Most leadership teams share a quiet assumption about AI. The results will come once the technology matures. A better model, cleaner data, the next release, and the value will follow.

The evidence says otherwise. In 2025, MIT's Project NANDA reviewed more than 300 public AI initiatives for its report, The GenAI Divide. It found that despite $30 to $40 billion in enterprise spending, 95% of organizations were seeing no measurable P&L return. The 5% that were seeing results were not using better models. They were using AI differently.

The technology has arrived. The way most companies deliver it has not.

Key takeaways

The model isn't the bottleneck. The work around it is. Most AI is added to workflows designed for people passing information by hand.

We have seen this before. Electricity took four decades to lift factory productivity because owners kept their old floor plans.

Automating a task inside a slow decision just makes the waiting faster. Value comes from redesigning around decisions, not tasks.

Leaders don't need to wait decades. Redesigning one workflow at a time produces measurable results in weeks.

We have been here before, and it took forty years.

In 1900, a factory owner who replaced his steam engine with an electric dynamo would have been disappointed. The new power was cleaner and more flexible. Output barely moved.

Power stations were running in New York and London by 1881, yet electricity took roughly four decades to show up in productivity numbers. When it did, in the 1920s, US manufacturing productivity grew at more than five percent a year. Economic historian Paul David explained why in his paper The Dynamo and the Computer.

Steam-era factories were built around one engine and a central shaft, with belts carrying power to every machine at the same speed. Machines sat where the shaft could reach them, not where the work needed them. The first electric factories kept all of it. They swapped the engine and left the shaft, the belts and the floor plan in place.

The gains came only when manufacturers redesigned the factory itself. Each machine got its own motor. Lines were laid out around how materials actually moved. Workers who now controlled their own machines made more decisions, which meant new training and new expectations. As David put it, the payoff came from reorganizing the work, not from installing the technology.

Most AI is being bolted onto steam-era workflows.

The MIT researchers did not blame the models for the 95%. They pointed to tools that don't learn from feedback, brittle workflows and a poor fit with how daily operations actually run. That is a new power source inside an old building.

We see the same pattern in our own work. In onsite workflow reviews across healthcare facilities, our teams found about two hours of inefficiency in every patient admission, including roughly 55 minutes of documentation delays and 32 minutes of nursing wait time. None of that is a model problem. Add an AI tool on top and those two hours stay exactly where they are.

It looks the same in every industry. A copilot is added to an approval chain built for people emailing documents. A forecast is generated in seconds, then waits for a weekly meeting. An agent drafts a risk summary that lands in the inbox it was meant to replace. Each tool works. The results don't change, because the work doesn't.

Automating a task inside a slow decision just makes the waiting faster.

Most AI projects start by asking which task to automate. The better question is which decision to change. Which supplier to re-rate. Which asset to service. Which shipment to reroute.

Starting with the decision forces the right follow-up questions. What data does it need? Who owns it? How fast should it happen? The workflow follows from the answers. A global chemical manufacturer we worked with didn't need more maintenance data. It needed the data it already had to reach decision-makers in real time instead of a week late. That shift cut maintenance cost per site by 76%.

The people around the AI decide whether it works.

When electric motors put control at each machine, factories had to change how they trained and trusted workers. AI creates the same pressure, and most organizations are underestimating it. In a 2026 Writer and Workplace Intelligence survey of 2,400 people, 44% of Gen Z employees said they work against their employer's AI rollout.

That resistance is a design problem, not an attitude problem. When roles are left unchanged, AI feels like something done to people. In an AI-native organization, the roles are redrawn on purpose. People set priorities and own the decisions that matter. Agents handle monitoring, analysis and repeatable work. Specialists keep both improving. Everyone can see where they fit.

Governance and cost belong in the floor plan, not the invoice.

In the old factory, the shaft set the limits of the building. In AI programs, governance and cost set the limits too, but usually late, after an incident or a surprising bill. Token spend climbs quietly. Agents run without a named owner. Nobody can trace why a recommendation was made.

Organizations that scale AI design these in from day one. Every agent has an owner and a policy. Every decision can be traced to the data behind it. Model choice, routing and token use are managed as operating decisions, the same way leaders manage any other cost.

An AI-native operating model is the redesign, not the add-on.

This is the lens we use at Aligned Automation. Our AI-Native Operating Model treats AI as a reason to rebuild how work gets done, one workflow at a time. A small pod of your people, our specialists and AI agents redesigns a single high-value workflow end to end. Through Automate30, it becomes a working, governed solution in 30 days, then gets measured before the next cycle begins. Every solution runs in AAxon, where governance, monitoring and cost control are part of the floor plan.

The results come from changing the work, not just the tools. On one global engineering program, a connected view cut reporting time by 75% and made decisions four times faster. For a Fortune 500 manufacturer, connected procurement intelligence identified and realized more than $220 million in value.

Three questions to bring to your next AI review.

Which decision does this change, and who owns it? If the answer is a task, such as summarizing reports, expect activity, not value. If no one owns the decision, expect the initiative to stall when the pilot team moves on.

If we designed this workflow today, with agents available, would it look like this? If the AI version keeps every handoff, review and meeting from the old process, you have kept the shaft and the belts.

What will we be able to measure in 30 days? If nothing can be measured in a month, the scope is too vague to succeed.

The lesson from the dynamo is not that new technology is slow. It is that the value arrives when leaders stop asking where to plug AI in and start asking what to rebuild around it.

Ready to find your first workflow? Talk to our team.

Related reading
Before you buy the AI tool, fix the workflow
When maintenance data is a week old, so are your decisions

‍

FAQ

Why do most AI pilots fail to deliver ROI?

Most pilots are added to workflows that were never redesigned for AI. MIT's 2025 GenAI Divide report found 95% of organizations saw no measurable P&L return, and pointed to brittle workflows, tools that don't learn from feedback, and poor fit with daily operations rather than model quality.

What is the productivity paradox, and how does it apply to AI?

The productivity paradox describes powerful new technology that fails to show up in productivity numbers for years. Electricity took about four decades to lift factory productivity because factories kept their old layouts. AI shows the same lag when companies add it to existing processes instead of redesigning them.

What does an AI-native operating model mean?

An AI-native operating model redesigns how work gets done around AI. People set priorities and own key decisions, AI agents handle monitoring, analysis and repeatable execution, and governance and cost control are built in from the start.

How can leaders tell if an AI initiative will scale?

Ask which decision it changes and who owns it, whether the workflow would look the same if designed today with agents, and what can be measured within 30 days. Initiatives without clear answers tend to stall after the pilot.