The Absorption Test: Will AI Absorb Your Business?

OpenAI shipped GPT-6 Astra this month and told everyone we had entered the AGI era. The headline capability was not another jump on a math benchmark. It was computer use: a model that opens your software, reads the screen, clicks the buttons, fills the forms, runs the checks, and models a house in Blender before walking you through it in a game engine. For a lot of founders that demo was not exciting. It was a gut check, because the thing on screen was doing the exact job their product charges for.

Here is the durable version of that gut check, the part that will still be true long after this month’s model is old news. Every few months the frontier reaches out and swallows a capability that used to be a business. It does not care how clever your prompt chain was or how nice your interface is. If your value lived inside the job the model just learned to do, your value is now a feature the model gives away. If it lived somewhere the model cannot reach, you just got a faster engine for free.

So the question that matters is not the one everyone asks. Everyone asks, can AI do what I do. Wrong question, because the answer is trending toward yes for almost everything. The question that actually predicts whether you have a business in two years is quieter and meaner. If the frontier did your core task perfectly and for free tomorrow, would anyone still pay you? I call that the Absorption Test, and this post is about how to run it on your own company, why most founders run the wrong test, and where value actually goes when a layer gets eaten.

Table of contents

• What absorption actually is
• The test most founders run, and why it lies
• The Absorption Test and the Absorption Map
• Where the profit goes when a layer is eaten
• The five things the frontier cannot absorb
• Why a better model can grow your market
• Running the test on your own business
• Four moves when the test comes back bad
• The contrarian take
• What to do Monday morning
• FAQ

What absorption actually is

Start with the mechanism, because the word gets used loosely. Absorption is when a capability that used to require a separate product becomes a native feature of the model itself. Not competed with. Absorbed. The model ships an update, and the thing you sold is now a checkbox inside the thing your customer already pays OpenAI or Anthropic or Google for.

We have watched this happen on a roughly quarterly clock. Early prompt-tip businesses were absorbed the moment models got better at following plain instructions. Basic retrieval wrappers were absorbed when long context and native file upload arrived. The little tools that summarized a PDF or cleaned a spreadsheet were absorbed when the chat products did it inline. Astra’s computer use is just the newest and largest bite. It aims squarely at the entire layer of products whose pitch was we click around software for you, and it took that bite in a single launch.

The numbers around the wrapper wave tell you how brutal the pattern is. By most accounts, somewhere north of 60 percent of thin AI wrappers make no revenue at all, only a few percent ever cross ten thousand dollars a month, and a large share of the 2024 and 2025 cohort is expected to be gone by the end of this year. That is not a failure of hustle. Many of those teams shipped fast and marketed well. They built on a patch of ground the frontier was always going to pave.

The reason this feels new is that in normal software, your competitor is another company that has to find you, target you, and out-execute you. In the AI stack, your most dangerous competitor is the supplier you depend on, and it does not need to target you at all. It improves its model for its own reasons, and your business is collateral. Marc Andreessen put the strange economics of it plainly this year: AI gets more expensive to compete in as it commoditizes, because the model itself stops being the prize and the fight moves to everything around it.

The test most founders run, and why it lies

Almost everyone runs the capability test. They watch a launch, they feel the cold sweat, and they ask can the model do my thing now. Then they either panic because it can, or they relax because it cannot do it quite as well yet. Both reactions are traps, and they are traps for the same reason.

The capability test measures the wrong distance. It measures how close the model is to your feature. What you actually need to measure is how much of your value survives the model closing that gap. Those are different questions with different answers. A tax product does not lose to a model that can do arithmetic, even though arithmetic is most of what it looks like it does. A generic chat-with-your-docs tool loses to a model that can chat with docs, even though its team is far more talented than the tax company’s.

The difference is not capability. It is what the customer is really buying. If the customer is buying the capability, absorption kills you the day the model matches it. If the customer is buying something wrapped around the capability that the model cannot ship, you are fine, and a better model actually helps you. So the honest test is not can it do my thing. It is this: if it did my thing perfectly and for free, what would I still be selling? If the honest answer is nothing, you do not have a company. You have a countdown.

This is the same error I wrote about with the service-as-software trap, where founders confuse selling a task for selling an outcome, and it rhymes with the data moat test. Absorption is the sharper, more current version of that same family of questions, because the frontier is now moving fast enough that the countdown is measured in quarters.

The Absorption Test and the Absorption Map

Here is the test in one line. Imagine the frontier model does your core task perfectly and for free tomorrow. Now list, in concrete terms, what customers would still pay you for. That surviving list is your actual business. Everything else was rented from the model vendor, and the rent is going to zero.

To make it usable, map any product on two axes. The first axis is where your value sits: inside the job the model does, or above it. The second axis is whether that value can be copied by a model update, or whether it is anchored in something the model cannot ship. Those two questions produce four quadrants, and knowing which one you are in tells you exactly how worried to be.

The bottom left is Absorbed. Your value is a capability the model has, and a model update copies it. This is where the dead wrappers live. The bottom right is Thin Layer. You sit above the model’s raw job, but what you added is a skin anyone can rebuild in a weekend, including the vendor when it decides to ship a nicer default. The top left is Borrowed Edge. You genuinely do the job better today, maybe because you have a small data lead or a clever pipeline, but it is inside the model’s lane and the model is walking up it. That edge has an expiry date, and the only smart move is to convert it into something anchored before it expires.

The top right is the Adjacent Moat, marked with a star, and it is the only quadrant that is actually safe. Your value sits above the model’s job, and it is anchored in something a model update cannot ship: a body of proprietary data your own usage generates, a workflow you have become the system of record for, a regulated liability you are willing to carry, a distribution channel the vendor does not own, or a switching cost you have earned. The whole game is to notice which quadrant each of your product lines is in, and to keep dragging the important ones toward the top right on purpose, because drift is always toward the bottom.

Where the profit goes when a layer is eaten

The most useful thing I know about absorption is that the profit does not vanish when a layer gets commoditized. It moves. Clayton Christensen named this the law of conservation of attractive profits back in 2003, long before any of this, and it has held up across every technology shift since. When one stage of a value chain becomes modular and cheap, the ability to earn attractive profit does not leave the system. It migrates to an adjacent stage that is still proprietary, still interdependent, and still hard.

You have watched this movie before even if you did not know its name. In the early PC industry the money was in hardware. Hardware commoditized, and the money moved up to the operating system and applications. Software commoditized in many categories as open source spread, and the money moved again, toward data and distribution and the integrated experience. At each step the people who died were the ones standing on the layer that was being commoditized, and the people who won were one layer over, holding the thing that was newly scarce.

Apply that lens to the AI stack and the picture gets clear fast. The model is the layer being commoditized right now, on purpose, by the people who make it. Its price is falling and its capability is rising at the same time, which is the exact signature of a stage becoming modular. Christensen tells you not to stand there. The attractive profit is migrating to the adjacent stage, and the adjacent stage in this shift is everything the model touches but cannot own: the proprietary data a product generates as a byproduct of being used, the workflow it becomes the official record for, the distribution that puts it in front of buyers, and the accountability someone has to sign for. Absorption and conservation of profit are the same event seen from two sides. The layer that gets absorbed is the layer whose profit is leaving. Your job is to already be standing on the layer it is leaving toward.

The practical trap is that the migration is obvious in hindsight and invisible in the moment. Nobody rang a bell when the money left PC hardware, and nobody is ringing one now as it leaves the raw model call. You feel it as margin pressure, as customers asking why they should pay for something the chat product almost does, as a competitor undercutting you with the same underlying capability. Those are not separate problems. They are the sound of a layer commoditizing, and they are the signal to move before the move is forced. The founders who read that signal early get to choose their adjacent layer deliberately. The ones who wait get whatever layer is left after everyone else has claimed the good ground.

The five things the frontier cannot absorb

Absorption has limits, and the limits are not about capability. They are about ownership and accountability. A model update can copy any skill. It cannot copy a fact about the world that only your usage knows, and it cannot volunteer to be sued. Those two facts generate a short, specific list of things the frontier structurally cannot eat, no matter how smart it gets.

The un-absorbable asset	Why a model update cannot ship it
Proprietary data from use	The data only exists because customers use you. The vendor was not in the room.
Workflow system of record	Being the official place work happens is a position, not a feature. It has to be earned account by account.
Regulated liability surface	Someone has to be accountable and insurable. A model will not sign the contract or carry the risk.
Distribution you own	Attention, trust, and a channel to real buyers are not in the training set. The vendor sells the engine, not your relationships.
Earned switching cost	Integrations, history, and habit make leaving expensive. A better model does not reset a customer’s sunk investment in you.

Notice what these have in common. Every one of them lives above the model’s job and is anchored in the real world rather than in the weights. Proprietary data from use is the strongest of the five because it compounds. The more the product is used, the more it knows, and that knowledge is not in any training set because it did not exist until your customers created it inside your walls. This is the same asset I unpacked in the data moat test, and it is worth being honest with yourself about, because most founders claim a data moat they do not have.

The regulated liability surface is the most underrated, and it is why AI landed in regulated work first. In medicine, law, accounting, and finance, the value is not only doing the task. It is being the accountable party who signs for it. A model can draft the filing. It cannot be the entity the regulator holds responsible when the filing is wrong. That accountability is a business, and it is one the frontier cannot absorb by getting smarter, because getting smarter is not the same as being liable.

Why a better model can grow your market

Here is the part that flips the fear, and it comes from a nineteenth century observation about coal. William Jevons noticed that when steam engines got more efficient, England did not burn less coal. It burned far more, because efficiency made coal useful for so many more things that total demand exploded. Cheaper units, wildly more usage. Economists have watched the same thing happen every time a core input gets cheap.

Cheap, capable models are the same story. When the marginal cost of the underlying task falls toward zero, demand for everything built on top of it does not shrink. It expands, often by orders of magnitude. The classic small example is the spreadsheet. When VisiCalc drove the cost of a financial calculation to nearly nothing, companies did not do less financial modeling. They did vastly more of it, and the demand for analysts who could exercise judgment over all those new models went up, not down. The cheap thing multiplied the market for the judgment sitting above it.

That is the upside of absorption, and most founders miss it because they are staring at the thing being eaten instead of the thing being multiplied. If you sit in the Adjacent Moat, a stronger frontier is not your enemy. It is a free upgrade to your engine and a bigger market for your anchored value at the same time. The businesses that own proprietary data, own the workflow, and own distribution get to ride every model improvement as pure tailwind, because the vendor is spending billions to make their complement better and cheaper. I made a version of this argument about cost in the pricing under cheap inference piece. The same force that destroys the thin layer subsidizes the thick one.

So the frontier is not simply a threat. It is a sorting machine. It punishes everyone whose value was inside the model’s job and rewards everyone whose value was above it, and it does both harder every quarter. Whether Astra is a threat or a gift to you is not a fact about Astra. It is a fact about where your value sits.

Running the test on your own business

Enough theory. Here is the test as a thing you actually do, per product line, with a pen. The mistake is running it on the company as a whole, because most companies are a bundle of one durable thing and several absorbable ones, and the average hides the truth.

Walk each product line through the tree. First question: if the frontier did the core task perfectly and free tomorrow, would this line still get paid? If no, it is Absorbed. Stop defending it and either move its value up or wind it down, because you are spending effort protecting a countdown. If yes, ask the second question: what specifically are they still paying for, and is it anchored in one of the five un-absorbables? If it is not anchored, you are a Thin Layer, and your job this quarter is to convert that soft advantage into a hard one before the vendor ships the default. If it is anchored, you are in the Adjacent Moat, and your job is to widen it and to treat every model release as a gift.

The output of this exercise is usually uncomfortable and always useful. Most founders discover that the feature they demo, the one they are proudest of, is the Absorbed one, and the thing that actually keeps customers is something boring they never talk about: the integration that took a year to build, the data history nobody wants to re-enter, the fact that the whole team already lives in the tool. That boring thing is the business. The demo was the marketing.

Four moves when the test comes back bad

Suppose you ran the test honestly and a real product line came back Absorbed. Not a fringe experiment, but something that pays salaries today. Panic is not a plan, and neither is pretending the launch you just watched did not happen. There are only four real moves, and the sooner you pick one deliberately, the more of the value you keep.

The first move is to go up a layer. If the model absorbed your capability, ask what job sits directly on top of the one it now does for free, and go own that. The team that sold document summarization moves up to owning the decision the summary feeds, the record of what was decided, and the accountability for acting on it. This is conservation of attractive profits applied to your own roadmap. You are deliberately abandoning the layer whose profit is leaving and planting yourself on the one it is arriving at. The hard part is emotional, because the absorbed layer is usually the thing you are known for, and going up means letting the frontier have the demo while you quietly take the part that pays.

The second move is to become the distribution. If you cannot get above the capability, you can sometimes own the path to the customer so completely that it no longer matters that the capability is a commodity. A commodity sold through a channel you own is a fine business, and the frontier does not ship your relationships, your brand, or your place in a buyer’s routine. This move only works if the distribution is genuinely yours and hard to route around, not a paid ad channel you rent from a platform that can also absorb you. If your distribution is itself borrowed, you have simply moved the absorption risk one supplier over.

The third move is to sell into the vendor’s blind spot. The frontier labs optimize for broad, horizontal capability. They are structurally uninterested in the narrow, ugly, regulated corners where the value is being the accountable party rather than being clever. A business built on carrying liability, meeting a specific compliance regime, or integrating with the one legacy system nobody at a lab will ever touch is a business the frontier can make more capable without making more competitive. The blind spot is not a place the model is weak. It is a place the vendor will not go because the market is too specific to be worth its attention, and that specificity is your cover.

The fourth move is the one founders resist most, and it is often the right one: exit the line early, while it still has value. An Absorbed capability is worth the most the day before everyone agrees it has been absorbed. If a product line has no path up, no distribution to hide behind, and no blind spot to occupy, the disciplined thing is to harvest it, fold its customers into a line that does have a moat, or sell it while the multiple still reflects a future that is not coming. Spending two more years defending a countdown is how good teams turn a soft landing into a hard one. Calling the absorption early is not defeat. It is the same judgment that let you see the opportunity in the first place, pointed at your own portfolio.

Most companies will use more than one of these at once, because most companies are a mix of quadrants. The point of naming the four is that it turns a vague dread into a decision. When a launch absorbs a line, you are not choosing between panic and denial. You are choosing which of four moves fits that specific line, and you are choosing it on purpose while you still have a hand in the outcome rather than after the market has chosen for you.

The contrarian take

Now let me argue against the neat version of my own framework, because the map is too comforting if you stop here. The first honest objection is that above the layer is not a permanent address. The frontier’s reach moves up over time. What is a safe adjacent moat this year can be the absorbed layer in two years, because the vendors keep climbing. Workflow orchestration felt un-absorbable until models started doing multi-step computer use. So the top right quadrant is not a fortress you reach and rest in. It is a position you have to keep re-earning as the absorption line rises underneath you. The test is not something you pass once. It is something you re-run every launch.

The second objection cuts the other way. Being absorbed is not automatically fatal, and treating it as fatal makes founders too precious about defensibility. If you own the distribution, you can sell an absorbable capability quite happily, because your moat was never the capability. It was the fact that you are how the customer reaches it. Plenty of durable businesses sell a commodity through a channel they own, and a commodity delivered through owned distribution is a fine business. The mistake is only fatal when the absorbable capability is the entire reason anyone showed up.

There is a subtler trap too, and it is the mirror of the capability panic. Some founders decide the answer is to stay as far from the model as possible, to be pure workflow and data and never touch the frontier’s fast-moving edge. That over-corrects. The Jevons upside only reaches you if you actually ride the improving model as a complement. Hide from it and you keep your moat but forfeit the tailwind, and a competitor who has the same moat plus the frontier engine will pass you. The goal is not distance from the model. It is the right relationship to it: your value above, its capability underneath, pulling you forward.

So the law I keep landing on is narrower and more demanding than a tidy quadrant. The frontier does not kill businesses. It absorbs capabilities, and it does so on a clock that keeps speeding up. Ask not whether AI can do what you do, because soon it can do almost anything you do. Ask whether anyone will still pay you after it does, keep dragging your value above the rising absorption line, and treat every model launch as the test being administered again whether you scheduled it or not.

What to do Monday morning

None of this matters unless it changes what you do this week. Here is the practice, and it is deliberately concrete.

List your product lines and run the test on each one. Do not average. For every line write the one-sentence answer to what would customers still pay for if the model did the core task free. Put each line in a quadrant on the map. You will likely find one true Adjacent Moat carrying several Absorbed and Thin passengers.

Name your one anchored asset out loud. If you cannot finish the sentence customers cannot easily leave because, in terms of data, workflow, liability, distribution, or switching cost, then you have found the real work. Building that anchor is now the roadmap, above whatever feature you were about to ship.

Instrument the data-from-use loop. If any part of your value is proprietary data, make sure you are actually capturing and compounding it as a byproduct of normal usage, not hoping it accumulates. The compounding is the moat. An uncaptured data exhaust is just exhaust.

Convert one Borrowed Edge before it expires. Pick the product line where you are ahead today only because you are inside the model’s lane, and this quarter turn that temporary lead into something anchored, or plan its exit. A lead you do not convert is a lead the vendor collects.

Re-run the test on every frontier launch. Put a standing entry on your calendar. When the next model ships, do not read the benchmark thread. Re-run the Absorption Test on each line and see what just moved from safe to absorbed. This is the same discipline behind keeping your opportunity map current, and it pairs with protecting the incompressible core of what only you can do. Deciding what to build next is itself the skill I called the initiative problem, and absorption is where that skill earns its keep.

If you want the wider frame this sits inside, it is the whole point of the AI-native founder playbook: build so the frontier is your engine, not your competitor.

FAQ

What is the Absorption Test? It is a one-line check on whether AI will destroy your business. Imagine the frontier model did your core task perfectly and for free tomorrow, then list what customers would still pay you for. That surviving list is your real business. If the list is empty, your value has been absorbed and you have a countdown rather than a company.

How is it different from asking whether AI can do what I do? The capability question measures how close the model is to your feature, which is the wrong distance. The Absorption Test measures how much of your value survives the model closing that gap. A tax product survives a model that can do arithmetic. A generic summarizer does not survive a model that can summarize. Capability is not the moat. What survives capability is.

What can the frontier not absorb? Five things, because they live above the model’s job and are anchored in the real world rather than in the weights: proprietary data your own usage generates, a workflow you are the official system of record for, a regulated liability someone has to sign for, distribution you own, and switching costs your customers have already paid. A model update can copy a skill. It cannot copy your usage data or volunteer to be liable.

Are all AI wrapper businesses doomed? No. A thin wrapper whose only value is a prompt over one API call is doomed, and most of that cohort is expected to be gone by the end of the year. A thick business that happens to use a model, with proprietary data, workflow lock-in, and owned distribution, is just a normal software company with a great engine. The word wrapper hides the only distinction that matters, which is whether your value sits inside or above the model’s job.

Where does the money go when a layer gets commoditized? It migrates one layer over. Christensen’s law of conservation of attractive profits says that when one stage of a value chain becomes modular and cheap, attractive profit does not disappear, it moves to an adjacent stage that is still proprietary and hard. In the AI stack the model is being commoditized on purpose, so the profit is migrating to the data, workflow, distribution, and liability layers around it. Stand where the profit is going, not where it is leaving.

Does a stronger model help or hurt me? It depends entirely on where your value sits. If you are inside the model’s job, a stronger model absorbs you. If you are above it, a stronger model is a free upgrade to your engine and, by the Jevons effect, a bigger market for the judgment and integration you sell on top. The same frontier launch is a threat to one company and a tailwind to another, and the difference is not the launch.

Is a position in the Adjacent Moat permanent? No. The absorption line keeps rising as vendors climb, so a safe adjacent position this year can be the absorbed layer in two years. The Adjacent Moat is not a fortress you reach and rest in. It is a position you re-earn on every model launch by dragging your value further above the rising line.

What should I actually do this week? Run the test on each product line separately rather than on the company as a whole, name the single anchored asset customers cannot easily leave, make sure you are capturing the proprietary data your usage generates, convert one temporary lead into something anchored, and put a recurring reminder to re-run the whole test every time a new frontier model ships.