"Signal" or "Noise"?

After that weekend in New York with Anil, after the board approved the transformation plan, after I had committed — publicly, irrevocably — to tearing down the architecture I had spent thirteen years building, I did something I should have done two years earlier. I went back and tried to reconstruct, honestly, the signals I had received and the ones I had ignored. I wanted to understand my own failure of perception. Not to punish myself, but because I knew that the next inflection point would come eventually, and I did not want to be slow again.

The cruel trick of a Strategic Inflection Point is that it is obvious only in retrospect. In real time, it is shrouded in noise. And between 2023 and 2025, the AI landscape produced more noise per unit of signal than any technology shift I have ever witnessed. Every week brought a new model, a new benchmark, a new startup, a new fundraising round, a new apocalyptic prediction, a new utopian promise. The discourse was simultaneously hysterical and banal. It was possible to read about AI for two hours a day and come away less informed than when you started.

And yet, buried in the noise, the signals were there. They were always there. The question is why I did not act on them.

Let me separate them cleanly, with the benefit of hindsight that I did not have at the time — and with the honesty to admit which ones I saw and dismissed, and which ones I did not see at all.

Signal: The cost curve.

This was the most important signal, and it was hiding in plain sight. The cost of running a state-of-the-art language model dropped by roughly a hundred times between early 2023 and late 2025. Not a hundred percent — a hundred times. GPT-4 at launch cost sixty dollars per million output tokens. By the end of 2025, models of comparable or superior capability were available for under a dollar for the same volume.

I knew this. I tracked the pricing. I read the announcements. And I failed to internalize what it meant, because my mental model for technology cost curves was shaped by decades of gradual decline — Moore’s Law, storage costs, bandwidth costs — where the trajectory was predictable and the planning horizon was long. A hundredfold cost reduction in two years was not on any curve I had ever seen. It was not a trend. It was a discontinuity.

When the cost of intelligence drops by two orders of magnitude, every business model that depends on expensive intelligence breaks. This is not an opinion. It is arithmetic. I should have done the arithmetic in 2023. I did it in 2025. The delay cost me everything I described in the last chapter.

Signal: The capability curve.

Each generation of models was not incrementally better. It was qualitatively different. GPT-3.5, which had launched with ChatGPT in late 2022, could write passable text and answer questions. GPT-4, which arrived a few months later, could reason about complex problems, write functional code, and pass professional exams. By mid-2025, frontier models could operate as autonomous coding agents for hours at a stretch — reading codebases, writing and executing tests, debugging their own errors, and producing production-quality software with minimal human oversight.

The gap between “impressive demo” and “production-ready tool” closed faster than almost anyone predicted. Certainly faster than I predicted. I had told myself in early 2024 that we had “two to three years” before AI was ready for enterprise procurement workflows. I was wrong by at least eighteen months. The capability curve was not a line. It was a staircase, and each step was higher than the last.

Signal: Adoption by sophisticated users.

This was the signal that should have hit me hardest, because it spoke directly to my own world. When the best engineers at the best companies — the people with the deepest technical expertise and the least patience for hype — started restructuring their daily workflows around AI tools, that was not hype. That was a verdict.

By mid-2024, every serious engineering team I knew was using AI coding assistants. Not as toys. Not as experiments. As load-bearing tools in their development process. Engineers who had spent years mastering their craft — people who could have written the code faster by hand than by explaining what they wanted to an AI — were choosing the AI anyway, because the combination of human judgment and AI execution was producing better results than either alone.

Amateurs adopt hype. Experts adopt capability. When the experts adopt, the debate is over.

Signal: Revenue shifts.

By the second half of 2024, the financial data was unambiguous. Companies that had integrated AI deeply into their core product — not as a feature, not as a checkbox, but as a fundamental redesign of their workflows — were seeing measurable improvements in retention, conversion, and customer satisfaction. Companies that had not were seeing the opposite. The pipeline data at Arcline told this story clearly: our new-logo win rate was declining, and the deals we were losing were going to AI-native competitors. When the P&L starts moving, the debate is over. I had the P&L data. I debated anyway.

Those were the signals. Now let me describe the noise — the things that sounded important, consumed enormous attention, and were almost entirely irrelevant to the strategic decisions I needed to make.

Noise: The AGI timeline debates.

Will artificial general intelligence arrive in 2027? In 2030? In 2040? Never? This question consumed an extraordinary amount of intellectual energy between 2023 and 2025. Conferences were organized around it. Books were written about it. Careers were built on having strong opinions about it.

It was completely irrelevant to my situation.

Whether AGI arrives in five years or fifty years has no bearing on the strategic decisions an operating executive needs to make today. The AI that existed right now — not the hypothetical superintelligence, but the actual, shipping, available-via-API models — was sufficient to reshape my industry. Planning for a hypothetical future was a distraction from the very real present. And yet I spent hours reading AGI timeline debates, because they gave me the illusion of engaging with the AI transition without actually having to do anything about it. Intellectual entertainment disguised as strategic thinking.

Noise: The alignment panic.

I do not dismiss AI safety concerns. They are serious and they deserve serious research and serious attention from serious people. But for the operating executive — for me, trying to decide what to do with a six-hundred-million-dollar software company — the safety discourse was often noise. Not because the concerns were invalid, but because they were being used, consciously or unconsciously, as a reason to delay.

I watched executives at other companies cite “responsible AI” as a justification for doing nothing. They formed ethics committees. They commissioned risk assessments. They published AI principles. And while they were doing all of this very respectable, very well-intentioned work, their competitors were shipping AI-native products and signing their customers.

Responsibility and urgency are not opposites. You can move fast and move carefully. The executives who understood this did both. The ones who used responsibility as a synonym for paralysis are now explaining to their boards why they are two years behind.

I used it too. In late 2023, I raised “safety and reliability concerns” as a reason to slow our AI adoption. The concerns were real. My use of them was not entirely honest. I was wrapping hesitation in the language of responsibility, because hesitation sounds weak and responsibility sounds virtuous.

Noise: The mass unemployment narrative.

“AI will take all the jobs.” This prediction was everywhere in 2023 and 2024. It generated enormous media coverage, political attention, and public anxiety. It was also, as a description of what was actually happening, largely wrong.

Jobs were being transformed, not eliminated wholesale. The nature of work was changing — from production to supervision, from execution to evaluation — but the apocalyptic vision of mass technological unemployment was not materializing in any sector I could observe. What was happening was more subtle and more interesting: role compression. One person with AI could do what three people did without AI. This did not mean two people were fired. It often meant that two people were redeployed to work that had been deferred, neglected, or considered too expensive to pursue.

The mass unemployment narrative was noise because it encouraged a binary framing — AI replaces humans, or it does not — that obscured the more important question: how does the relationship between humans and AI actually work? The answer, which Klarna had learned the hard way and which I was about to learn myself, was that the relationship is a partnership, not a substitution. And designing that partnership well is the real strategic challenge.

I have described the signals and the noise, and I have admitted which ones I saw and which ones I used as cover. But there is one more part of this story that I need to tell, and it is the part that shames me most.

The Cassandras were right. And I ignored them.

There were two engineers at Arcline — I will call them Priya and Tomás — who saw the signal clearly, months before I did. Priya was a senior engineer on our data pipeline team. Tomás was a mid-level engineer in our integrations group. Neither of them reported to me directly. Neither of them had “AI” in their title or their job description.

What they had was curiosity and proximity to the work.

Starting in early 2024, without being asked, without any mandate or budget, Priya and Tomás had begun rebuilding Arcline’s internal tools with AI. They used Claude to generate data migration scripts that previously took days to write by hand. They built an internal chatbot that could answer questions about our codebase — “where is the retry logic for failed supplier API calls?” — by reading our repository and documentation. They prototyped a procurement anomaly detector that used a language model to flag unusual patterns in purchase orders, and it outperformed the rule-based system we had spent two years building.

They presented their work at an internal demo day in June 2024. I attended. I was impressed. I told them it was “really interesting” and that I would “find a way to incorporate it into our roadmap.” And then I did nothing. I went back to my Phase 1 slide and my careful experiments and my sophisticated procrastination.

Priya sent me a follow-up email a week later. “Vikram,” she wrote, “I think what Tomás and I built could be the foundation for a fundamentally different product. Not an add-on to the existing platform. A replacement for parts of it. I’d love to talk about this if you have time.”

I did not make time. I replied with a polite note about how we needed to be “thoughtful” about how we integrated AI into our core product and that I appreciated her initiative. It was the corporate equivalent of a pat on the head.

Priya was right. About everything. The tools she and Tomás had built in their spare time, with no budget and no support, were closer to Arcline’s future architecture than anything my official AI strategy had produced. They had seen the signal because they were close to the work — because they used the tools every day and could feel, in their hands, how the nature of their work was changing. They did not need a framework or a board meeting or a six-force analysis. They needed only to pay attention to what was happening in front of them.

Grove wrote about this phenomenon. He called it the knowledge of people “in the trenches.” The salespeople, the engineers, the individual contributors who interact with reality every day — they often see the inflection point before senior leadership does, because senior leadership is insulated by layers of abstraction, by dashboards and reports and quarterly reviews that smooth out the jagged edges of what is actually happening on the ground.

When I finally committed to the transformation in January 2025, the first thing I did was find Priya and Tomás and put them at the center of the new architecture team. I told Priya that I owed her an apology. She was gracious about it. She said she understood. But I could see in her face the frustration of someone who had been right and had not been heard, and I made a promise to myself that I would never again dismiss a Cassandra because their message was inconvenient.

If you are reading this and wondering how to distinguish signal from noise in your own industry, here is the only advice I have, earned at a cost I would rather not have paid:

Do not look at the technology. Look at the behavior.

When the best people in your organization start quietly changing how they work — not because they were told to, but because the new way is obviously better — that is the signal. When your most sophisticated customers start asking questions that your product cannot answer — not complaints, but expectations — that is the signal. When startups that did not exist eighteen months ago start winning deals that you used to win without trying — that is the signal.

Everything else — the conferences, the analyst reports, the vendor pitches, the timeline debates, the thinkpieces — is noise. Useful noise, occasionally. Entertaining noise, often. But noise.

The signal is in the work. It is always in the work. Go talk to the people doing it.