In 1899, electric motors provided less than 5% of the mechanical drive power in American factories. The first Edison power stations had been running for almost two decades by then. Owners were buying the new technology. They just installed it the old way: motors bolted onto plants built for steam, with one central shaft, belts running to every machine, and the whole floor arranged around the power source instead of the work.
Productivity barely moved for another twenty years. The gains arrived in the 1920s, once engineers began designing factories around what electricity made possible. The economist Paul David told this story in 1990 to explain why computers were showing up everywhere except the productivity numbers.
In 2026, it reads like a warning label for AI. Companies are buying intelligence and bolting it onto processes built for a different era. The results look much the way factories looked in 1910.
Key Takeaways
- Almost 90% of organizations say they are at least experimenting with AI. Only 7% report scaling it across the enterprise (McKinsey, June 2026).
- 56% of CEOs say their companies have seen no significant financial benefit from AI so far (PwC 29th Global CEO Survey, January 2026).
- AI leaders are twice as likely to redesign workflows to incorporate AI rather than simply add AI tools to the way they already work (PwC, April 2026).
Everyone Is Adopting AI. Few See a Return.
Adoption stopped being the story a while ago. In McKinsey's June 2026 survey of 1,000 executives across 696 businesses, almost 90% of organizations said they were at least experimenting with AI. Only 7% reported scaling it across the enterprise.
The value that does exist pools in a small group. PwC's 2026 AI Performance Study, published in April, surveyed 1,217 senior executives and found that just 20% of companies capture nearly three quarters of the economic value AI generates.
The people signing the checks have noticed. In PwC's 29th Global CEO Survey, published in January 2026, 56% of 4,454 CEOs said their companies have seen no significant financial benefit from AI: no revenue lift, no cost reduction. Only 12% reported both. And in a Gartner survey of 782 infrastructure and operations leaders published in April 2026, only 28% of AI use cases fully succeeded and delivered the return on investment leaders expected. One in five failed outright.
The Problem Is Not the Model
When researchers look at what separates the companies that see returns from the ones that do not, the model barely features.
Microsoft's 2026 Work Trend Index, built on surveys of 20,000 knowledge workers who use AI at work, found that organizational factors like culture, manager support and talent practices account for 67% of AI's reported impact. Individual factors account for 32%. The same report found that 45% of those workers say it feels safer to focus on current goals than to redesign how work gets done with AI.
Ask the technologists and the answer is similar. In a June 2026 vendor survey of 4,625 IT leaders by Confluent, the top-ranked barrier to success with agentic AI (software that carries out multi-step work toward a goal) was a skills gap and limited organizational readiness, at 69%. The reliability of the models themselves ranked below it. So did data quality.
of CEOs say their companies have seen no significant financial benefit from AI so far.
of AI's economic value is captured by just 20% of companies.
of AI use cases in IT infrastructure and operations fall short of fully meeting ROI expectations.
of companies already running agentic AI report stalled projects.
Michael Hammer made the same argument in 1990, in a Harvard Business Review essay titled "Reengineering Work: Don't Automate, Obliterate." Companies were using computers to speed up processes designed for paper, and he called the result what it was: the same mess, faster.
What the Winners Do First
PwC's study puts the difference in one sentence: AI leaders are twice as likely to redesign workflows to incorporate AI rather than simply add AI tools to the way they already work. They are also 2.8 times more likely to have increased the number of decisions made without human intervention. That only works when the process defines which decisions those are.
McKinsey's June 2026 study points the same direction. Companies that embed AI across multiple functions report nearly double the profit margins of peers that use it in a few departments, and a three-year return on invested capital (the profit earned on the money put to work) more than five times higher. These are self-reported numbers, and embedding AI widely is not identical to redesigning a process. But the study's conclusion is blunt: the operating model around the technology matters as much as the technology.
Read the failure data and the success data together and the pattern is hard to miss. AI bolted onto the existing way of working produces pilots. AI given a redesigned process to run produces margin.
But Don't AI Agents Adapt to Any Process?
The strongest objection to all this comes from the agent vendors themselves. The pitch: you no longer script software step by step. You give an agent a goal and guardrails, it works across the tools you already have, and it handles ambiguity instead of breaking on the first exception. On that view, process redesign is the old reflex from the big enterprise-software rollouts, months of mapping workshops before any value, exactly what this generation of software was supposed to end.
The objection deserves to be taken seriously. Nobody should relive the era when the system imposed its logic on the business and called it best practice.
But the early returns point the other way. In the same Confluent survey, only 32% of organizations had agentic AI in production at all. Among those already running it, 77% reported stalled projects and 61% reported project abandonment as a problem. The companies furthest along are the ones hitting the wall, and the wall is not the model. Gartner's read on failed AI use cases in operations is that leaders expected too much, too fast, from initiatives that were overly ambitious or poorly scoped.
There is also a quieter logical problem. An agent needs someone to define the goal, set the guardrails and decide what done looks like. That is a process specification by another name. Skip it, and the agent learns the current mess, exceptions, workarounds and all.
So concede the vendors their real point: agents have collapsed the cost of running a well-defined process. What they have not done is remove the need to define one. The difference from the reengineering era is sequencing. You no longer need a multi-year program before value shows up. You need one process, mapped and fixed, with outcomes the AI can be measured against. That is weeks of work. Which is exactly why doing it first has become affordable.
The Playbook: Fix the Process, Then Run AI
The playbook is four things in the right order, one process at a time.
Start where money leaks: deductions, freight invoices, supplier statements. One process, one owner, one number to move.
The documented process and the real process are never the same. Capture the actual steps, handoffs and workarounds before anything gets automated.
Cut the steps that only exist because of an old system limit. Decide where a person approves. Define what a good outcome looks like for every run.
Give the improved process to the AI and score every run against the outcome you defined. Scale to the next process once the first one pays.
The order is the whole trick. Map before you fix, fix before you automate, measure from day one. Run it in reverse, automation first, and you get the 2026 scorecard above.
Why Duvo Starts With the Process
This playbook is how Duvo works. Duvo maps how a process actually runs, including the workarounds nobody documented, then designs the better version and runs it across the systems you already have. No rip and replace. The first process is live in 4 weeks, for one fixed price per process run.
The sequence matters more than the software. When Rohlik Group put supplier reconciliation through this loop, it protected €2.1M in revenue and €1.4M in margin a year.
If you want to see your own process mapped first, Duvo can deliver the map and the savings case as a standalone step. Or bring the process that keeps slipping and book a demo to see it run.
Sources
- PwC, 2026 AI Performance Study (April 2026)
- PwC, 29th Global CEO Survey (January 2026)
- McKinsey, Putting AI to Work: The Operational Excellence Imperative (June 2026)
- Gartner, AI projects in infrastructure and operations stall ahead of meaningful ROI (April 2026)
- Microsoft, 2026 Work Trend Index (May 2026)
- Confluent, 2026 Data Streaming Report (June 2026)
- Paul A. David, "The Dynamo and the Computer", American Economic Review (1990)
- Michael Hammer, "Reengineering Work: Don't Automate, Obliterate", Harvard Business Review (1990)
Frequently Asked Questions
Why do most AI initiatives fail to show ROI?
Mostly because they automate work as it currently runs. The reasons executives give for weak AI results cluster around organization and process, not model quality: Microsoft's 2026 Work Trend Index attributes 67% of AI's reported impact to organizational factors, and IT leaders in Confluent's 2026 survey rank a skills gap and limited organizational readiness above model reliability as the top barrier. An AI tool dropped into a broken process inherits the broken process. We broke down the recurring failure patterns in more detail in why AI pilots fail.
Do AI agents remove the need for process redesign?
No. Agents make a well-defined process much cheaper to run, but someone still has to define the goal, the guardrails and what a good outcome looks like, and that definition is process design. Among companies already running agentic AI, 77% report stalled projects (Confluent, June 2026). What agents do change is the effort: defining and fixing one process now takes weeks.
What does process transformation mean before AI adoption?
Three things, done for one process at a time: capture how the work actually runs today, including the handoffs and workarounds that never made it into documentation; remove or redesign the steps that exist only for historical reasons; and define a measurable outcome for every run. AI then executes the improved process and gets scored against those outcomes. For how this mapping works without clean system logs, see process mining vs process discovery.