AI ROI Metrics are Coming, Even if They are Essentially "Soft" Measures of Impact

We might as well be honest and predict that enterprises are going to develop all sorts of metrics that purportedly show the positive impact of their artificial intelligence investments, but that the metrics will quite probably be proxies that measure all sorts of things other than direct AI impact.

Still, some common metrics are a starting point:

• Cost per unit of output — Does AI reduce the labor or compute cost to produce a document, resolve a ticket, process a claim, underwrite a loan?

• Throughput / cycle time — How many units processed per hour, or how much time shaved off a workflow (e.g., code review, contract drafting, customer onboarding)?

• Error rates and rework costs — Does AI reduce defect rates, compliance exceptions, or manual correction loops?

• Headcount avoidance — the ability to scale output without proportional headcount growth. Often measured as "FTE equivalents automated."

Revenue-side metrics are less common, but might include:

• Conversion lift — Does AI-personalized outreach or recommendation improve sales conversion rates?

• Revenue per sales rep — If AI handles pipeline qualification or proposal drafting, does rep productivity improve?

• Customer retention / churn reduction — Does AI-assisted support or proactive intervention improve net revenue retention?

• Time-to-market — Does AI-accelerated R&D or software development compress product cycles in ways that generate earlier revenue?

Other operational outcomes also sometimes are quantified:

• Accuracy or precision rates on specific tasks (e.g., document classification, anomaly detection in fraud)

• Audit findings or compliance exceptions reduced

• Model risk KPIs — false positive/negative rates in detection systems

• Employee time recaptured — hours per week freed from low-value tasks, redirected to higher-value work

• Employee satisfaction / retention — particularly in roles prone to burnout from repetitive work

• Decision quality — harder to measure, but some firms track downstream outcomes of AI-assisted decisions against historical baselines

As rational as all that sounds, the metrics are “soft.” The attribution problem is severe, as AI is almost never the sole variable changing in a deployment. 

AI might be deployed while other changes also are occurring:

• Process redesign — Most AI deployments force workflow reengineering. Efficiency gains may be 60% process change and 40% AI

• Training and change management — The same tool deployed with weak adoption programs vs. strong ones produces dramatically different outcomes

• Data quality improvements — Organizations often clean and structure data as a precondition to AI deployment; that alone drives gains

• Personnel changes — New hires, role restructuring, or management changes co-occur with AI rollouts

• Macroeconomic or market tailwinds — Revenue gains during an AI deployment may reflect market growth, not AI impact.

• Hawthorne effects — Measuring a team's performance changes behavior regardless of the tool.

The point is that it can be almost impossible to isolate the impact of AI cleanly. So most enterprise AI ROI figures are really "ROI of the initiative that included AI," not AI's marginal contribution.

The more interesting question might be "which specific processes have changed in ways we can measure, and do we understand why?"

Skeptics are correct to argue that attributing success purely to AI is often an oversimplification. But enterprises will have to try and do so, as investors will demand such “proof.”

So firms will supply such “proof” as best they can, even if the outcomes are not, strictly speaking, solely because of AI use. 

And that is not an unusual case. 

Research highlights that AI’s impact is heavily moderated by "complementary assets.” In other words, a firm’s  organizational structure, existing data quality and worker skill levels often do more to determine the outcome than the AI model itself.

Study Focus	Key Finding Regarding Attribution	Source
Productivity Paradox	AI adoption does not guarantee boosts; results are contingent on organizational structure and worker attributes.	Cho et al. (2026)
Social Penalty/Bias	Using AI for assistance causes observers to attribute success to the tool rather than the person, leading to negative competence assessments.	Reif (2025)
Supply Chain/Bias	In complex systems, responsibility is fragmented across vendors/platforms, making it nearly impossible to attribute specific outcomes to one source.	Sharma et al. (2026)
Task-based Impact	AI improves performance within its "capability frontier" but degrades it outside that range; attributing net gains requires granular task-level data.	Brynjolfsson et al. (2023)

The difficulty in quantifying the immediate return on investment for new technologies is a recurring theme in economic history.

During the 1970s and 1980s, despite massive corporate investment in information and communications technology, overall productivity growth in many industrialized nations remained stagnant. This led economists to question whether computers were truly providing the expected value.

Eventually, results were observed, but:

• Results lagged deployment: it took decades for firms to fully "reimagine" their organizational structures, business models, and workflows to leverage the new technology effective

• Value was indirect: better management, more efficient coordination or improved service quality, but correlation, not causation, remains a question. 

The measurable financial benefits of a transformative technology often became clear only after business processes were redesigned.

Technology	Scope of Impact	Key Findings	Source
ICT / General IT	U.S. Economy (1995–2000)	ICT accounted for 56% of labor productivity growth; added 1.18 percentage points to GDP growth.	Oliner & Sichel (2000)
Emerging Tech (AI/ML)	U.S. Public Firms (2009–2019)	Over a three-year period, “neither the mean nor the median abnormal ROE (expected performance) reaches statistical significance in the post-implementation period.” “The mean abnormal inventory turnover is −1.06, which is not significantly different from zero.” “Overall, our results…indicate no significant difference in performance between sample and control firms during the implementation period of emerging digital technologies.”	Li et al. (2024)
Internet / ICT	SME Growth (Global)	Web-savvy SMEs grew more than twice as fast as those with minimal web presence.	McKinsey (2011)

Still, in the meantime, we will see all sorts of metrics “demonstrating” AI impact. Enterprises making the investments have no choice but to try to do so, even if those metrics are “soft.”

U.S. Productivity is Rising, but AI Doesn't Seem the Reason

U.S. productivity has been rising for several years, but artificial intelligence is probably not the reason, at least, not yet. 

According to a report published by the Federal Reserve Bank of San Francisco, the U.S. economy expanded at a relatively steady pace of around 2.5 percent per year over the past three years, even though employment growth slowed to near zero. 

Almost by definition, higher output with the same input means higher productivity. But it is not clear artificial intelligence has much to do with the increases.

A survey of nearly 6,000 senior business executives in the United States, United Kingdom, Germany and Australia published by the National Bureau of Economic Research found:

• 69 percent of firms actively use AI

• 66 percent of executives regularly use AI

• Average use is about 1.5 hours a week

• 90 percent of executives report little own-firm impact of AI over the last three years

• 90 percent report no impact on employment or productivity

• Over the next three years, respondents predict that AI will boost productivity at their firms by an average of 1.4 percent

• Will raise output 0.8 percent

• Cut employment 0.7 percent

• Employees believe AI will raise employment 0.5 percent in the next three years.

Perhaps the most-unexpected result is the employee belief that AI will actually boost employment at their firms over a three-year period. That findings seems at odds with the usual press reports suggesting employee angst about AI impact on employment. 

The least-surprising result should be the inability to pinpoint AI productivity gains. 

For starters, U.S. productivity has recently been rising since about 2019, well before AI emerged as a potential driver. 

Labor productivity measures how efficiently workers use the capital available to them, such as equipment or software. The data suggests workers are doing so. 

Total factor productivity uses a broader view, measuring how efficiently the economy uses all inputs together, including both labor and capital.

One interpretation might be that workers have been using tools effectively, but that the gains have not yet shown up in TFP metrics. 

Think about your own work. Many of us would absolutely agree that AI has boosted our own personal productivity. But few of us can point to measurable gains in economic `outputs. 

Federal Reserve Bank 

And U.S. productivity had been rising since about 1992 as well, to 2000. 

Federal Reserve Bank 

For some observers, past experience suggests a productivity gain will happen. The U.S. economy has experienced several distinct productivity regimes over the past 70 years, including a high-growth period in the late 1990s, with the proliferation of computers and the internet, and a lengthy period of low average growth during the 2010s.

Federal Reserve Bank

Right now, it appears there is a significant disconnect between labor productivity and TFP. 

“The divergence between strong labor productivity growth and more modest TFP growth suggests that recent investments related to AI might be making workers more productive by providing them with better tools, such as new software and expanded computing capacity, but broader efficiency gains remain unrealized so far,” the report says.

But the report also says the pattern (Labor productivity and TFP misaligned) resembles what we saw when the internet became important. 

There was a lag then, and there is arguably a lag now. As the adage goes, one can see the impact everywhere but in the outcomes (paraphrasing the Solow Paradox: "You can see the computer age everywhere but in the productivity statistics.").      

J Curve and Solow Productivity Paradox are at Work with AI

Investors are going to keep challenging firms to show evidence their heavy artificial intelligence investments really are boosting productivity.

That is going to continue being a tough challenge, as history suggests the real output gains will take some time to develop.

So AI "productivity," or the "lack of quantifiable gains," are currently the most significant contemporary case of the Solow productivity paradox. 

In 1987, Nobel laureate Robert Solow famously remarked, "You can see the computer age everywhere but in the productivity statistics."

Recent research suggests productivity might actually decline for a time as firms deploy AI. 

The reason is the J curve. 

“We find causal evidence of J-curve-shaped returns, where short-term performance losses precede longer-term gains,” say economists Kristina McElheran; Mu-Jeung Yang; Zachary Kroff and Erik Brynjolfsson. “Consistent with costly adjustment taking place within core production processes, industrial AI use increases work-in-progress inventory, investment in industrial robots, and labor shedding,

while harming productivity and profitability in the short run.”

In other words, it takes time for enterprises to retool their business processes for the new technologies. And the more profound the innovations, perhaps the longer it takes to integrate those tools. 

Also, much of the reported AI adoption is horizontal rather than vertical; personal rather than systematic. In other words, individuals might be using chatbots, but workflows have yet to be transformed. 

So “personal productivity” has not yet been matched by an applied transformation of key work processes. And personal productivity gains are hard to measure, in terms of impact on firm performance. 

Agentic AI should help, as they can affect complex business processes. 

source: Forbe

Many have noted that  U.S. labor productivity significantly slowed in the 1970s and 1980s, despite rapid information technology investment. 

Then starting in the mid 1990s a decade of faster growth returned arguably because business process re-engineering had taken place.

A similar productivity paradox surrounds AI. As explained by economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson in a 2017 working paper, AI and the Modern Productivity Paradox,” the paradox is primarily due to the time lag between technology advances and their impact on the economy. 

While technologies may advance rapidly, humans and our institutions change slowly. 

Moreover, the more transformative the technologies, the longer it takes for them to be embraced by companies and industries across the economy.

Translating technological advances into productivity gains requires major transformations, and therefore time.

Today, we see a "Modern AI Paradox": while Large Language Models (LLMs) and Generative AI are ubiquitous in headlines and corporate pilots, global aggregate productivity growth  remains sluggish.

Economists like Erik Brynjolfsson argue that the paradox isn't a failure of the technology, but a timing and structural issue. He identifies four main reasons for this lag:

1. Mismeasurement: AI often improves quality, variety, or speed in ways that traditional GDP (which tracks "units produced") fails to capture.

2. Redistribution: AI may be used for "rent-seeking" (competing for market share) rather than increasing total industry output.

3. Implementation Lags: Significant "General Purpose Technologies" (like electricity or the steam engine) require decades of organizational restructuring before they move the needle.

4. Mismanagement: Companies often use AI to automate old processes rather than inventing new, more efficient business models.

Study	Target Group	Productivity Impact Found	Notes on Enterprise Deployment Gaps
MIT/Stanford (NBER)	Customer Support Agents	14% increase in issues resolved per hour.	High-skilled workers saw less gain; impact was greatest on novices. Enterprises often fail to use AI as a "leveler" for training.
Harvard/BCG (SSRN)	Management Consultants	40% higher quality; 25% faster task completion.	"Jagged Frontier": AI failed spectacularly on certain logic tasks where humans over-relied on it, leading to "falling off the cliff" errors.
Microsoft/GitHub	Software Developers	55% faster at completing coding tasks.	Gains are often eaten by "code bloat" and increased technical debt if not managed by senior architects.
Goldman Sachs Research	Aggregate US Economy	Projected 1.5% annual increase over 10 years.	Real-world adoption is currently hindered by power grid constraints and data center infrastructure delays.
NBER / Brynjolfsson et al.	Generative AI & the "J-Curve"	Initial 0% or negative impact.	The "Productivity J-Curve": Measured productivity dips initially as firms invest in "intangible capital" (retraining, restructuring) before the payoff.

While individual tasks show gains, enterprise-wide productivity often remains flat for several reasons:

• The "Pilot Trap": According to recent Adobe/Business research, 86 percent of IT leaders see potential, but only a fraction have moved beyond "isolated experiments" to organization-wide workflows

• Inertial Workflows: Companies often use AI to "do the old thing faster" (e.g., writing more emails) rather than "doing the right thing" (e.g., eliminating the need for those emails entirely). This results in "Digital Overload"

• The Human Bottleneck: AI can generate a report in seconds, but a human still takes hours to verify, edit, and approve it. Without changing the governance and approval structures, the AI speed gain is neutralized

• Data Fragmentation: Most AI models are effective only if they can access clean, centralized data. Most enterprises still have "siloed" data, leading to AI hallucinations or irrelevant outputs

• Skills Gap: Enterprises frequently treat AI as a "plug-and-play" tool like a calculator, failing to realize it requires a new type of "AI Literacy" to prompt and integrate effectively into complex projects.

None of that will be too comforting for suppliers who must justify their heavy AI capital investment. 

But history suggests the payoff is coming. It just will take some time. It always does.