bottom-line results

The five-layer framework described below provides a structured approach to planning, measuring, and managing the value of AI investments. It creates clarity on who owns what, what to measure at each level, and how all layers connect to produce enterprise impact (table). Below, we examine each layer in reverse order—from Layer 5 (the basic technical infrastructure needed to support any AI initiative) through enabling capabilities such as user engagement, strategy, and key metrics, to Layer 1, bottom-line financial results.

Layer 5—Technical performance: Necessary, but not sufficient

Technical performance is the foundation of any AI system. Performance metrics indicate whether the model is operating as intended, staying within safety and cost guardrails, and maintaining quality over time. They are the “health stats” of AI: essential for reliable operation, but not sufficient on their own to demonstrate business value.

In practice, these stats typically show up in core system measures that vary by use case but follow common themes:

• Risk and safety: hallucination rates or instances of toxic or noncompliant outputs
• Cost efficiency: token spend per interaction or model call frequency
• Output quality and trust: percentage of responses accepted without significant edits
• Performance and reliability: response latency under load
• Stability over time: signs of model drift or degradation in output quality

These technical indicators are critical for keeping systems safe, reliable, and economically viable. However, they gain real meaning only when viewed alongside adoption and operational impact.

Layer 4—User adoption and engagement: The missing link in most deployments

User adoption and engagement show whether people are actually using and trusting AI in their daily work. Even highly capable models create little value if they are not consistently used in day-to-day workflows. In practice, adoption is one of the most common failure points in capturing value from AI efforts: without sustained engagement and trust, the downstream operational KPIs simply do not move.

In practice, this typically shows up in observable user behavior:

• Adoption and reach: number of daily active AI users, segmented by role or function
• Workflow penetration: percentage of eligible tasks completed with AI support
• Engagement depth: number of features adopted compared with those that are ignored
• Trust and reliance: AI acceptance rates compared with instances of overrides or substantial edits

When these measures improve, it signals that AI is becoming embedded in real work rather than remaining an occasional experiment. When they lag or vary widely across roles, they reveal trust gaps, usability issues, or training needs that must be addressed before value can scale. Tracking these patterns helps leaders boost enablement, product refinement, and change efforts where they will have the greatest impact.

Layer 3—Operational KPIs: The daily pulse of the business

Operational KPIs show whether the AI being used is improving how work actually gets done. This layer focuses on the process-level results AI is designed to improve—the measures that show whether tasks are getting faster, smoother, and more effective. If these indicators are not moving, it is a sign that AI may be active, but it is not yet changing how the business operates.

In practice, this typically shows up in metrics drawn directly from frontline systems, including:

• Speed and efficiency: shorter cycle times or lower cost per case, order, or ticket
• Quality and accuracy: reduced defect or rework rates
• Customer flow performance: lower abandonment in service journeys or higher first-contact resolution
• Retention signals: reduced churn within targeted segments

When these operational measures improve, organizations can see that AI is producing tangible, observable impact and not just generating activity or outputs.

Layer 2—Strategic outcomes and key results: Where strategy meets execution

Strategic outcomes show whether AI is moving the business in ways that matter to function and business leaders. These metrics sit closer to day-to-day performance than enterprise financials and, as a result, offer a level of granularity that broad indicators like EBIT often can’t provide. Tracking them helps organizations understand whether AI is improving customer experience, operational execution, or commercial effectiveness in ways that matter to specific functions.

In practice, this typically shows up in several types of outcomes:

• Customer experience: higher customer satisfaction scores as AI enables faster, more accurate service
• Operational performance: better on-time delivery or fewer unplanned equipment outages through AI-enabled forecasting and monitoring
• Service effectiveness: higher first-contact resolution as AI assists frontline teams
• Commercial results: sales uplift or improved customer retention driven by AI-informed targeting

These indicators are usually reviewed quarterly and tied explicitly to the organization’s strategic priorities, creating a bridge between daily operations and long-term value creation.

Layer 1—Financial impact: The enterprise outcome

Financial impact shows whether AI is delivering measurable enterprise value. That means translating technical and workflow improvements into clear, auditable business outcomes tied to the P&L and balance sheet. The most effective organizations define expected value before implementation begins and track results against a living business case.

In practice, this typically shows up in four places:

• Lower cost to serve: fewer human hours per support ticket as AI handles first-line resolution
• Revenue uplift: higher conversion rates or faster sales cycles driven by AI
• Margin expansion: engineering and operations teams delivering the same output with fewer hours
• Total cost of ownership: model usage costs and vendor or licensing fees

When benefits and costs sit in the same ledger, ROI withstands scrutiny and becomes a reliable input into budgeting and strategy discussions.

Rewired: Second edition

This updated edition offers brand-new insights into cutting-edge AI solutions—and what it takes to implement them—as well as the new economics of digital and AI transformations.

Learn more

How to make the framework real: The importance of governance

A measurement framework is only as effective as the governance that brings it to life. Even the best metrics lose impact if they are not embedded in how leaders review performance, make decisions, and allocate resources. High-performing organizations address this with a simple, disciplined structure: recurring routines and shared artifacts that keep AI value creation on track.

Importantly, this is not added bureaucracy; it is a way to reduce noise. Standard forums, shared definitions, and consistent evidence replace ad hoc meetings, one-off decks, and conflicting data, leading to clearer trade-offs and faster decisions.

In practice, this governance structure rests on two simple ingredients:

• a consistent monthly and quarterly cadence that creates accountability and maintains momentum
• a shared evidence pack—a single source of truth that anchors every discussion across the five layers, pulling together benefits, total cost of ownership, adoption metrics, and technical health

To make the cadence actionable, leading organizations also define decision gates: explicit checkpoints where projects must demonstrate progress on a small set of agreed metrics before receiving more funding, broader rollout, or additional engineering capacity. In other words, gates are how the organization turns measurement into decisions.

A typical set of gates might include: “Is the model safe and stable enough to put in front of users?” “Are users actually adopting it in a real workflow?” “Is there a measurable operational and financial impact that justifies scaling?” The evidence pack provides the facts; the cadence provides the rhythm; and the gates provide the decision logic.

The key gates: From idea to enterprise value

Having a clear review process and governance only helps if teams also know what they are trying to achieve at any given time. No AI project will move bottom-line metrics on day one, but all deployments need to be tracking toward value to receive continued investment. Typically, AI projects follow the same four main phases, similar to what we see in other tech projects.

Pilot phase: Execution begins with a tightly scoped pilot designed to prove technical and practical feasibility. A small group of users tests a prototype against clear technical KPIs, safety and cost guardrails, and early adoption signals (Layer 1 and 2 metrics). Teams also lock in how impact will be attributed (for example, A/B testing or staggered rollout) and evaluate whether the original operational value hypothesis still holds. The goal is not scale but proof: the solution must show credible technical performance and early user pull before moving forward.

MVP phase: This is where the solution enters real workflows, with limited but live exposure. Unlike the pilot, where metrics need to be generated and reviewed manually, measurement is now built directly into the AI tool. The system automatically tracks core technical indicators (such as response time and error rates), user behavior (who is using it and how often), and early signs of workflow impact (such as changes in cycle time or defect rates). Dashboards and automated reports replace data exports pulled from back-end systems. Human-in-the-loop safeguards are clearly defined, such as when AI outputs require review or approval, and ownership is assigned for monitoring performance, resolving issues, and escalating risks. Monthly reviews focus on whether the system works consistently in practice and whether usage and workflow impact are building as expected.

Initial scaling: This stage is when many initiatives either succeed or stall. With broader rollout, organizations can rigorously attribute impact and evaluate economics. Adoption should extend beyond early enthusiasts; operational improvements should be statistically significant; and financial benefits should at least offset the total cost of ownership. Technical performance must also hold under higher load, with monitoring and retraining processes functioning reliably. If ROI is not evident at this stage, it is a signal to refine or stop before full-scale investment (Layers 4 and early indicators on Layer 5).

Full scale: A deployment is considered scaled when AI shifts from an initiative to business as usual. The solution is embedded into standard workflows, governance, and budgeting cycles, with sustained adoption and structurally improved operational KPIs. Financial impact is reflected in plans, risk and compliance requirements are fully met, and long-term support and retraining are resourced. At this point, the full measurement framework becomes part of ongoing performance management, ensuring AI remains a durable source of enterprise value rather than a one-off project.

Taken together, these stages show how the five-layer framework moves from concept to operating model. Early phases focus on technical performance and adoption, establishing whether the system is safe, reliable, and used in real work. As deployments mature, measurement shifts toward operational impact, strategic outcomes, and ultimately, financial performance. The governance cadence and decision gates ensure that progress up the pyramid is evidence based: projects advance only when signals at one layer credibly support moving to the next. In this way, the framework does more than track AI performance. It creates a structured path from experimentation to durable enterprise value.

The discipline that separates leaders from followers

The next phase of AI adoption won’t be won by those who experiment the most, but by those who can turn experimentation into measurable, repeatable performance. As gen AI moves from pilots to the core of how work gets done, leaders will need more than powerful models—they’ll need a management system that can distinguish real impact from noise and enables them to scale only what proves its value. The organizations that embrace this discipline will move faster, spend smarter, and build a deeper conviction in where AI truly creates advantage. They’ll also be better positioned for the next generation of AI capabilities, where the gaps between early movers and everyone else will widen quickly.

AI’s promise is no longer in question. The differentiator now is proof, and the ability to turn that proof into sustained performance. The companies that build this muscle today will define the benchmark for value in the years ahead.