Key Takeaways
- Only 28 percent see a return: More than half of all corporate AI projects are already live, yet the clear majority never produce a measurable business result.
- Budgets keep growing anyway: 86 percent of executives are increasing their AI investments, but only 21 percent use AI inside their core processes.
- The skills gap is the main brake: Missing know-how, not missing technology, is the single biggest obstacle to successful adoption.
- Time savings stay small: 44 percent of users save less than four hours per week, and only 44 percent receive structured training.
- Generative AI rarely pays on its own: Just 29 percent of companies report a significant return from generative AI alone.
- Agentic AI is still mostly a promise: Only 23 percent of agent projects deliver a demonstrable financial benefit today.
Introduction
Only 28 percent of AI projects deliver a measurable return on investment. More than half of all companies are already working with artificial intelligence in production, and still the tangible value is missing in most cases. That gap is the most uncomfortable number in enterprise technology right now.
Why does digital change fail so often, even when the technology works? Six studies published in recent weeks point in the same direction. The bottleneck is almost never the model. It is process design, leadership attention and the skills of the people who are supposed to use the tools every day.
My working principle for midsize companies is simple: relevance beats gadgets. This article walks through what the numbers actually say, why more budget does not automatically create more value, and which levers a management team can pull in the next quarter.
The 28 Percent Problem
Gartner data reported by Olakai puts the number bluntly. Only 28 percent of companies reach the return they hoped for. At the same time 52 percent of AI projects are live in some form. In other words, most organisations have successfully shipped something and still cannot point to a line in the profit and loss statement that moved.
The same data set contains a second observation that deserves attention in every leadership meeting: in 80 percent of the companies running autonomous AI pilots, headcount reductions followed. Whatever your position on that, it means AI projects are no longer a quiet IT topic. They touch the workforce, and they need to be communicated as such.
A measurable return needs a baseline. If nobody wrote down how long a process took before the pilot started, no dashboard will prove an improvement afterwards. That sounds trivial. In practice it is the most common reason a successful pilot cannot be defended in a budget review.
Why More Budget Does Not Equal More Value
Accenture reports that 86 percent of executives are raising their AI investments and 78 percent name revenue growth as the main goal. Only 21 percent, however, use AI inside their core processes. That is the mismatch in one sentence: the money goes into experiments at the edge of the business while the expectation sits at the centre of it.
Edge experiments are useful for learning. They are a poor source of return, because the volume is too small. A summarisation assistant used by four people in marketing cannot move a company result, no matter how good it is. The processes that carry real volume are order intake, service requests, quoting, invoicing and reporting. Those are also the processes with the most rules, the most exceptions and the most internal resistance, which is exactly why they get postponed.
McKinsey adds the counterweight. Only 39 percent of companies see an EBIT contribution from AI, but mature organisations achieve a 5.8 times return within roughly 14 months of going into production. The spread between the two groups is not explained by better models. Everyone buys from the same handful of providers. It is explained by how deeply the technology is wired into daily work.
The Skills Gap Is a Leadership Problem
Deloitte identifies the AI skills gap as the largest single obstacle. Most companies respond with education and training, which is the right instinct. What most of them do not do is rethink the workflow itself. Training people to use a chat window faster does not change a process that was designed for paper forms in 1998.
There is a useful distinction here. Tool training teaches which button to press. Process training teaches which decisions can now be made differently. The second kind is harder to organise and it is the one that produces a return. It also requires managers who understand the tool well enough to redesign the work, which is why AI literacy at the leadership level matters more than another licence for the team.
Productivity Gains Are Lagging Behind Adoption
MIT Sloan Management Review found that around 50 percent of respondents now have access to AI tools, but 44 percent of those users save less than four hours per week. Only 44 percent receive structured training at all. Access has been solved. Effect has not.
Four hours a week is not nothing. Across a department it is a meaningful number. The problem is that scattered minutes rarely convert into capacity unless somebody decides what the freed time is for. Without that decision the saved time disappears into the general noise of the working day and no controller will ever find it.
From Pilot Project to Core Process
A meta study by Paul Okhrem, drawing on McKinsey, Deloitte, Stanford, IBM, Gartner and MIT, reports that 72 percent of companies use generative AI in at least one area, that only 29 percent achieve a significant return, and that agentic AI produces a financial benefit in just 23 percent of cases. Falling behind is not an option, but piloting alone is clearly not a strategy either.
The transition from pilot to production is where most initiatives quietly die. A pilot has an enthusiastic sponsor, a small group of volunteers and no service level agreement. A core process needs ownership, documentation, a fallback when the system is unavailable, and somebody who is accountable for quality. Those are unglamorous requirements, and they are the actual difference between the 28 percent and everyone else.
A Practical Plan for Midsize Companies
Start with one process that carries real volume and has a clear owner. Measure it for two weeks before you change anything, so that the baseline exists. Define in advance what success looks like in hours, in error rate or in response time, and write that number down where the management team can see it.
Then decide what happens with the time you free up. Reassigning it to a task the team never got to is usually more valuable, and far easier to defend, than promising a headcount saving. Train the managers first, the team second, and review the numbers after 90 days. If the process did not improve, stop it and pick another one. A short list of finished, measured projects beats a long list of pilots every single time.
Mentioned Tools & Resources
- Gartner via Olakai, Only 28% of AI Projects Deliver ROI: olakai.com
- Accenture, Pulse of Change 2026: accenture.com
- Deloitte, The State of AI in the Enterprise 2026: deloitte.com
- McKinsey QuantumBlack, The State of AI Global Survey 2026: mckinsey.com
- MIT Sloan Management Review, AI Use Is Rising at Work but Productivity Gains Lag: mitsloanme.com
- Paul Okhrem, Enterprise AI Adoption & ROI Benchmarks 2026: paul-okhrem.com
Frequently Asked Questions
How many companies actually achieve a return from AI?
Current studies put the figure at 28 percent. Roughly 52 percent of projects are live in some form, so the gap is not about shipping the technology. It is about connecting it to a business result that somebody measures.
Why do so many AI projects fail?
The dominant reasons are missing process design and missing measurement, combined with a focus on tools rather than on value creation. The model is rarely the bottleneck. Know-how, leadership attention and a clear owner for the process are.
How much time do AI tools really save?
For 44 percent of users the saving is less than four hours per week. The number rises when people receive structured training and when the tool is embedded in a specific workflow rather than offered as a general assistant.
Is generative AI enough on its own?
Only 29 percent of companies report a significant return from generative AI by itself. It works best as one component inside a redesigned process, not as a standalone productivity layer bolted on top of unchanged work.
What about AI agents?
Agentic AI currently delivers a demonstrable financial benefit in about 23 percent of cases. Agents act on their own, which increases both the potential and the attack surface, so access rules, logging and a human in the loop belong in the design from day one.
How can a midsize company close the AI skills gap?
Train managers before teams, and teach process decisions rather than button clicks. Education and training are the most common response in the market, but they only pay off when the underlying workflow is redesigned at the same time.
Conclusion
The studies disagree on many details and agree on one thing: the return on AI is not bought, it is designed. Companies that reach the top quartile are not the ones with the largest budget. They are the ones that picked a process with volume, measured it before and after, trained their managers, and stayed with it long enough to move a number.
If you take one action from this article, make it the baseline measurement. Without it, every discussion about AI value stays an opinion. With it, the conversation becomes a business case, and business cases are what survive the next budget round.
