5 metrics to drive successful AI outcomes

Despite massive investments in AI last year, most businesses have struggled to achieve measurable results. In PwC’s January 2026 Global CEO Survey, 56% of CEOs reported neither increased revenue nor decreased costs from AI in the last 12 months — a statistic backed by Gartner figures that show only 5% of CFOs have reported cost reductions due to AI, and only 6% have reported revenue increases. So what’s behind those dismal numbers?

“They’re really laser-focused on measuring the wrong things,” says Shamim Mohammad, executive VP and CITO at CarMax. “There’s a fundamental misunderstanding of how to measure AI.”

Companies also tend to focus on the wrong business goals, such as boosting employee productivity without fully capturing business outcomes. An AI that saves two hours of time per software developer means nothing unless you can tell the CFO what they did with that time, and that can be hard to do. Also, some corporations have underestimated the level of organizational change management required to achieve efficiency gains. And the metrics to assess performance, such as what percentage of any gains are directly attributable to AI and predicting the ongoing costs of running AI projects at scale, can be difficult to nail down.

Some CEOs who thought AI productivity gains might demonstrate value by reducing headcount or new hires have also come away disappointed. Organizations need at least a 50 to 70% productivity gain before headcount can be reduced, according to Gartner. “We’re not seeing many use cases this high,” says Nate Suda, the research firm’s senior director analyst, with most falling between no gain at all to about 30%.

But some organizations are achieving results by aligning with strategic business outcomes, taking a long view on ROI, and focusing on critical business goals that may or may not fall inside the purview of traditional financial metrics. So here are five key benchmarks to apply to, or check against, your AI pursuits.

Choose the right projects and align with strategic business goals

Before considering success metrics and milestones, make sure you have use cases aligned with business goals, and you’ve prioritized the most strategically important business outcomes. The business cases for AI should come from the top down, says Keith Sarbaugh, EVP and chief digital and technology officer at animal health firm Zoetis. “Working with the board, we prioritized use cases in R&D and commercial business areas,” he says.

Also, benchmark against strategic business outcomes, and benchmark spend against best-in-class organizations in your industry, adds Zar Toolan, former chief data and AI officer at Edward Jones, and current executive-in-residence at Practitioners for Practitioners, a network of expert AI and data advisors. Based on his research of category-leading organizations in the financial services sector, for example, data and AI consumes 15 to 20% of overall IT budgets, which itself constitutes 10 to12% of topline revenue for the business. “While these numbers will vary by organization, it’s important to know where and how much your key competitors are investing and divesting,” he says.

Non-tech leaders are often the ones deciding on AI spend, while tech leaders are responsible for implementation, Suda says, which necessitates a shared understanding of value expectations and timescales.

That’s what’s happening at CarMax. “The business owners sign up for the benefits, and the CFO and I review the use cases,” Mohammad says. “The cost comes out of the budget for that business unit, and accountability lies in the hands of the business unit head.”

When Zoetis decided to pilot an AI project to make the customer service team more productive, it measured time spent on each call, and the time required to generate a follow up. The former stayed constant while the latter improved dramatically due to AI-generated follow up recommendations. The quality of the experience improved, and the team was able to reach more customers, which was the primary success metric. Zoetis also deployed Microsoft Copilot and additional AI capabilities to democratize AI, but is now focused on bigger transformative things in the areas of supply chain and contract lifecycle management.

So having an overall AI strategy aligned with your organization’s business strategy is crucial. At Edward Jones, the process of setting up its data and AI strategy stretched over two years, says Toolan. The company armed advisors with gen AI tools trained against the firm’s own data, and developed advice agents to help its financial advisors quickly create plans for client portfolios using inputs from meeting notes and data in the CRM system.

“It takes internal research, matches that with the client’s plan, and gives recommendations about how to implement the plan for the client,” he says. Advisor teams saved about one person day per week and could get back to clients faster. “They could go deeper with clients, had better client satisfaction scores, and client outcomes were better aligned with what was most important to their goals,” he adds.

AI that improves customer service can also have a direct impact on the P&L sheet. For example, a reduction in customer churn avoids revenue losses, while an increase in conversions leads to higher sales. The trick lies in measuring the change in revenues and determining what percentage of any impacts are directly related to AI.

Still, other projects may have strong financial benefits that are less visible. For example, using AI to improve capital use through better sales and AP forecasting can substantially improve the working capital cycle and average revolving debt utilization, effectively increasing liquidity. “As you increase accuracy, the amount of working capital or revolving debt interest payments decreases.

“Both can increase free cash flow and therefore are often considered functionally the same as cost savings, even though they show up in different places,” Suda says, “like working capital reduction in cash flow, and revolving debt interest expense reduction on the income statement/P&L.”

For some AI projects the goal may be simply keeping ahead of users already familiar with tools like ChatGPT. There’s the risk that if you don’t provide basic gen AI tools in the enterprise, employees will push company data into publicly available tools for analysis. That risk may be a larger one than having your own Copilot or GPT solutions.

But gen AI is becoming a must-have part of every office worker’s toolset. “How do you measure the value of Excel in an organization,” asks Afshean Talasaz, former SVP and CTO at Colonial Pipeline. “Everyone uses it, but no one asks if you need it anymore. We’re not there yet with gen AI, but it’s certainly a possible outcome as ways of working, costs, and expectations around AI change in the enterprise.” 

Know your true costs

When projecting ROI, the investment or cost denominator can be tricky to nail down. Unlike traditional IT projects, the costs to build an AI are smaller than the ongoing run costs, especially at scale. “AI flips the script, and you need to plan for that,” says Suda. “For CIOs, this isn’t a cost dynamic they’re used to.”

That’s especially problematic because the stakes can be so high as the cost of both individual AI projects and the underlying data, infrastructure, and governance work that must be completed to support those efforts heighten risks of failure. “I’ve been in technology for 25 years and this is the most expensive and complex thing I’ve ever worked on,” says Sarbaugh.

But not all projects are expensive, says Talasaz. “I’ve seen projects that don’t cost a lot and I’ve seen others in the millions.” The cost of a given AI project depends in part on if an organization has internal staff that can build and develop AI projects or rely on external partners; if it’s already addressed data and infrastructure requirements or if it’s fully digital and in the cloud.  

How you measure success matters when it comes to ROI, says Talasaz. Most projects sit on top of common platforms that can be costly to set up, and how those costs are allocated affects the economics of AI projects. That’s why, he says, project economics matter, but it can be challenging to get the full picture depending on the project. There’s a degree of innovation and uncertainty in those that can take on characteristics of risk capital projects, similar to drug discovery in pharma, or oil and gas exploration. “Articulating financial metrics for AI projects is more effective at a portfolio level, where data and AI infrastructure investments, platforming, scalability, expectations, time to adoption uncertainties, and interconnectedness to other use cases can be more effectively measured,” he says.

The FinOps Foundation’s FinOps for AI working group is hashing out how to measure the value of AI projects at a more granular level since 98% of enterprises are managing AI spends, up from 31% two year ago, according to the Foundation’s recent State of FinOps Survey.

While the financial reporting indicators are the same, the data needed to compile those metrics differ, which can be difficult to obtain. “With AI, success is measured in tokens instead of minutes, or API calls instead of usage hours and minutes,” says Rob Martin, FinOps fellow at the Foundation. “And, unfortunately, a lot of the cost and usage data doesn’t come with the billing data.” So this leaves organizations with no choice but to collect it through API proxies or by using third-party products.

But models can’t always tell you how many tokens were used. “AI costs and service availability are more volatile, so you may not be able to predict the costs of operating your model at scale,” Martin adds. That’s forced many IT executives to forecast costs in ranges rather than using fixed numbers. So he recommends forming an AI investment council or tiger team chaired by the CIO that includes all stakeholders involved in AI spending, and an experienced executive who can help create a FinOps scope for spending, incrementally fund competing projects, develop a plan to gather and analyze data, and translate performance metrics to the board.

Prioritize for success 

Of the 10 AI value metrics Gartner has identified, productivity is on the lowest end of the spectrum. Projects that improve capital use, reduce losses, improve customer experience, or help with new products have a much greater potential for revenue growth.

Then operational metrics like net promote score, output per hour, and backlog reduction are all about return on employees, while financial measures such as average labor cost per worker, sales and conversion rates, and losses avoided focus on ROI. Leading indicators like traffic and clicks, and market share growth are all about what Gartner calls return on future.

For Toolan, there are certain AI scorecard benchmarks that every business should follow, tailored to each set of circumstances. Evaluate the AI strategy, technology stack, or platform, assess AI readiness relating to mindsets and skillsets of employees adopting the technology, and measure the results against desired business outcomes. “These need to be very industry- and sector-specific, and tailored to where your organization is on its AI journey,” he says.

But traditional financial metrics aren’t always the most important measures for desired business outcomes. For example, the business goal may be to improve customer satisfaction through measures such as NPS surveys. But they have limits. “You have to rely on survey answers, and it’s tempting to overstate the value of the improvement,” says Sarbaugh.

At CarMax, the most important metrics for its two major AI projects include customer satisfaction, conversions, and sales. The online auto retailer built its Skye AI assistant to help customers navigate the buying process; and Rhodes, a virtual agent, guides call center associates by providing questions to ask customers, quickly surfacing needed documents, and helping them navigate state rules, regulations, and policies.

“Time savings is a good metric to have, but what we look at are the business outcomes it’s driving,” says Mohammad. “What are we doing with the time that’s freed up to create more value for the company.”

NPS scores show a better customer experience, while the combination of Skye and Rhodes has generated higher conversion rates and elevated sales. “It’s about the elevation of the associates’ role from an information retriever to a trusted advisor,” he adds. “Roles are converging, what people do is getting redefined and processes are getting reengineered.”

Prepare your users and track acceptance rates

Even the best laid plans for AI will fail if users won’t adopt it, so user acceptance is yet another critical metric that requires advanced planning. It’s not enough to just train the AI models — you have to train people how to use AI and personalize it to their role in the organization. “It has to be a great experience in their flow of work,” Toolan says. “If it’s not, you’ll see a lower sentiment score and poor AI adoption.”

At first, says Sarbaugh, Zoetis underestimated the investment needed to make in change management since early on it was about how to move quickly. “We assumed if we built it they would come, but you won’t see value if users don’t understand how to incorporate it into their day-to-day work,” he says.

Think long-term value — and trust your judgement

At CarMax, the AI project metrics conveyed to top management fall into four areas: ROI, revenue impact, cost efficiency, and strategic value or the capability to gain a competitive advantage.

When presenting the business value of AI projects, include a value exposure measure, Talasaz says. Each AI use case may require changes to data, processes, and platforms or create new capabilities that may enable other projects to leverage those investments. “It opens the door for other things,” he says. “It’s usually a qualitative element as you may not know the costs and benefits for those other projects in detail yet, but there’s value in knowing that the current project is strategically opening up other capabilities.” And if a project will be a competitive differentiator, you need success measures for that as well.

When pitching a project, Talasaz adds he sets expectations transparently. “It’s important to provide scenarios, not just one financial metric, so provide a range of outcomes and the underlying assumptions and factors that impact those outcomes,” he says. For example, he’ll paint a picture that includes expected outcomes based on different adoption or implementation rates. “Leaders want to understand how all of the dots connect and what the bigger picture is,” he continues.

Ultimately the success metrics that matter are those that show the impact on P&L, and how the AI use case affects the employee experience. “Any chance you get to put a quantitative metric at any step in the process, do it,” says Sarbaugh. “But with the newness of AI, and being in the early stage of adoption, there’s an element to things that can’t be measured at the moment. So there’ll be times when you have to trust the judgments of the people who understand the business. Just use common sense and make sure you don’t overstate the results.”