The AI Carry-Forward Test

Callable assets cross. Inert assets strand. And intellectual property is the transport layer through which everything else you own enters the AI economy.

Scott Farrell · LeverageAI · leverageai.com.au

Everyone I talk to about AI eventually arrives at the same quiet fear, usually somewhere in the second coffee. If AI is becoming the oracle — all-seeing, all-knowing, cheap for everyone — then what's the value of a thought? What's the value of thirty years of judgement, or an archive, or a workshop full of machines, or a licence it took a decade to earn? The fear has a respectable pedigree: when something becomes abundant, the things near it usually get cheaper.

I think the fear has the sign backwards. AI is repricing intellectual property and thought — but not downwards. If you can get your intellectual property down, callable, accessible and usable, AI can compound it for you. And the same repricing that lifts compiled thought quietly decides the fate of everything else you own. That's the part almost nobody is looking at: the boundary AI actually draws doesn't run between "knowledge businesses" and "physical businesses". It runs through the middle of every asset register, and most owners can't yet see where.

This article names that boundary, shows the mechanism for crossing it, and hands you the instrument: a seven-question test you can run on any asset you hold — a framework, an archive, a machine fleet, a licence — that tells you whether it crosses, what it needs to cross, or whether you should stop defending it and harvest.

The boundary is callable versus inert

Start with the boundary everyone already knows about, because it moved once before. In 1975, tangible assets — property, plant, equipment, inventory — represented 83% of the market value of the S&P 500. By the end of 2025, that relationship had completely inverted: intangibles now constitute roughly 92% of S&P 500 market capitalisation¹. Ocean Tomo calls it "economic inversion" — value migrated from what can be touched to what can be thought. And the flow is still accelerating: global investment in intangible assets crossed USD 10 trillion for the first time in 2025, growing almost twice as fast as tangible investment².

So "intangibles win" is not news. Here is the news: AI is now drawing a second boundary, and it runs straight through that 92%. Most of the intangible value on the planet — the judgement, the archives, the playbooks, the relationships — is inert. It sits in documents, slide decks and people's heads, waiting for someone to remember it exists, interpret it correctly and manually apply it. A machine cannot find it, cannot read it, cannot use it. As far as the AI economy is concerned, it isn't there.

AI creates a new asset boundary: assets that are callable, and assets that remain inert.

Callable means something specific. A callable asset is legible (a machine can read what it is and what state it's in), addressable (an agent can locate it at the moment it's needed), and invocable (it can participate in live work without its owner in the loop). An inert asset may be enormously valuable on paper. But the machines now doing a growing share of the world's cognitive work cannot see it, so it earns nothing from them.

Notice what this boundary is not. It is not physical versus intangible. I know a practice owner who wrote down every question her staff asked, and every answer she gave, for over ten years — hundreds of pages of pure, disciplined knowledge capture. Her staff still asked her the questions. The document was intangible, and utterly inert: she remained the retrieval system, because writing it down had made the knowledge exist without making it findable or usable³. Meanwhile, a lump of capital equipment with a maintained digital passport — identity, configuration, history, constraints, all machine-readable — is physical, and callable. The boundary ignores the accounting categories entirely.

The inert side of the boundary is where almost everything currently lives, and we can price it. Gartner estimates 70–80% of enterprise knowledge is tacit — never written down in any retrievable form⁴. Panopto's workplace study found 42% of institutional knowledge is unique to a single person: when they leave, their colleagues can't do 42% of that job⁵. Employees spend around 1.8 hours a day just searching for information — hire five people and only four show up; the fifth is off looking for answers⁶. Deloitte puts the annual cost of institutional knowledge loss in the US at about $1.3 trillion⁷. That is what inert judgement costs before AI arrives. After AI arrives, the cost changes character: it stops being friction and becomes forfeiture, because the inert asset is now locked out of the economy where the work is increasingly done.

The oracle inversion

Now to the fear itself, because the logic deserves to be dismantled rather than dismissed. The naïve oracle story runs: (1) AI will know everything; (2) everyone will have access to the same intelligence; (3) therefore an individual thought or body of expertise becomes less valuable. The trouble is that the first two premises actually produce the opposite conclusion.

When everyone rents the same increasingly capable models, model intelligence becomes shared infrastructure — and shared infrastructure cannot be the source of durable differentiation. This is now measurable. For a language model of equivalent performance, inference cost is falling roughly 10x every year⁸. The cost of GPT-3.5-level performance dropped over 280-fold in two years⁹. Open-weight models trail the closed frontier by an average of about three months¹⁰. Whatever a frontier release confers, it confers on everyone, at collapsing cost, almost at once. Capability is symmetric — and symmetry is the exact opposite of a moat¹¹.

So AI separates two kinds of thought and prices them in opposite directions. Generic, reproducible thought — the first-pass analysis, the standard report, the competent summary — becomes radically cheaper, because the shared oracle really does produce it for everyone. But distinctive judgement, private context, accumulated discrimination and opinionated frameworks become more valuable — provided, and only provided, they are made usable by machines.

The model is the multiplier. Your compiled judgement is the multiplicand.

A vastly more powerful multiplier does nothing for an organisation whose proprietary multiplicand is zero — generic, or trapped in people's heads. And each new model release pays its dividend to whoever has already compiled something for the better model to operate over; without a substrate to pour capability into, a cheaper, smarter model is just a cheaper, smarter chatbot¹¹. The consulting data agrees from the other direction: nearly 80% of organisations now use generative AI, yet only 39% report any enterprise-level EBIT impact¹² — and McKinsey's own account of where the exceptions come from is privileged, cumulative, closed-loop data and context that competitors cannot rent¹³. Everyone has the multiplier. Almost nobody has a multiplicand.

The oracle does not eliminate the value of having a worldview. It eliminates the value of leaving that worldview implicit.

The asset ladder: from a thought to compounding capital

"Get your IP down and callable" compresses several distinct steps, and the distinctions carry real money, because each step is where a different self-congratulation goes wrong. Treat "a thought", "IP" and "capital" as synonyms and you will believe you've built an asset when you've built a filing cabinet. They are stages:

Four hard distinctions fall out of the ladder — four ways of saying "not yet":

A thought is not yet capital. A thought is the most perishable asset a person owns; its half-life is measured in hours, and even when written down, what returns years later is a fossil you can cite but not re-enter¹⁴. Retention with provenance is the first rung, not the prize.

Capture is not yet compilation. The ten-year Word document is the controlled experiment: everything captured, nothing callable, and the owner still the retrieval layer³. Compilation is what capture is missing — canonical claims, resolved contradictions, version chains, relationships, a navigable map. It's the difference between a pile of receipts and a set of accounts.

Callability is not yet value. A callable framework nobody invokes is latent. Callability gives the asset potential energy; the economic event happens when it meets a live problem. This is why "we digitised our knowledge base" is not a result. It's a precondition.

Application is not yet compounding. A framework used repeatedly without evaluation is merely reusable. It compounds only when outcomes, corrections, exceptions and rejected paths flow back and improve the substrate the next piece of work loads. No write-back, no compounding — just reuse with better marketing.

A thought becomes intellectual property when it gains durable identity. It becomes productive capital when it becomes callable, changes real work, and learns from the result.

The unit of value is the collision

Where, exactly, does the money appear? Not when the file is created. The value of latent IP is driven by the probability that the right idea meets the right context — the right person, company, live problem and still-open window — before the moment closes¹⁵. You can encrypt the drive that holds your frameworks; congratulations, you have protected a cupboard. The value event is a collision, and there are two collision surfaces.

Internal collisions — thought meets thought. A compiled body of judgement lets ideas formed years apart, in different contexts, for different purposes, become co-present on one problem. That's not better memory; it's recombination. The arithmetic is striking: seventeen co-present thoughts allow 136 possible pairings (n(n−1)/2 — a fact about pairs, not a measured statistic), and each new thought added to the room introduces itself to everything already there¹⁶. Unaided working memory holds about four items; no thinking person, ever, has had their whole accumulated history present at the moment of a new problem. Now that room exists — and internal collisions produce new hypotheses, sharper distinctions, exposed contradictions, cross-domain transfers.

External collisions — thought meets reality. The second surface is where a compiled framework meets a person, a company, a live problem and an open opportunity window. That's where a matched idea becomes a changed decision, an offer, a product thesis, a company. A compiled canon is not a vault inventory; it is a matching surface¹⁵.

Chain the two surfaces and you get the engine: thought × thought produces new possibility; possibility × live reality produces economic consequence; consequence × evidence produces better future judgement — which flows back into the substrate and raises the value of every future collision. This is stronger than "knowledge management", and stronger than "reusing IP". It is collision engineering.

The unit of value is not the thought stored. It is the verified collision the thought enables.

Cognitive concurrency: thought's new economic property

Callability gives thought an economic property it has never had in human history. Historically, expert judgement had the throughput of the expert's calendar. However brilliant the framework, it was applied at the speed of the one person who remembered it, interpreted it and drove it — one serious problem at a time.

Once judgement is compiled into a machine-callable substrate, it becomes non-rival at the production layer. One agent invokes it for a consulting engagement; another tests it against an industrial distributor's problem; ten more run boundary cases in parallel; none of those invocations consumes the original or queues behind the others. Economists will recognise the property — Romer's ideas-as-nonrival-inputs is the foundation of modern growth theory¹⁷, and Haskel and Westlake's "scalability" (intangibles "can be used repeatedly and in multiple places at the same time") is the same property at asset level¹⁸. What's new is not the economics. What's new is the invocation mechanism: for the first time, a single firm's — even a single person's — private judgement can actually be exercised concurrently, because AI agents can call it the way software calls a function.

A human can personally inhabit one serious problem at a time. Callable judgement can participate in many problems at once.

Be precise about what gets cloned, because the overclaim is where this idea goes to die. Concurrency clones access to the selected, explicit parts of a person's discrimination — the parts that survived compilation. The non-callable remainder still matters enormously: taste, responsibility, relationships, politics, moral judgement, authority, the willingness to bear consequences. None of that parallelises. AI parallelises the reusable discrimination. It does not parallelise accountability.

So the expert's role doesn't disappear — it moves upstream and outward: maintain and improve the source judgement; choose which variables matter; resolve the true exceptions; control what gets promoted into the canon; hold authority and make the final call where stakes require it. The expert stops being the retrieval layer and becomes the editor, the governor, and the court of appeal.

The physical-asset correction — and the scarcity anchors

At this point the argument sounds like bad news for anyone whose balance sheet is made of steel, land and licences. My own first instinct ran that way: physical assets are slow, can't be recombined, can't be parallelised, can't be taken into the AI future very effectively. That instinct needs a correction, and the correction is where the most contrarian — and most valuable — part of the framework lives.

The boundary that matters is not physical versus intangible. It is inert versus addressable; fixed versus reconfigurable; isolated versus composable; undocumented versus machine-legible; static versus learning; generic versus scarce; consequence-free versus authority-bearing. Run those tests and physical assets split into two utterly different fates.

Some physical assets are about to appreciate. When cognition becomes abundant, the genuinely scarce things next to it become the binding constraint — and value migrates to the bottleneck. Watch it happening in real time across the AI build-out itself: the IEA describes "a scramble for electricity, grid connections, manufacturing capacity, chips and capital", with data-centre electricity consumption projected to roughly double from 485 TWh in 2025 to 950 TWh by 2030¹⁹. Grid connection queues in America's largest market average eight years — the queue itself is now the moat²⁰. Microsoft signed a twenty-year power purchase agreement to restart a retired nuclear reactor at Three Mile Island²¹. Advanced chip packaging, high-bandwidth memory and leading-edge foundry capacity are constrained through 2027²². Energy, grid access, land in the right place, manufacturing capacity, inventory, logistics, workshops, licensed operating authority, customer access: the mistake would be to assume that because these assets cannot be parallelised, they lose value. Quite often they become the new bottleneck to which value migrates. I call them scarcity anchors.

But an anchor only collects that migrating value if it can be joined to the cognitive economy. Which brings us to how a physical asset actually travels.

The cognitive twin: how a physical asset crosses

A physical asset cannot become callable itself — the steel does not become digital. What crosses is its cognitive counterpart: a maintained representation carrying stable identity, current state, configuration, operating history, constraints, relationships, compatible options, authority over who may do what, and an interface through which agents can inspect and propose action.

We built exactly this for a national capital-equipment distributor: a Machine Passport for every enrolled unit — serial, configuration, manuals, service history, inspections, parts relationships, contacts²³. The reframe is blunt: previously, the customer explained the problem before the supplier knew which machine was speaking; with passports, the system knows the machine before the customer starts talking. The machine's identity, history and possible actions became legible and callable — so the physical capacity could be recombined intelligently across customers, projects and time.

The industrial world is converging on the same conclusion from every direction. Organisations working with digital twins report about 15% improvement in sales and operational metrics and up to 25% in system performance²⁴. Nearly nine out of ten industrial companies call the digital thread essential — and the top driver they name for improving it is enabling AI²⁵. And regulators have arrived independently: under the EU's Ecodesign for Sustainable Products Regulation, a growing list of products must carry a Digital Product Passport — a machine-readable lifecycle record reached through a data carrier — making traceability "a gate on the right to sell", with the battery passport mandatory from February 2027²⁶. Read that carefully: in the EU's flagship market rules, a physical asset now literally cannot cross the market boundary without its cognitive counterpart.

The representation is not the territory. It is how the territory becomes addressable — and addressable is what the anchor needs to collect the value migrating toward it.

Trust, governance and licences: the hybrid assets

One class of assets refuses to collapse into either category, and it's worth being honest about why. Trust, governance, licences and institutional standing are hybrids. Their informational content compiles beautifully: rules, precedents, evidence, decision pathways, compliance history — all of it can be made legible, addressable and callable, and should be. But the economic asset is partly external to any representation: an institution's recognised authority; a licence granted by a regulator; a reputation built through past conduct; a party the law can reach; a balance sheet that can carry failure; customer permission accumulated over years.

AI can make those assets legible and operational. It cannot generate their legitimacy. You can compile the rulebook; you cannot compile the right to enforce it. Even where AI is being aimed at the machinery of authority — the US Department of Energy is using it to compress nuclear reactor licensing timelines²⁷ — the effect is to make the paperwork cheaper while making the licence itself more clearly the scarce asset. Which means the strategy for hybrid assets is both/and: compile the compilable layer (that's what makes the authority usable at machine speed), and recognise that the legitimacy underneath is a scarcity anchor — un-parallelisable, appreciating, and yours.

Thought is not the only asset that survives. Thought is the transport layer that makes the other surviving assets intelligible and usable.

The Carry-Forward Test: seven questions

Now the instrument. For every significant asset you hold — a framework, an archive, a fleet, a licence, a relationship, a brand — ask seven questions:

1. What scarcity currently produces its value? And is AI dissolving that scarcity, leaving it unchanged, or making it more intense? This is the question that separates doomed from durable, and most owners have never asked it explicitly.
2. Can the asset be made legible and addressable? Could an agent identify it, understand its state and locate the relevant evidence — without depending on the person who "just knows"?
3. Can it be invoked independently of its original holder? Can the judgement, history or capability participate in work without its creator reconstructing it each time?
4. Can it be recombined? Can it join new customers, problems, products, workflows or physical resources without a full manual rebuild?
5. Can it operate concurrently? Can it serve multiple situations at once — or is it constrained to one person, one location, one machine, one authority holder at a time?
6. Does use improve it? Does each application leave evidence, corrections, exceptions or new judgement that makes the next use stronger?
7. What irreducible scarcity remains underneath it? Physical capacity, trust, legal authority, relationships, capital, consequence — what stays valuable even after the cognitive layer is built?

The answers land every asset in one of four postures:

Readers of our earlier work will recognise the ancestry. The Three Asset Classes taxonomy — stranded, convertible, compounding — classifies assets by how their value behaves under improving AI²⁸, and The Re-Roll runs the same repricing at industry altitude, where every participant's character sheet gets re-rolled at once²⁹. What the Carry-Forward Test adds is the mechanism: what, precisely, allows this asset to cross from its current form into the next value architecture? Behaviour tells you what the asset's value will do. The test tells you what to do to the asset. And "scarcity anchor" is the posture the original taxonomy lacked — the contrarian recognition that the un-parallelisable can appreciate.

The restraint: callability is leverage, not truth

One honest boundary before the pitch for action, because the failure mode is already visible in the wild. Callability is leverage, not truth. AI can parallelise bad judgement exactly as efficiently as good judgement — a wrong framework, compiled and callable, becomes wrong at scale, with the confidence of infrastructure. The consulting graveyard of the next five years will contain a lot of beautifully compiled nonsense.

What converts scalable opinion into compounding capital is the unglamorous machinery around the substrate: provenance (where did this claim come from), contradiction (what disagrees with it, and how was that resolved), evidence (what happened when it met reality), authority (who is allowed to promote what into canon), and world-loop write-back (do outcomes actually flow back in). Question six of the test — does use improve it? — is doing more work than it looks. An asset that scales without learning is not compounding capital. It's amplified guesswork.

A working specimen

I'll close with the asset I know best, because I've walked it across the boundary myself and the transition is the proof the framework runs. My IP wiki is my best ideas — formulated, structured, compounded, written down — with an MCP server over the top, so they're callable: an active substrate an AI agent can query mid-task. The power is not just that they're good ideas. It's not just that they're written down. It's that they're callable — and their biggest value is when they collide with a real problem or a real customer. They can be invoked unlimited times, in parallel, at the same time, by AI. Run the seven questions on it: the scarcity it rests on (decades of accumulated discrimination) is one AI intensifies rather than dissolves; it is legible, addressable, invocable without me, recombinable, concurrent; use improves it, because engagements write evidence back in; and the irreducible scarcity underneath — taste, authority, the willingness to sign — stays with me. Natively portable, posture one, and compounding.

The specimen chain is the asset ladder run end to end: experience → thought → deliberation → frameworks → claims and edges → a callable MCP surface → live customer applications → new evidence and further frameworks. It's why a venture studio consuming that substrate can compile a serious thesis against a live industry problem in thirty days: the thirty days were never the thinking. The cognition was capitalised over decades; thirty days is the compile-and-contact-with-reality period.

We turned intellectual property from something you read into something that works.

What to do next

The oracle story told you your accumulated judgement was about to be worthless. The truth is stranger and far more actionable: your judgement is the one asset class that crosses the new boundary natively — and everything else you own crosses only by riding on it. The machines, the archives, the licences, the trust: each needs its cognitive counterpart written, structured and made callable before the value migrating in its direction can actually land on it.

So here is the homework, and it fits on one page. List your ten most significant assets — don't forget the invisible ones: the archive, the relationships, the licence, the founder's judgement. Run the seven questions on each. Assign each a posture: natively portable, convertible, scarcity anchor, stranded. Then act on the postures: compile the portable, convert the convertible (starting with the highest-value passport, not the easiest), anchor the anchors to a cognitive layer, and harvest the stranded — without sentiment, and without reinvesting in their decline.

All assets need the test of how to be taken into the future. Some have clearer paths than others. Intellectual property — we know the path. And it turns out to be the path everything else takes too.