Something Changed

It was a Thursday in December 2022, two weeks before Christmas, and I was doing what I always do at that hour — reviewing pull requests on my second monitor while half-listening to a product sync on my first. I had been CTO of Arcline for thirteen years. I had written the first ten thousand lines of our codebase myself. I knew every architectural decision, every tradeoff, every load-bearing wall in our system. Arcline was a $600 million enterprise procurement platform used by some of the largest manufacturers in the world, and I had built it with my own hands, one commit at a time.

I had come to the United States from Hyderabad in 1995 with a suitcase and a Stanford admission letter. I had spent a decade at Oracle learning how enterprise software really worked — not the elegant theory, but the brutal reality of deployment, integration, and customer politics. And in 2009, when Anil and I started Arcline from his apartment in Sunnyvale, I had poured everything I knew into a system that would prove, to the industry and to myself, that an immigrant engineer from India could build something that the largest companies in the world would depend on. Arcline was not just my company. It was my proof.

Ravi, a twenty-four-year-old engineer who had joined us five months earlier, walked up to my desk. He had that look — the one young engineers get when they think they’ve found something important and aren’t sure if they’re allowed to be excited about it.

“Have you tried this ChatGPT thing?” he asked.

I had heard of it. A chatbot from OpenAI. Sam Altman’s company. I had seen a few tweets. “Played with it briefly,” I said, which was a lie. I had not played with it at all.

Ravi sat down uninvited and opened his laptop. He typed a prompt — something about generating a Python function that parsed procurement order data from a CSV, validated it against a schema, and flagged anomalies. The kind of task that, in our world, would be a well-scoped ticket for a mid-level engineer. Maybe half a day’s work, including tests.

ChatGPT produced the function in about eight seconds.

It was not perfect. The error handling was sloppy. It made an assumption about date formats that would have broken on European inputs. But the structure was right. The logic was right. The bones were right. A competent engineer could have cleaned it up in twenty minutes.

“Interesting,” I said, in the tone that CTOs use when they mean I need you to go away so I can think about this. “A nice demo.”

Ravi nodded, but he didn’t look convinced by my nonchalance. He shouldn’t have been. I wasn’t convinced by it either.

I want to be honest about what happened next, because honesty is the point of this essay. What happened next is: nothing. I did nothing. For months.

I told myself the stories that every experienced technologist tells themselves when a new technology appears that threatens the foundations of their work. It’s impressive but brittle. It hallucinates. It can’t handle enterprise complexity. It doesn’t understand our domain. It’s a toy.

Every one of these statements contained a grain of truth, which is exactly what made them so dangerous. They were not wrong. They were insufficient. They described the technology’s current limitations while completely ignoring its trajectory. It was like standing on a railroad track, noting that the train is still a mile away, and concluding that you have plenty of time — without checking how fast it’s moving.

I had built a career on my ability to evaluate technology with precision. I could tell you the difference between a genuine technical breakthrough and a well-marketed demo. I had watched hype cycles come and go — blockchain, the metaverse, Web3. I had been right to ignore all of them. And that track record of being right became the very thing that made me wrong this time.

Because this time, the demo was the breakthrough.

The first crack in my confidence came in May 2023. I was on a flight to Chicago — we had a customer meeting at a manufacturing company’s headquarters — and I was scrolling through my news feed when I saw the headline about Chegg.

Chegg was an education technology company. Online tutoring, homework help, textbook solutions. Not my industry. Not my problem. But the headline stopped me cold.

Chegg’s stock had fallen forty-nine percent. In a single day. One billion dollars in market capitalization, gone between the opening bell and the close. The cause? On their earnings call, Chegg’s CEO Dan Rosensweig had said something that no public company CEO had ever said before. He told analysts that ChatGPT was hurting their business.

His exact words haunted me. In the first part of the year, he said, they had seen no impact from ChatGPT. They were meeting expectations on new signups. But since March — just two months — they had seen a significant spike in student interest in ChatGPT. It was having an impact on their new customer growth rate.

Two months. That was all it took for a free chatbot to begin displacing a company with seven hundred million dollars in annual revenue and a database of seventy-nine million solved problems built over nearly two decades.

I read the analyst reactions. Goldman Sachs slashed their price target. Jefferies downgraded the stock. The Jefferies analyst wrote something that stayed with me: the concern was not just that students were leaving Chegg for ChatGPT, but that the shift could become viral — spreading across campuses the same way Chegg itself had once spread. The very network effects that had built the company could now work in reverse, pulling students toward a free alternative with zero switching cost.

But here is the detail that truly disturbed me, the one that most of the coverage missed. Before ChatGPT launched, Chegg’s own employees had suggested building AI tools to automate their answer generation. Their engineers had seen the potential. They had raised the alarm. And leadership had denied the request. They had looked at the most important technology shift of their era and said: not yet.

After ChatGPT launched, some inside the company still were not worried, because the chatbot sometimes gave wrong answers. They pointed to hallucinations. They pointed to inaccuracy. They were right about the flaws and catastrophically wrong about the trajectory. Students, it turned out, did not need perfect answers. They needed good enough, fast, and free. The same way people had not needed Wikipedia to be as reliable as Encyclopaedia Britannica. They needed it to be accessible.

I put my phone down and stared out the airplane window. Chegg was a homework help company. I ran an enterprise procurement platform. The domains could not be more different. But the underlying economics were identical. Chegg had built a business on the assumption that generating answers to questions required expensive human labor. I had built a business on the assumption that managing complex procurement workflows required expensive human configuration, expensive human analysis, expensive human judgment.

What if that assumption was wrong?

Not wrong today. Not wrong this quarter. But wrong soon enough to matter.

I landed in Chicago and gave the customer meeting my full attention. It went well. Our platform was deeply embedded in their operations. They had spent eighteen months implementing it. The switching costs were enormous. I told myself this was our moat.

But on the flight home, I opened my laptop and did something I had been avoiding. I went to ChatGPT — GPT-4 had launched by then — and I started testing it on our domain. Not on toy problems. On real problems. I took actual procurement scenarios from our customer support logs — the messy, ambiguous, domain-specific questions that our product was designed to handle — and I fed them to the model.

The results were uneven. On some queries, it was shockingly good. On others, it was confidently wrong. On a few, it hallucinated specifications and compliance requirements that did not exist. I could have stopped there. I could have written a memo to my leadership team titled “AI Assessment: Not Ready for Enterprise Procurement” and moved on with my life.

Instead, I did something that I now recognize as the most important decision I made that year. I kept going. I spent the entire six-hour flight prompting, refining, testing. I gave the model context. I gave it examples. I gave it our documentation. And I watched the quality of its answers climb — not to perfect, but to disturbingly competent.

By the time we landed in San Francisco, I had a knot in my stomach that would not go away for months.

Here is what I have learned about Strategic Inflection Points, having now lived through one: the hardest part is not the change itself. It is the period before you accept that the change is real.

Andy Grove wrote about this in the original Only the Paranoid Survive. He described the years when Intel was losing the memory chip business to Japanese competitors. The data was there. The market share numbers were there. The pricing pressure was there. But Intel’s identity was memory chips. That was who they were. That was what they had built. And so, for years, very smart people looked at very clear data and found reasons to believe that the situation was temporary, manageable, recoverable.

I was doing the same thing. I was looking at a technology that could already do twenty percent of what my four-hundred-person engineering team did, and I was focusing on the eighty percent it could not do. I was comforting myself with the gap instead of terrifying myself with the trajectory.

Because here is the thing about exponential curves that every engineer understands mathematically and almost no one understands emotionally: by the time the change is obvious, it is too late to respond. The moment when the technology goes from “interesting but limited” to “good enough to threaten your business” is not a moment at all. It is a phase transition. And phase transitions, by definition, happen faster than the systems they transform can adapt.

Chegg’s stock would continue to fall. By the end of 2024, the company that had been worth fourteen billion dollars would be worth less than two hundred million. They would lay off forty-five percent of their workforce. They would face delisting from the New York Stock Exchange. Their CEO would step down. And the most telling detail of all: their attempt to fight back by building their own AI product — a GPT-4-powered tool called CheggMate — would be described by the new CEO with two words: It was never a thing.

Chegg was the canary in the coal mine. They were the first publicly traded company to bleed out from AI disruption, in full view of the market. And the lesson was unmistakable: this was not a technology that would politely wait for incumbents to formulate a response. This was a technology that would eat your business while you were still scheduling the meeting to discuss whether it was a threat.

I did not act on that lesson. Not immediately. I went back to running Arcline the way I had always run it. I reviewed pull requests. I debated architectural decisions. I planned our next release. The unease from that flight stayed with me — a low hum beneath the surface of every meeting, every roadmap discussion, every quarterly plan — but I learned to talk over it. I told myself that enterprise software was different. That our customers needed reliability, compliance, auditability — things that a language model could not provide. That our competitive moat was deep and our switching costs were high.

All of this was true. And none of it would save us.

Something had changed. Not in the way a new product launch changes a competitive landscape, or the way a recession reshuffles market share. Something more fundamental. The cost of cognitive labor — the thinking, synthesizing, drafting, analyzing work that had defined the value of knowledge workers for seventy years — was collapsing toward zero. Not gradually. Not linearly. Exponentially.

And I, the CTO who had built a six-hundred-million-dollar business on the assumption that cognitive labor was scarce and expensive, was watching it happen from the window of a United Airlines flight and telling myself it didn’t apply to me.

It applied to me. It applied to everyone.

I just didn’t know it yet.