The largest untouched market on earth

Last year Klarna replaced roughly 700 customer-service agents with artificial intelligence. This year it started hiring people back. Quality had slipped, the company said, and customers preferred talking to someone.

Klarna is not an outlier. MIT's NANDA initiative, which examined more than 300 corporate AI programs and interviewed 52 senior leaders, reported last year that 95% of organizations were getting no measurable return on their generative-AI pilots. Only about 5% of custom tools survived the move from pilot into production.

Now set that against the other half of the picture. The Census Bureau reported in May that only 19.8% of U.S. businesses use AI in any part of their operation. Among companies with 250 or more employees the figure is 37%. Among those with fewer than 20 employees it is under 20%, and it has not moved in six months while every larger category climbed.

Sitting out the revolution

When the Census Bureau last asked why, nearly 82% of businesses with fewer than five employees said AI wasn't relevant to them. Cost and privacy barely registered. They simply did not see what it had to do with their work.

For a one-person business, that is often true. But you hear the same thing from companies running fifty employees and a fleet.

The reason is simpler than it looks. Most of what gets sold to these companies is badly engineered for the job. The workflow assumptions are wrong, the fit is poor, and the systems design behind it is thin. A company with six trucks and one dispatcher looks at what's on offer and finds nothing built for how it actually works.

Klarna went the other way

Klarna's error was the opposite. Customer service is the most exposed part of a company. Every mistake happens in front of the customer, and there is no chance to catch AI mistakes before rollout. It is the last place a careful operator would start. But it is also where the labor cost is concentrated, which is why so many companies went there first.

Also, and more importantly, Klarna moved before the tools were ready. Voice AI in 2024 was not what voice AI is today. Grok and its peers now handle conversation at a level that simply did not exist when Klarna committed. The ground under this technology shifts every few months. Knowing what these tools can and cannot do this quarter is the difference between success and failure.

What the winners did

Some companies are making it work. MIT found the difference had little to do with the technology. It was setup: narrow scope, back-office first, systems built to flag what they were unsure about instead of guessing. Nothing exotic, and nothing a competent operator couldn't run.

The price of entry

For a century, automation required scale.

Building any automation cost real money. Someone had to define the process, connect the systems, test it, and keep it running. To justify that expense you needed to do the same thing thousands of times. So automation went to payroll departments cutting 50,000 checks and insurers handling a million claims a month.

Everyone else was priced out. A business handling 400 orders a month could never justify the cost of automating them. That is why four decades of automation reached the largest companies and stopped there.

Agents change the arithmetic. Teaching a system a process now costs roughly what it costs to explain that process to a new employee. At that price, automating something you do 40 times a month makes sense.

Traditional automation only worked on processes that had already been standardized. If the work varied from job to job, you had to eliminate that variation before a machine could handle it, and doing so usually cost more than the automation saved. Agents work with the variation. The process no longer has to be cleaned up first.

The durable margin

The assumption is that the returns go to whoever builds the models. I would take the other side. Model capability is converging quickly, and any advantage built on being slightly ahead of it is temporary. The durable margin belongs to whoever aims the technology at a specific, ugly problem inside a working business.

That advantage builds over time, because it runs on knowledge nobody outside the company has: how the work really gets done, which rules matter, and which are just habit, and what the exceptions are. Experienced operators carry decades of it in their heads. None of it was ever worth writing down, because no technology could use it.

Small firms employ 62.3 million Americans, roughly 46% of the private workforce, and account for 43.5% of GDP, according to the Small Business Administration. Through the entire software era they were served last. In the AI era they own what the technology needs most: detailed knowledge of how the work actually gets done.

What actually works

I have had success with this, and so have the companies MIT identified, for the same unglamorous reasons: deploy first where errors are cheap and contained; keep autonomous processes away from anything irreversible; never let the system that recommends an action be the one that approves it; maintain an audit trail good enough to reconstruct a failure; and re-benchmark the tools every quarter, because the answer keeps changing.

None of that is technical. Think of these systems as a capable new hire who follows instructions exactly as written and has no instinct for when something looks wrong. You would not hand that person the checkbook on day one. You would give them narrow work, check the output, and widen the scope as they earned it. The same judgment applies here, and any experienced operator already has it.

Klarna did not fail because AI cannot handle customer service. It failed because it handed the most exposed job in the company to a tool that wasn't ready, with nothing in place to catch what went wrong. The technology has since caught up. What Klarna lacked was not a better model. It was someone who could tell them the tools weren't ready yet.

The tools are ready now, and they are cheap enough to change the arithmetic on the other side of the transaction too. Serving 6.4 million small operations was never worth anyone's time, because the cost of building anything for one of them could never be recovered. That is no longer true. What is still missing is judgment: knowing an operation well enough to see where this belongs, what to ask of it, and when not to believe the answer. That is discernment built from experience, and no model has it.

That judgment is already sitting inside those businesses. What nobody has shown them is how to use it, how to turn what they already know into systems that scale the operation and pay for themselves.