Same AI Tools, Nearly Double the Revenue

Researchers gave 515 young companies around the world the same package: access to leading AI tools, technical training, mentors and API credits.

Half of them, picked at random, also got examples of how other organizations had redesigned their workflows and business models around AI. Then every company was asked where AI could help in its own work.

Compared with the others, the companies that saw the examples:

• identified 44 percent more ways to use AI
• completed 12 percent more tasks
• were 18 percent more likely to land paying customers
• made nearly double the revenue, and needed significantly less outside capital

The study is by Hyunjin Kim and Dahyeon Kim of INSEAD, and by Rembrand Koning.¹

They call the obstacle the mapping problem. Managers know AI can do useful work. They can’t see where it fits in their own business, or what has to change around it. The research on why AI adoption stalls traces the same gap from the other end: push the tool without touching the workflow, and the cost lands on the people inside it.

The examples widened what managers thought AI could do. Those who got them looked past the obvious tasks, into product development, operations, strategy and administration.

We see the mapping problem from the other side, in client work.

At Yuno, the studio where we work, clients usually arrive wanting to optimize one process they already have in mind. A common one: invoices arrive by email, someone opens each one, finds the data and pastes it into the accounting system.

Then we map it, step by step, through every team the invoice touches.

The typing turns out to be the fast part. The days go to steps in other departments that nobody takes care of, and that some people would fight to keep.

The steps nobody takes care of

A step costs more than its minutes. Everything behind it waits.

At JD Group, a customs company on the US–Mexico border, classifying a new product meant finding the right tariff code. In practice, that meant asking five other people, while the rest of the operation waited anywhere from 30 minutes to 4 hours.

A model trained on the company’s own 25 years of tariff decisions now does it in 10 seconds or less.

That’s why AI that speeds up a task often barely changes how long the job takes. If a request sits in a queue for three days, cutting the task from an hour to a minute saves an hour, not three days.

Why the step survives

In our experience this is especially strong in Latin America. No org chart is immune: the more activities, people and budget sit under your area, the more power you have.

Nobody is protecting the bottleneck. They’re protecting the activity. A step that lives in your department brings people and budget with it, and taking the step away takes those too.

That’s one reason companies skip the question “should this step exist?” and go straight to automating it. Michael Hammer, who started the reengineering movement, warned about this in 1990: “It is time to stop paving the cow paths.”

When you automate a step that shouldn’t exist, you’re scaling a mistake. People save time. The return rarely shows up, because nothing about the business changed.

When we map a workflow, every step goes into one of four boxes:

• Delete. Nobody would miss it.
• Automate. Rules-based work that software can do the same way every time.
• AI agent. Reading, checking and drafting, with a person approving what matters.
• Human judgment. The calls that should stay with a person.

Delete comes first. When a step disappears, the reaction we see most is delight: no more copying invoice data, or asking five people where the right information lives.

A manufacturer’s design engineers spent hours on every customer bid, in a case reported this year by Bain’s Arjun Dutt and economist Aaron Chatterji. In some segments, more than half of those bids were unlikely to win.

Now a 20-minute estimate handles the early bids, and full designs go only to the likely ones. The authors report quotes about 15 times faster, and a higher win rate.

The gain came from a person with a spreadsheet, and from the work they stopped doing.

When someone defends the step

The delight comes from the people doing the work. The people who own the step are another story. Some are protecting their turf. Others see something you don’t.

Start with the ones who see something you don’t. Change researchers John Kotter and Leonard Schlesinger found that people resist because they see the situation differently from the people pushing the change, and sometimes they see it better. A defended step can be the one catching a mistake nobody else sees.

If the step really is waste, take away the fear behind the defense. In 1992, Art Byrne, the new CEO of Wiremold, a maker of cable-management systems, told employees that nobody would lose their job because of process improvement.

Because the company was growing, people whose work disappeared moved to where they were needed, and headcount came down through attrition. Jobs could still change. Nobody would lose one for fixing the work.

In a shrinking business that promise is harder to make, and it doesn’t answer the owner who would lose budget. That conversation belongs to the CEO.

The bottleneck will move

Years ago we helped with a transformation at a large media company in Mexico, the publishing house with the rights to DC and Marvel comics.

The CEO wanted more sales. There was huge demand, but we weren’t selling. The only point of sale was newsstands around the cities.

So we opened a Shopify store and cut out the newsstands. A reader clicked, and the order went straight to us.

We did check with the warehouse first. But digital sales had fallen for three straight years, and the warehouse had planned its capacity on that decline.

Sales took off.

The warehouse asked us to stop selling. The bottleneck had moved.

At JD Group, the next question is whether every step after classification can keep up with hundreds more products a day.

Hyunjin Kim gives the test we’d use for any AI project: does it let you reach more customers, make better products, or use fewer resources? “Otherwise,” she says, “you’re measuring activity rather than impact.”

Borrow the map

In 1954, Sam Walton, who went on to found Walmart, took a 12-hour bus ride to see a new kind of checkout at Ben Franklin variety stores: one central line instead of separate counters. He put it in his own stores right away, and later admitted that almost everything he’d done was copied from someone else.

UBS Wealth Management took something different. Around 2000, its leaders saw that budget negotiations between headquarters and the business units were slowing growth. One executive pointed to Handelsbanken, the Swedish bank that had dropped budgeting in the 1970s.

Senior managers visited. The two banks’ business models differed, but the UBS team borrowed the principles: less oversight from headquarters, more responsibility at the front line, friendly competition between peer units. Then it built its own version, down to its own bonus plan.

Walton looked outside his own stores and copied a practice. UBS looked outside its own industry and took the principle instead of the plan.

The winners in the study did the same with their examples: they took the reasoning and rebuilt it for their own business.

Most AI examples you’ll see are someone else’s practice. Look for the principle underneath.

A first map in ninety minutes

Take one workflow that crosses two teams, one that people complain about.

1. Before the session, collect two or three examples of how organizations in other industries redesigned something similar. Look for the principle in each.
2. Walk the workflow step by step with the people who do it. Ask where work waits, not what work gets done.
3. For each step, ask who owns it and what they would lose if it disappeared. That answer tells you which steps will be defended.
4. Put every step into one of the four boxes. Pick one change, and name how you’ll measure it: hours or money.

The tools were the same

Every company in the study had the same tools. The gap opened on what their managers had seen.

These were early-stage companies. The paper is still under review, and it doesn’t say how much revenue “nearly double” started from. A large organization has more process to move, and more people whose budget depends on it.

The last time your team planned where AI goes next, whose redesigns had they studied?