They're Everywhere

I might have continued telling myself that this was a software industry problem — my problem, my industry’s problem, containable and specific — if it were not for what I started hearing from our customers.

Arcline sold procurement software to manufacturers, logistics companies, pharmaceutical firms, energy conglomerates. These were not technology companies. They did not read Hacker News. Their CTOs did not spend weekends evaluating foundation models. They were operational businesses that moved physical things through physical supply chains, and they had always treated software as a necessary cost of doing business, not as a strategic weapon.

By the second half of 2024, the conversations changed.

The first call that shook me came from a customer I will call Dietrich. He ran supply chain operations for a European industrial manufacturer — one of our oldest accounts, a company that had been on Arcline since 2015. Dietrich was not a man who got excited about technology. He was a man who got excited about on-time delivery rates and container utilization. I liked him precisely because he did not care about our architecture. He cared about whether his parts arrived on time.

“Vikram,” he said, with the directness that I had come to expect from him, “I need to tell you something as a friend. We had a demo last week from a company I had never heard of. Three people, based in London. They showed us an AI agent that does what your platform does for our spot procurement — identifies suppliers, compares quotes, checks compliance, generates the PO. They did the demo live, on our data. It took eleven minutes.”

He paused. I said nothing.

“Our implementation of Arcline for that same workflow took fourteen months.”

I asked him if the quality was comparable. He said it was about seventy percent as robust, maybe eighty. The compliance checks missed some edge cases. The supplier matching was not as nuanced. But for spot procurement — the high-volume, lower-stakes purchases that made up sixty percent of their transactions — it was more than good enough. And the three-person company was offering it at a price point that was, in Dietrich’s words, “almost embarrassing compared to what we pay you.”

Dietrich was not leaving Arcline. Not yet. Our platform handled his strategic procurement — the complex, high-value, multi-stakeholder negotiations where the stakes were too high for an AI agent operating with seventy percent accuracy. But his message was clear: the bottom of our market was being eaten. And bottoms have a way of becoming middles.

This was not a competitive threat from within my industry. The three-person London company was not a procurement software firm. They were three engineers with domain-agnostic AI skills who had pointed a foundation model at a market they found interesting. They could just as easily have pointed it at legal contracts or insurance claims or medical billing. I later looked up their website. It had a single page. Their “About Us” section was three sentences long. I had four hundred engineers and a twelve-page capabilities brochure, and these three people with a one-page website had just replicated the majority of my spot procurement workflow in a live demo on a customer’s actual data.

There is a particular kind of vertigo that comes from realizing your moat is not a moat. I had felt something like it once before, twenty years earlier, when I watched an early cloud software demo and understood for the first time that on-premise was going to lose. That feeling was back, and it was worse, because this time I was the incumbent.

The SIP was not producing competitors in the traditional sense. It was producing capability — raw, transferable, domain-agnostic capability — that could be aimed at any knowledge-work market by anyone with a laptop and sufficient curiosity.

Once I started looking, I saw it everywhere. Not in the technology press, which was saturated with AI coverage to the point of meaninglessness. I saw it in the operational reality of the industries our customers occupied.

Legal services. One of our pharmaceutical customers mentioned, almost in passing, that their legal department had cut outside counsel spending by thirty percent in a single year. Not by hiring cheaper lawyers. By using AI to do first-pass contract review, regulatory analysis, and litigation research. Work that had been billed at four hundred to eight hundred dollars per hour by associates at major law firms was now being done by a system that cost a few dollars per document. The general counsel told me that the quality of the AI’s first pass was “better than most second-year associates, worse than most fifth-year associates.” She had restructured her entire outside counsel strategy around this assessment: AI handles the first pass, experienced lawyers handle the exceptions.

The large law firms, she said, were in a state of visible discomfort. Their economic model — the leveraged pyramid, where a small number of partners profit from the billable hours of a large number of associates — was designed for a world where junior legal work required junior lawyers. That world was ending. Some firms had adapted. Many were still debating whether AI was “ready for real legal work” — the same phrase, almost word for word, that I had used about enterprise procurement eighteen months earlier.

Healthcare. A medical devices customer told me about the radiology department at a hospital system they worked with. For years, the radiology profession had debated whether AI could match human diagnostic accuracy. By 2025, that debate was over. In several narrow diagnostic tasks — detecting certain lung nodules, identifying early-stage diabetic retinopathy, flagging suspicious mammographic findings — AI systems were matching or exceeding the accuracy of experienced radiologists. The debate had moved from “can AI do this?” to “is it ethical not to use AI when it demonstrably improves patient outcomes?”

But radiology was just the visible edge. Behind the scenes, AI was reshaping drug discovery timelines, automating clinical trial enrollment, generating patient communication, and consuming the administrative work that had been devouring clinician time for decades. A hospital CIO I spoke with estimated that forty percent of the administrative labor in their system — the documentation, the coding, the prior authorizations, the compliance paperwork — could be substantially automated within three years.

Education. The Chegg story had been my wake-up call, but the disruption in education extended far beyond one company’s stock price. The entire assessment model of modern education — essays, problem sets, take-home exams — was built on a single assumption: that the student’s submitted work reflects the student’s own capability. AI broke that assumption beyond repair.

Every university, every school system, every credentialing body was now wrestling with a question they had never expected to face: how do you evaluate a human when the human has access to superhuman assistance? The institutions that answered thoughtfully — by shifting toward live demonstration, process-based evaluation, and assessments that tested the ability to direct and evaluate AI rather than replicate its output — were emerging stronger. Those clinging to the take-home essay were engaged in a losing war against a technology that could produce a passable essay on any topic in thirty seconds.

And Chegg, the canary I had been watching since that flight to Chicago? The story only got worse. By late 2025, the company had cut nearly half its remaining workforce, its stock was flirting with delisting, and the CEO who had first sounded the alarm had stepped down. Their attempt to fight back — an AI product called CheggMate — never gained traction. The new CEO’s assessment was devastating in its brevity: It was never a thing. Nearly two decades of market dominance, undone in less than three years by a technology that was not even trying to compete with them.

Financial services. Junior analyst roles at investment banks were being redefined so fast that the job descriptions from 2022 were unrecognizable by 2025. Quantitative analysis, risk modeling, earnings summary generation, client communication drafts — every layer of a financial institution’s value chain was being compressed. An analyst at a mid-tier bank told me that the first version of every research memo in their group was now AI-generated. The analyst’s job was no longer to write the memo. It was to evaluate, refine, and take responsibility for the memo the AI had written.

Creative industries. Perhaps the most emotionally charged disruption of all. I am an engineer, not an artist, but I have friends in design and content creation, and the conversations I had with them in 2024 and 2025 were painful. Writers, designers, illustrators, musicians — entire professions built on the premise that creative work is uniquely and irreducibly human — were confronting a technology that could produce passable creative output at industrial scale. The key word was passable. AI-generated creative work was rarely great. But “great” is not the standard for most commercial creative work. The standard is “good enough, fast, and cheap.” And by that standard, the market had already shifted.

But it was the Klarna story that crystallized the real lesson — not just that AI could disrupt industries, but that how you deploy it determines whether you survive the disruption or become a cautionary tale.

Klarna, the Swedish buy-now-pay-later company, had gone further and faster than almost anyone. In early 2024, they deployed an AI customer service assistant built with OpenAI. The numbers were staggering. In its first month, the assistant handled 2.3 million conversations across twenty-three markets and thirty-five languages. It was doing the equivalent work of seven hundred full-time customer service agents. Resolution times dropped from eleven minutes to under two. Repeat inquiries fell by twenty-five percent. The company projected forty million dollars in profit improvement.

Klarna’s CEO, Sebastian Siemiatkowski, became the global poster child for AI-driven workforce transformation. He made the strategy public. He talked about it on earnings calls, in interviews, at conferences. The company laid off staff. They paused hiring. They let their headcount fall from roughly five thousand to thirty-four hundred. The narrative was intoxicating: AI replaces humans, costs drop, efficiency soars, investors celebrate.

And then reality intervened.

By mid-2025, Siemiatkowski was saying something different. Customer satisfaction had declined. The AI handled simple queries brilliantly — payment status checks, basic account questions, routine transactions. But Klarna’s core product involved payment disputes, credit decisions, and sensitive financial conversations. These were interactions where a customer was anxious, frustrated, or confused. Interactions where nuance and empathy were not nice-to-haves but essential. The AI could not do empathy. It could simulate empathy, but simulation is not the same thing when a customer is worried about a charge they do not recognize on a joint bank account or a credit decision that affects their ability to make rent.

The company reversed course. Siemiatkowski acknowledged publicly that cost had been “too predominant” a factor in organizing their customer service. Quality had suffered. The company began rehiring human agents — not to replace the AI, but to work alongside it. They moved to a hybrid model: AI for the routine, humans for the complex. The AI continued to handle two-thirds of all inquiries, but the remaining third — the hard cases, the emotional cases, the cases that required judgment — went to humans who were better supported and better trained than the outsourced agents they had replaced.

I studied the Klarna reversal obsessively because it contained two lessons, not one.

The first lesson was obvious: full automation fails. AI augmentation works. You cannot replace human judgment wholesale and expect quality to hold. This was important, and I took it seriously.

But the second lesson was subtler, and I think more important. Klarna’s AI was still doing the work of eight hundred agents even after the reversal. The company had saved sixty million dollars. The hybrid model — AI for the routine, humans for the complex — was demonstrably superior to either pure automation or pure human service. The narrative that emerged in the press was “Klarna’s AI strategy failed.” The reality was more nuanced: Klarna’s full replacement strategy failed. Their AI strategy — modified, humbled, made hybrid — was working better than anything they had done before.

The lesson was not that AI doesn’t work. The lesson was that deploying AI as a wholesale replacement for humans is a different thing from deploying AI as an amplifier of humans, and confusing the two is dangerous. One destroys quality. The other transforms economics.

This is the pattern of a true Strategic Inflection Point. It does not respect industry boundaries. It does not politely confine itself to one sector while others watch from a safe distance. It spreads. It morphs. It finds every crack in every business model built on the assumption that cognitive labor is expensive, scarce, and irreplaceable.

Dietrich’s call about the three-person London startup. The pharmaceutical GC cutting outside counsel by thirty percent. The radiologists confronting a machine that sees tumors better than they do. The investment analysts editing memos written by AI. The Chegg CEO admitting defeat. Siemiatkowski pivoting from triumphant automation to humbled hybrid.

These were not separate stories. They were the same story, playing out across every industry simultaneously. And every one of them carried the same implication for Arcline, and for every company like Arcline:

The question was no longer whether AI would reshape your industry. The question was whether you would reshape yourself first.