Most AI business cases begin with a familiar promise: save time, reduce cost, do more with less.
Those outcomes matter. They are also incomplete.
If a firm measures AI only by the hours it removes, it may miss the larger return: serving more clients, responding faster, improving conversion, increasing capacity, strengthening retention, or creating a service the business could not deliver before.
That is the lesson inside this week’s strongest research. Departments that included growth outcomes in their AI scorecards were far more likely to report strong returns than departments that measured efficiency alone.
The study does not prove that adding a growth metric causes ROI to rise. It shows something more practical: the way an organization defines value is closely associated with the kind of value it reports—and likely influences the workflows it chooses to redesign.
At the same time, Gartner is questioning whether the enormous infrastructure build-out beneath enterprise AI can earn an adequate return. That risk will eventually reach buyers through vendor consolidation, changing prices, and harder procurement decisions.
The message at both levels is the same: AI cannot be justified by activity. It needs a return model broad enough to capture value and disciplined enough to survive scrutiny.
Here are the numbers—and what they mean for professionals putting AI to work.
The numbers worth knowing this week
80% of enterprise leaders reported positive AI ROI, but only 28% called the return strongly positive. Wharton Human-AI Research and GBK Collective surveyed 1,416 U.S. enterprise employees, including 1,007 managers and senior leaders. The high positive number shows that AI is producing perceived value; the much smaller strong-return group shows how rarely that value has become material. These are self-reported departmental assessments, not audited financial results. (Knowledge at Wharton)
Strong ROI was reported by 35% of departments measuring at least one growth outcome, compared with 13% of departments measuring efficiency alone. The study classified net-new revenue, sales conversion, output volume, and customer satisfaction as growth measures; efficiency measures included operating cost, hours saved, turnaround time, and error rates. This is an association, not a causal finding, but the nearly threefold difference is a strong reason to examine whether an efficiency-only scorecard is setting the ambition too low. (Knowledge at Wharton)
Only 7% of organizations succeeded across all three of Wharton–GBK’s markers: proving value, transforming workflows, and aligning the organization. Seventy-three percent of managers said workflows had changed significantly, yet only 28% described AI as fully embedded in core business processes. The gap matters: a tool can save time inside an old process without changing the outcome that process produces. Durable ROI requires the workflow, the people, and the measurement system to move together. (Knowledge at Wharton)
Gartner estimates AI data-center owners would need $8.5 trillion in cumulative revenue through 2030 to cover planned spending, financing costs, and a 12% return—against no more than $3.6 trillion in attributable enterprise and consumer AI revenue. Gartner says more than $3.5 trillion of AI infrastructure spending could be exposed to impairment by 2030 if the current capital-spending pace continues, and predicts more than 80% of neocloud providers will cease operating independently. That is a provider forecast, not an enterprise ROI calculation. For buyers, it signals future pricing, continuity, and concentration risk. (Gartner)
The pattern that matters
Efficiency is the floor, not the ceiling.
An efficiency metric asks whether AI made the existing work faster or less expensive. A growth metric asks whether the redesigned workflow produced more value.
Both belong in the business case.
The mistake is treating hours saved as if they automatically become financial return. An hour released from administrative work creates capacity. It becomes ROI only when the organization can show what happened to that capacity: more client conversations, faster delivery, additional volume, better retention, lower risk, or a cash cost that genuinely disappeared.
That is why every AI workflow needs a balanced scorecard with four layers:
- Efficiency: What is the full cost per accepted outcome, including technology, integration, verification, correction, and human time?
- Effectiveness: Did accuracy, quality, turnaround, compliance, or client experience improve?
- Growth: Did the workflow increase revenue, conversion, service capacity, output, retention, or access to a new opportunity?
- Payback: How long will it take the observed benefit—not the projected benefit—to recover the full cost?
The layers prevent two common accounting errors.
The first is declaring victory because a task became faster, even though the saved time was absorbed by checking, rework, or unused capacity. Last week’s report addressed that hidden verification cost.
The second is declaring failure because no headcount or software expense disappeared, even though the same team now serves more clients or produces better outcomes. That is the measurement gap this week’s research exposes.
Recent Gartner research on finance reinforces the need for a portfolio view. Fifty-five percent of CFOs reported positive overall returns from their 2025 AI initiatives, while 57% said the return on individual use cases was unclear. Straightforward applications such as data extraction, accounts-payable and receivable automation, and report creation typically reached expected value in nine to ten months. More complex work took longer. A six-month expectation applied to every use case can be as misleading as no deadline at all. (CFO Dive)
A good scorecard therefore does not make every project look successful. It makes the decision clearer: continue, redesign, scale, or stop.
The EMDUIT take — for the professional putting AI to work
Consider a brokerage using AI to improve lead follow-up.
The efficiency case is easy to see. AI can organize inquiries, draft personalized responses, suggest the next action, and reduce the time agents spend preparing routine communication.
If the brokerage stops there, the report may show hundreds of hours saved. But that number does not tell leadership whether the workflow improved the business.
A stronger scorecard starts with a baseline and follows the outcome through the work:
- How quickly did a new inquiry receive a useful response?
- What percentage of leads entered a real conversation?
- How many conversations became appointments?
- How many appointments became signed clients or closed transactions?
- Did response quality, compliance, and client satisfaction hold steady or improve?
- After technology and review costs, what did each accepted outcome cost?
Now the brokerage can distinguish capacity from growth.
If agents save time but contact and conversion rates remain flat, the workflow produced efficiency. Leadership can decide whether that capacity is useful enough on its own.
If faster, more consistent follow-up increases qualified conversations and appointments, the workflow may be contributing to growth. The team should still avoid claiming causation too quickly. Compare a defined period, control for lead source and seasonality where possible, and keep the same outcome definitions from baseline through review.
The point is not to force every AI use case to produce revenue. Risk reduction, quality, and service improvement can be valuable returns. The point is to name the intended outcome before implementation and measure the result the workflow was designed to create.
That discipline also protects the organization if the AI market becomes more expensive or concentrated. A company that knows its cost per accepted outcome, its contribution to growth, and its payback window can evaluate a vendor change. A company tracking only seats, tokens, or hours saved cannot.
Savings can prove that AI made the work cheaper. Growth proves that the redesigned work matters.
Which growth or effectiveness measure is missing from your AI scorecard?
If you are a licensed professional or business leader, choose one AI-supported workflow and add one outcome beyond time saved. Define the baseline, the owner, the measurement window, and the decision you will make when the evidence arrives.
Sources & methodology
Sources: Wharton Human-AI Research and GBK Collective, via Knowledge at Wharton (knowledge.wharton.upenn.edu) | Gartner (gartner.com) | CFO Dive, on Gartner's CFO AI research.