Is AI ROI Real? Read the Surveys by Who Paid for Them

Ask whether AI delivers a return on investment and you can find a credible-looking survey to support any answer you like. One says 88% of companies see revenue gains. Another says 56% of chief executives have seen no financial benefit at all. A third says only 14% of finance chiefs see meaningful returns today.

These are not fringe studies. They come from NVIDIA, PwC, RSM, RGP and MIT. They cannot all be describing the same reality, so the useful question is not "what does the research say about AI ROI?" It is "who paid for each survey, who did they ask, and what did they count as a return?"

I run engineering at Chronexa, an AI implementation firm, so I am not neutral either. Our business depends on AI working. That is exactly why I think buyers deserve a straight reading of the evidence, including the parts that do not flatter companies like mine.

Here's what matters most:

• The headline AI ROI numbers disagree wildly, from 88% of companies reporting revenue gains (NVIDIA) to 56% of CEOs reporting neither revenue nor cost benefits (PwC).
• The gap is mostly explained by three things: who sponsored the survey, who answered it, and whether "return" meant satisfaction or a measurable profit.
• The neutral data is sobering. The US Census Bureau found about one in five US businesses used AI as of May 2026.
• Real AI ROI shows up in countable, high-volume work where you can measure cost per unit before and after.
• The only AI ROI number that matters is your own, measured against a baseline you recorded before the project started.

Is AI ROI real? What the major 2026 surveys say

Here are the most-cited recent studies, with the sponsor and sample next to each figure.

NVIDIA, "State of AI" 2026. NVIDIA surveyed more than 3,200 organizations in financial services, retail, healthcare, telecom and manufacturing. It reported that 88% saw measurable annual revenue increases and 87% saw AI-driven cost reductions. NVIDIA sells the chips that most AI runs on.

RSM, Middle Market AI Survey 2026. RSM surveyed 1,030 managers and executives at US and Canadian companies with $30 million to $10 billion in revenue, in March 2026. It found 97% satisfied with their AI investments and 54% saying AI had beaten their ROI expectations. Every respondent already used AI. RSM sells AI consulting.

PwC, 29th Annual Global CEO Survey. PwC surveyed 4,454 chief executives in 95 countries and territories between September and November 2025. It found 56% reported neither increased revenue nor lower costs from AI in the previous 12 months. About 30% reported higher revenue, 26% lower costs, and only 12% both.

RGP, 2026 CFO research. RGP surveyed 200 senior US finance executives at companies with $500 million to over $10 billion in revenue, in October and November 2025. It found 66% expect measurable AI impact within two years, but only 14% see meaningful AI ROI today.

MIT NANDA, "The GenAI Divide: State of AI in Business 2025." Based on interviews, a survey of leaders and a review of more than 300 public AI deployments, it found about 95% of generative AI pilots produced no measurable profit-and-loss impact. It defined success narrowly (measurable financial impact within six months of deployment) and used a modest sample, which critics rightly point out.

Line them up and the pattern is almost comic. The chip seller finds the highest returns. The consulting firm, surveying its own buyers, finds near-universal satisfaction. The CEOs and CFOs, who answer to shareholders for actual profit, find the lowest.

Why do the AI ROI numbers disagree so much?

I do not think any of these organizations is lying. I think three ordinary things explain nearly all of the gap.

1. Who answered. A technology manager who championed an AI rollout will rate it differently from a CFO who has to find it in the income statement. RSM's respondents were mid-level managers and above with influence over technology decisions. PwC asked CEOs. RGP asked CFOs. The further from the project and the closer to the profit line, the lower the reported return.

2. What counted as a return. "Satisfied with our AI investment" and "AI increased our revenue in the last 12 months" are completely different questions. Satisfaction can come from faster drafting, happier staff or simply having started. Measurable return needs a before-and-after number someone actually tracked. RSM's own framing acknowledges this; achieving satisfaction is not the same as making a significant impact.

3. Who was in the sample. Surveys of companies that already use AI exclude the ones that tried and stopped. Every RSM respondent was an AI user. NVIDIA's respondents were organizations engaged enough with AI to answer an AI survey. Neither design captures the abandoned projects.

And yes, sponsor interest matters at the margin. A company that sells chips, software or consulting has every reason to publish the encouraging numbers prominently. To RSM's credit, its own report includes a figure that cuts against consultants: among companies increasing AI spending, 40% are shifting budget away from external consulting services. That is the kind of detail that makes a survey more believable, not less.

What does neutral data say about AI adoption?

When I want a number with no product attached, I look at government data. The US Census Bureau runs the Business Trends and Outlook Survey every two weeks across a broad sample of US employers. As of May 3, 2026, 19.8% of US businesses reported using AI to produce goods or services, a figure that stayed between 17% and 20% from December 2025 through May 2026. Use was far higher among larger employers: 37% of firms with 250 or more employees and 32% of firms with 100 to 249 employees.

Two things follow from that. First, despite the noise, most US businesses are not using AI in operations yet, so any survey of "AI users" is a survey of the leading fifth. Second, adoption is concentrated in firms large enough to have someone whose job includes making it work. That is a clue about where returns come from.

Where does AI deliver a real return?

Across the more careful studies and our own projects, AI ROI shows up most reliably when four conditions are true.

The work is high-volume and repetitive. Reading invoices, sorting client documents, extracting fields from forms, drafting routine correspondence, answering the same phone questions. Volume turns a small saving per item into a large total.

The unit is countable. You can count documents processed, hours of review saved or calls answered, and put a cost on each. When the unit is countable, the before-and-after is undeniable.

The process is embedded, not bolted on. PwC found that CEOs reporting both revenue and cost gains were more likely to have applied AI widely across their business, and that companies with strong AI foundations were about three times more likely to report meaningful financial returns. A tool sitting beside the real workflow rarely pays; a system that is the workflow often does.

People stay in the loop. The returns hold up when AI does the reading, chasing and first drafts and your team keeps the review and judgment. The returns collapse when errors reach clients and someone has to clean up.

The opposite is just as instructive. In our own business, the AI systems with the clearest return are the ones that process countable work. The weakest return came from an outbound sales engine we built for ourselves. The system worked exactly as designed, but when it wrote a different pitch for every prospect we got roughly one reply in 560 emails, and a later campaign reached 895 emails with zero replies. The software was not the problem. The outcome depended on our offer and our market, which no software can fix. Any AI project whose return depends mostly on strategy will show a weak ROI if the strategy is weak.

How do you measure AI ROI in your own firm?

Survey numbers are interesting. Your own number is the only one that should drive a decision. Here is the method I recommend, and the one we use when scoping work.

1. Record the baseline before you start. For the chosen process, write down today's volume, the staff hours it takes, the error or rework rate and how long it takes from start to finish. Without a baseline, every result becomes a matter of opinion. This is the step almost everyone skips.

2. Measure cost per unit, not total savings. Calculate what one unit of work costs today (one invoice, one client file, one intake) and what it costs with the AI system, including the AI usage bill, the software, the maintenance fee and, most often forgotten, the staff time spent reviewing the AI's work. Usage costs can grow faster than expected, as we explain in why AI is so expensive even as token prices fall.

3. Count capacity, not just cost. In professional firms the return usually shows up as capacity: the same team handling more clients in busy season, or partners spending less time on administration. Put a value on the extra capacity you would otherwise have had to hire for.

4. Check at 90 days, then again at six months. The first weeks overstate returns (novelty and attention) and understate them (staff still learning). A second checkpoint catches drift.

5. Decide what failure looks like in advance. If cost per unit does not fall below a set threshold by the second checkpoint, stop or change course. We cover how to write these stop rules in our guide to whether you should pay for an AI proof of concept.

For a quick first estimate on a specific process, the AI ROI calculator gives a payback figure in about two minutes without asking for your email.

Does data security affect AI ROI?

Yes, and mid-sized firms tend to underweight it until a client asks the question. A single incident (client documents sent to a consumer AI tool, an output with someone else's data in it, a missing record of who approved what) can wipe out a year of efficiency gains in lost trust and remediation.

Build the controls into the return calculation from the start: run AI through an enterprise agreement that excludes training on your data, keep it inside your own cloud account where possible, log every action the system takes and every human approval, and give one person ownership of both the cost and the risk. These controls cost a little up front and protect the return you are trying to measure.

Frequently Asked Questions

What is the ROI of AI for businesses?

It depends heavily on the work and on how return is measured. Vendor surveys report very high returns, with NVIDIA citing 88% of organizations seeing revenue increases, while PwC found 56% of CEOs saw neither revenue nor cost benefits in the past year. The most reliable returns come from high-volume, countable work measured against a baseline.

Does AI have a good ROI?

For specific, high-volume processes with a human review step, often yes. As a broad, unmeasured rollout, often not. RGP found only 14% of US CFOs see meaningful AI ROI today, while 66% expect measurable impact within two years, which suggests most firms have not yet set up the measurement to prove it.

Why aren't most businesses seeing ROI from AI yet?

Mainly because pilots are not connected to real workflows and nobody recorded a baseline. Firms in RSM's 2026 survey whose pilots underperformed most often blamed data quality and integration. PwC also found companies with strong AI foundations were about three times more likely to report meaningful financial returns.

How do you calculate ROI on an AI project?

Record the current cost per unit of work before the project, then measure the new cost per unit including AI usage, software, maintenance and staff review time. Add the value of extra capacity freed up, and compare at 90 days and six months against a threshold you set in advance.

The bottom line

AI ROI is real, but not evenly distributed, and not proven by any single survey. Read every study by who paid for it, who answered it and what it counted. Then ignore all of them for your own decision and measure your own process, before and after, per unit, with the review time and the AI bill included.

If you want help picking the process where the return is most likely and setting up the baseline, book a 30-minute call. For industry-specific ranges, see our AI automation ROI benchmarks for law, accounting and financial services, and for what different budgets buy, our guide to AI automation agency pricing.