What Ford, FedEx, Progressive, and Dell understood about technology that most AI deployments still don’t.
In 1993, Ford Motor Company had 500 people processing invoices in accounts payable.
They set out to cut that number in half. They benchmarked. They flow-charted. They redesigned the process, step by careful step. Then someone looked at what Mazda was doing — a company Ford had a stake in — and discovered Mazda ran the same function with five people.
Not 250. Five.
The gap wasn’t efficiency. It was architecture. Mazda hadn’t automated invoice processing. It had made invoice processing irrelevant.
Ford eventually did the same. They implemented a system where payment triggered automatically when goods arrived at the warehouse and matched a purchase order. No invoice required. No three-way reconciliation. No 500-person department. The enabling technology was EDI — electronic data interchange — and real-time goods receipt data. That combination made it possible, for the first time, to pay a supplier without waiting for a piece of paper confirming what had already happened.
The door didn’t exist before. Ford walked through it. They came out with a 75% reduction in headcount, not the 25% they’d originally aimed for.
Actually, Let’s Go Back Further
Before EDI. Before the internet. Before GPS.
In 1979, Frederick Smith’s Federal Express was six years old and losing money at a pace that would have killed most companies. Smith had a thesis: businesses would pay a premium for overnight delivery if the package arrived reliably and on time. The problem was reliability. Packages disappeared into the logistics system. Nobody — not the sender, not the recipient, not FedEx itself — knew exactly where anything was at any given moment.
So Smith built COSMOS. Customers, Operations and Services Master Online System. The first real-time package tracking platform in the industry.
The enabling technology was barcode scanning and a centralized computing network that could track every package at every handling point. Those technologies existed. What didn’t exist, before COSMOS, was a process built around continuous visibility as the product. Every other carrier was selling delivery. FedEx decided to sell certainty.
COSMOS made it possible, for the first time, to tell a customer: your package is here, right now, at this facility, and it will be there tomorrow. That wasn’t a feature. It was a redesign of what the service was. The tracking number — which seems obvious now — didn’t exist before. FedEx invented it because COSMOS required it.
Competitors copied the delivery model. Nobody matched the certainty model for years. By the time they did, FedEx had built a logistics empire on the difference.
Same pattern. Different decade. Different technology. Same question: what does this make possible that wasn’t possible before?
The Question Most Companies Get Wrong
Every major technology wave creates a window. During that window, things that were previously impossible become possible — not faster, not cheaper, but actually possible for the first time. The companies that notice the window and redesign around it win permanently. The companies that use the technology to do their existing steps more efficiently win some margin points and then wonder why the gap keeps widening.
This is not a story about innovation culture. It’s not about agility or transformation mindset. It’s about something more specific: the difference between automating a process and making it obsolete.
The question most companies ask when a new technology arrives is: how do we use this in what we already do?
The question that creates separation is: what does this make possible that wasn’t possible before?
Those are not the same question. And right now, in the current AI era , most companies are asking the first one.
Progressive Insurance Didn’t Speed Up Claims. It Changed Where They Happened.
In 1994, Peter Lewis’s Progressive Insurance introduced the Immediate Response Vehicle — a specially outfitted truck that drove trained claims adjusters to the scene of an accident, often within hours.
The old model: accident happens, customer calls insurer, insurer schedules an adjuster, adjuster visits when available, paperwork follows, settlement eventually arrives. Each step handled sequentially, each one dependent on the previous.
The new model required two things that hadn’t existed together before at scale: mobile deployment infrastructure and GPS routing. With those, you could put a professional at the physical site of loss before the car had been towed away. You could assess damage in real time, on the ground, with the customer standing next to you.
What Progressive noticed wasn’t “we could route our adjusters more efficiently.” It was something more fundamental: the accident scene is a better place to process a claim than an office. The information is right there. The customer is right there. The car is right there. The moment GPS and mobile deployment made that feasible at scale, Progressive redesigned the entire claims model around the scene rather than the office.
Progressive grew through the late 1990s to become the third-largest auto insurer in the United States. Competitors who bought dispatch software to schedule their existing adjusters faster got modest improvements. Progressive got a different process altogether.
Dell Decided Not to Build a PC Until Someone Had Paid For It
In 1996, Dell moved its entire sales operation to the internet. The question they were answering wasn’t “how do we sell computers online?” It was: “what does computing actually need — and does any of this have to happen before a customer orders?”
The answer, it turned out, was no. No retail channel. No pre-built inventory sitting in warehouses depreciating. No distributor taking margin. The internet made it possible, for the first time at scale, to take a custom order, route it to a just-in-time supply chain, build the machine, and ship it directly — all without a finished product ever existing until someone had asked for it.
Dell’s competitors had warehouses full of pre-configured PCs they’d have to discount if they didn’t sell. Dell had near-zero inventory because internet connectivity made build-to-order viable. By 1997, Dell’s internet sales reached $4 million per day. Revenue grew more than 50% annually from 1995 to 1998. Dell didn’t become the world’s largest PC seller by making better computers. It became the world’s largest PC seller by redesigning what the supply chain needed to look like given what the internet had just made possible.
The Japanese electronics giants — Hitachi, Sony, Fujitsu — had stronger brands, manufactured their own components, and had every structural advantage. They lost to a college dropout from Texas who noticed that the channel no longer needed to exist.
What the Wrong Version Looks Like
CIGNA ran a massive business process reengineering program in the early 1990s — over 20 separate initiatives. They saved more than $100 million in aggregate. By almost any measure, a success.
Except roughly half their individual projects failed to deliver what was expected. And the pattern, when researchers looked at it, was consistent: the projects that failed tended to automate existing processes rather than replace them. They sped up the same steps. Reduced friction within the same logic. Got better at doing something that perhaps shouldn’t have been done the same way at all.
The projects that worked redesigned around what technology had made possible. The ones that didn’t gave CIGNA faster versions of what they already had.
That $100 million in savings is real money. But it’s not separation.
The Doors That Are Open Right Now
Klarna deployed an AI customer service assistant in early 2024. In its first month, it handled 2.3 million conversations, resolving most of them independently. Average resolution time fell from 11 minutes to under 2 minutes.
But here’s where it gets instructive: customer satisfaction scores dropped on complex interactions. The AI was handling volume without handling difficulty. Klarna subsequently moved toward a human-hybrid model for edge cases. The lesson isn’t that the technology failed. It’s that Klarna initially asked the wrong question — “can AI handle our support volume?” — rather than the better one: “which customer problems can AI now resolve that previously required a trained human, and which ones still don’t?”
That second question points toward a redesign. The first one points toward a faster version of the old model.
Legal contract review is another door that’s been quietly opening. Work that required a junior associate to spend two days reading, cross-referencing, and flagging clauses can now be done by AI in minutes, with comparable accuracy on standard commercial terms. The firms that are winning with this aren’t asking “how does AI help our associates work faster?” They’re asking “what is the associate’s job when AI handles the reading?”
Those are different jobs. They require different training, different hiring profiles, different pricing models.
The Practitioner Question
If you’re leading transformation inside an established organization right now, you’re probably being asked to find AI use cases. Build a roadmap. Identify efficiency gains. Show a business case.
That’s all reasonable. But it tends to produce a particular kind of thinking — the kind that looks at every existing process and asks: where can we slot AI in?
The harder and more valuable exercise is different. Take your most process-heavy functions. Ask: what does AI make possible here that wasn’t possible before? Not faster. Not cheaper. Actually possible for the first time.
First-contact resolution of genuinely complex customer queries — possible now, wasn’t before.
Claims assessment from photos in under two minutes — possible now, wasn’t before.
Legal review without bottlenecking on human hours — possible now, wasn’t before.
Real-time synthesis of customer signals across every touchpoint simultaneously, feeding into a personalized response in the moment — possible now, wasn’t before.
Each of those is a door. The question is whether you walk through it or stand in front of it installing a faster lock.
Ford didn’t streamline invoice processing. They made invoice processing unnecessary.
FedEx didn’t make deliveries more reliable. They made visibility the product.
Progressive didn’t schedule adjusters more efficiently. They moved the claims office to the accident.
Dell didn’t build a better retail channel. They eliminated the channel.
The technology enabled each of those moves. The move itself required someone to ask a different question than everyone else was asking.
I don’t think the hard part is the technology. I think the hard part is giving yourself permission to question why the process exists in the form it does — and whether that form was always a workaround for something the technology just fixed.
Most processes are. Most of them were designed around constraints that no longer apply.
Which means the more useful question isn’t “what can AI do?” It’s: “what constraint did we build this process around — and is that constraint still real?”
Answer that honestly, and the redesign usually becomes obvious.