Skip to content
AI-Native PM
Field Notes

Field Note

Strategy Used to Be a Document. Now It's a Discipline.

In late 2023, OpenAI shipped one ChatGPT update that turned a whole cohort of startups' only differentiator into a checkbox. This is what strategy has to become when the ground can move on the platform's schedule, not yours.

· 7 min read

Picture a founder in late October 2023, watching an OpenAI livestream the way you watch weather you cannot change. OpenAI tells the room that ChatGPT Plus can now take a file, read it, and answer questions about it. For most of the audience that is a convenience. For this founder it is the end of a company, because the one thing their product did, let you chat with a PDF, just became a built-in feature of the platform their whole business ran on top of. The teams behind a cohort of tools that had each raised money on exactly that capability, names like AskYourPDF, AI PDF, and ScholarAI, woke up to find their single differentiator demoted to a checkbox inside a product their own users already paid for. The press had a word for it within the day: they had been Sherlocked.

It matters less that this happened than how fast it happened. Kodak took the better part of two decades to lose the ground under film. Blockbuster had years to recognize the threat from mail-order DVDs and then streaming. This founder's differentiation did not erode over a long decline they could have managed around; it disappeared during a single keynote.

Strategy used to be a document you wrote once a year. Now it is a discipline you run every week, because a frontier lab can erase a year of your differentiation in an afternoon with a single model release.

This note is the field evidence for a claim we make in our essay The Discipline and then leave you to take on faith: that strategy is a set of deliberate exclusions, and that a frontier lab can vaporize your differentiation overnight with one model update. Here is that claim borne out in the open, with names and dates, and here is what the people who survived it did differently.

The annual plan worked while markets moved slowly

For twenty years, product strategy was an artifact. You wrote a document, usually annually, that named a three-to-five-year horizon, set quarterly checkpoints against it, and rested on defensibility you accumulated slowly: a feature lead measured in release cycles, an integration no competitor had built, a brand customers trusted. The document worked because the market moved slowly enough for you to execute the plan written on it.

That assumption is what broke. Categories that were stable for a decade can now be reshaped in a quarter, which means the strategic horizon got shorter at the same time the stakes got higher. An annual roadmap is obsolete before you finish presenting it, and a quarterly one is mostly obsolete by its own Q3. Planning did not stop mattering, but the document you save at the end is no longer the valuable part. Strategy became less like a blueprint you draw once and more like a thing you keep running, because the question is no longer "what is our three-year plan" but "is the plan we wrote last month still true this morning."

The disruption clock is collapsingFive horizontal bars, longest at top, shortest at bottom. Older incumbents took years to be displaced. AI-native cases close in months, weeks, or a single keynote. Caption: the window to erase your differentiation is collapsing.WHAT WAS DISPLACEDTIME TO DISPLACEMENTKodak, filmabout 20 yearsBlockbuster, rentalabout 7 yearsEdtech incumbent39 monthsModel wrapperweeks after a releaseDocument-chat cohortone keynoteYEARSAN AFTERNOONThe window to erase your differentiation is collapsing, on the platform's schedule.

The casualties, dated and varied

The PDF cohort is the cleanest single-keynote erasure, but it would be easy to dismiss as the fate of thin wrappers that were always going to get absorbed. So look at the mechanism across different kinds of products, each with real revenue and a real moat that turned out not to be one.

Jasper sold AI writing and marketing copy, and in October 2022 it raised at roughly a 1.5 billion dollar valuation, an early generative-AI winner with paying customers and momentum. Then OpenAI launched ChatGPT a month later, a free, general-purpose competitor that did much of what Jasper charged for. By September 2023 the company had cut its own internal valuation by about twenty percent, run layoffs over the summer, revised its forecast for the year down by roughly a third, and appointed a new chief executive. None of that came from a Jasper mistake; OpenAI had shipped a platform release, and within months the thing Jasper had been selling was a commodity.

The same thing happened to Chegg, an entrenched incumbent rather than a young startup. The homework-help company was worth about 14.7 billion dollars at its February 2021 peak, with a subscription business that looked like a durable moat: a decade of content, a brand students reached for by name. After OpenAI launched ChatGPT, students substituted toward it, and later toward Google's AI Overviews, and the moat dissolved. Chegg Services subscribers fell to about 3.6 million by the end of 2024, revenue slid from a 776 million dollar peak toward a roughly 290 million dollar run-rate, and the stock dropped near a dollar with delisting notices from the exchange. Analysts described it as among the fastest documented cases of generative-AI disruption, a span measured in a few years rather than the long declines of the companies it now gets compared to. Chegg's own chief executive named both culprits, saying Google's AI Overviews was as material to the decline as ChatGPT itself, which is the uncomfortable part: two other companies set the schedule for his decline, and he controlled neither.

Then there is the platform owner saying the quiet part out loud. On the 20VC podcast in April 2024, Sam Altman told founders that companies built as a thin layer over GPT-4 risk being steamrolled by the next model, "not because we don't like you, but because we have a mission." When OpenAI does what he called its fundamental job, making "the model and its tooling better with every crank," some founders get what he named "the OpenAI killed my startup meme." He framed it as a bet each builder is implicitly making: either you are wagering the model stays roughly where it is and you fill the gaps, or you are building something a stronger model makes better rather than redundant.

The common thread is not bad luck: each of these had a genuine product and genuine revenue, and none of them had an exclusion sharp enough to survive a platform update. They were differentiated by a capability, and a capability is exactly the thing a frontier lab can hand to everyone at once with a single model release.

The other half of the rule: the survivors chose what to refuse

Our essay's claim is that real strategy is deliberate exclusion, that a strategy which precludes nothing is just a wish. The casualties show what happens without that exclusion, and the survivors show it paying off.

Jasper appears twice in this story because the same company that took the platform shock also lived through it, and it did so by cutting, not adding. Instead of chasing the consumer market it was losing to a free chatbot, it walked away from that segment and bet the company on a single buyer, the enterprise marketing team. It brought in a new chief executive from Dropbox, and it acquired Stability AI's Clipdrop, closed in February 2024, to deepen that one bet rather than broaden its product. By late 2024 it reported more than 850 enterprise customers with enterprise revenue roughly doubling year over year, and it later reported enterprise ARR tripling in the year ending February 2025. Jasper recovered not through a clever feature but by refusing a market.

Document versus disciplineTwo panels. On the left, a single static page icon labeled annual strategy document with a one-way arrow to execute, a snapshot. On the right, a continuous loop of five steps, diagnose, exclude, ship, watch, adjust, drawn with a gap so it never closes. Caption: a document is a snapshot, a discipline is a loop you never stop running.DOCUMENTannual strategydocumentEXECUTEthen waitDISCIPLINENEVER CLOSESdiagnoseexcludeshipwatchadjustEVERYWEEKA document is a snapshot. A discipline is a loop you never stop running.

Cursor, the AI code editor from Anysphere, shows the same thing from the growth side. While the wrappers were disappearing, it became the fastest software product to reach 100 million dollars in annual recurring revenue, in roughly a year, and it did it by refusing to be a general assistant. It went deep on one job, reasoning across a whole repository and making agentic multi-file edits inside the editor, the kind of thing a frontier lab cannot casually reproduce with a horizontal model update, because it lives in the product itself, not just in a capability. In both cases, focus is the moat itself rather than the consolation prize you accept after the big idea fails, because what you refuse to build is the part a platform owner cannot copy from you with an update.

Strategy is a control loop, not a saved file

A document is a snapshot, true at the moment you saved it and decaying from there. A discipline is a control loop, something that runs continuously and corrects itself against the world. The whole change we are asking you to make is to move from the first to the second.

This is how we practice it: keep a small set of load-bearing assumptions written down, the few beliefs your strategy actually rests on. For each one, name the leading indicator that would tell you it just broke, the early signal rather than the lagging revenue figure you only get a quarter too late. Set a refresh cadence measured in weeks, not an annual offsite. The job is not to write the three-year plan and execute against it with discipline, but to build the muscle to re-choose quickly when the market shifts under you, which it now does on the platform's schedule, not yours. Strategy used to be about choosing where to compete and then defending that position for years. Now it is about choosing where to compete and building the muscle to re-choose, fast, the moment an assumption you wrote down stops being true.

Putting this into practice

Day to day, the discipline comes down to a few habits. Keep your strategy on a single page, structured around the five questions in The Discipline, and treat the last one as the real test: what does this strategy preclude us from doing? If the answer is "nothing," you have written a wish, not a strategy. Underneath each answer, list the assumptions it rests on, and beside each assumption, the leading indicator that would tell you it broke.

Then run every roadmap item through one gate before you give it a place on the page.

The exclusion test: moat or exposureOne box, a thing we could build, flows down into a single question: does this strengthen something only we can build? A yes branches left to a moat node, build. A no branches right to an exposure node, cut. Caption: what a frontier model can reproduce in a quarter is not strategy, it is exposure.ROADMAP ITEMa thing we could buildTHE QUESTIONDoes this strengthen somethingonly we can build?YESNOMOATbuild ita platform update cannot copy thisEXPOSUREcut ita frontier model reproduces itWhat a frontier model can reproduce in a quarter is not strategy. It is exposure.

The question is the moat question from the essay, asked of one feature at a time: does this strengthen something only we can build, our proprietary data, our distribution, our taste, our speed of iteration, our ecosystem position? If it does, it is strategy, and it belongs on the page. If instead it is a capability a frontier model could reproduce within a quarter, it is not strategy but exposure, and the casualties above are what exposure looks like when a lab ships the next model.

Where this goes next

This note is the field evidence for one stretch of our essay The Discipline, the parts on product strategy and product-market fit, where the five-question template and the full moat list live, and where we make the case that PMF is a state you can lose rather than a milestone you cross. Read this as proof, with names and dates attached, that the claim has already come true.

For the operating muscle underneath it, the part of The Builder's Stack on building an eval suite is the closest companion: the leading indicators that tell you an assumption broke are exactly the kind of signal a real evaluation harness is built to catch before your users feel it. Strategy as a control loop only works if you have something watching the loop.

Sources