The answer up front: AI creates leverage in workflows that repeat at volume, run on data you actually have, and produce an outcome you can measure. Sprinkled across everything, it fails: MIT found 95 percent of corporate GenAI pilots deliver no measurable return. Pointed at one right workflow, it compounds.
A century ago, factory owners bought electric motors and were furious.
They ripped out their steam engines, installed the shiny new technology, and productivity barely moved. It took decades before the numbers jumped, and the reason is the interesting part: early adopters kept their steam-era factory layouts. Same overhead shafts and belts, same workflow, new power source bolted on. The gains only showed up when someone redesigned the floor around what electricity made possible.
If that story feels familiar, it’s because most companies are currently doing it with AI. New motor, steam-era floor plan.
Why do most AI pilots actually fail?
The numbers are rough. MIT’s NANDA initiative studied corporate GenAI programs in 2025 and found 95 percent of pilots produced no measurable P&L impact. S&P Global Market Intelligence found the share of companies abandoning most of their AI initiatives jumped to 42 percent in 2025, up from 17 percent the year before.
Read past the headlines and the pattern is consistent: the technology mostly worked. The placement didn’t. Pilots got pointed at whatever was easy to demo instead of what the business actually repeats a thousand times a month. A chatbot for a question nobody asks. A summary of a report nobody reads. Impressive on stage, invisible on the P&L.
The failure mode isn’t “AI is overhyped.” It’s “we electrified the steam layout.”
What does real leverage look like?
Leverage means the same effort moves more weight. For AI in a business, that happens in workflows with three properties. This is the test we use, and you can run it on your own company in ten minutes:
- It repeats at volume. A task done 500 times a month beats a task done 5 times, every time. Automation compounds on repetition; one-offs give it nothing to grip.
- The data exists and is reachable. If the workflow’s knowledge lives in one person’s head and a shared drive nobody can navigate, fix that first. AI runs on your data the way that electric motor ran on wiring. No wiring, no motor.
- The outcome is measurable. Hours saved, response time, error rate, revenue per rep. If you can’t measure it, you can’t prove it worked, and (see above) unproven pilots get abandoned.
Score every candidate workflow against those three. Most companies discover their flashiest AI idea fails the test and some unglamorous back-office grind passes it easily. Follow the boring one. The boring one pays.
Where does AI NOT belong?
Honest answer, because this is where the 95 percent went to die:
- Judgment-heavy one-offs. The big strategic call you make once a year gets informed by AI, never delegated to it.
- Anything without a human owner. AI augments people. Remove the person accountable for the outcome and you’ve built an unsupervised intern with infinite confidence.
- Broken processes. Automating a mess produces faster mess. We wrote about the decision-plumbing version of this in The Hidden Cost of Messy Decisions.
Sometimes the honest recommendation is that simple tools are enough, and no build is needed at all. Anyone selling you AI for everything is selling you the motor without asking about your floor plan.
So where should a leader start?
Not with a technology decision. With an inventory. List the ten workflows your team repeats most, run each through the three-part test, and rank them. The winner is almost never the one from the conference keynote. It’s the one your ops manager mutters about.
Then solve exactly one. Prove it, measure it, and let the win fund the next one. Companies that get AI right in 2026 aren’t the ones doing the most. They’re the ones that redesigned one corner of the floor, saw the numbers move, and expanded from evidence.
That first workflow is the entire point of our audit: we look at how your business actually operates and find the one problem worth solving first. If it turns out you don’t need us after that, we’ll say so.
One short read like this, twice a month: what’s changing in AI, why it matters, and what to actually do about it. Get the Insights below.
Quick questions
Should we build custom AI or buy an off-the-shelf tool? Run the volume/data/measurement test first. Off-the-shelf wins for generic, low-volume needs. Custom wins when the workflow is specific to how you make money and repeats constantly. The test decides, not the vendor.
How do we avoid becoming part of the 95 percent? Pick one workflow that passes all three tests, put a human owner on it, and define the success metric before you build. Pilots with a number attached survive; demos don’t.
Do we need our data perfect before starting? No, just reachable for the one workflow you picked. Perfecting all your data before doing anything is how AI projects turn into data projects that never end.
Sources: MIT NANDA, “The GenAI Divide: State of AI in Business 2025” (Aug 2025; 95% of corporate GenAI pilots show no measurable ROI). S&P Global Market Intelligence, 2025 AI Trends survey (42% of companies abandoned most AI initiatives, up from 17% in 2024). The electrification story: economic historian Paul David’s research on the “productivity paradox” of the electric dynamo.
