The Absorption Frontier

Abstract

Every transformative technology has succeeded by making a scarce factor of production abundant. However, its benefits only arrive downstream of invention—once society has reorganised itself to take advantage of that abundance. AI will make intelligence abundant. To realise its potential, society must reorganise itself around what this makes possible. The AI absorption frontier for a sector is the point at which it has reorganised itself sufficiently to convert abundant intelligence into productive gains. Crossing it requires three things: data must be legible to AI, actionable, and outputs generated must be verifiable against ground truth. Of these, verifiability is hardest. Software has already crossed the absorption frontier through version control systems, interoperable interfaces, and automated testing built well before AI arrived. Other sectors will need to deliberately achieve this to derive value from this technology. AI supremacy, therefore, will depend not on who builds the most advanced models, but on which sectors—and ultimately which nations—can absorb AI.

Introduction

Every transformative technology has succeeded by making a scarce factor of production abundant.¹ Electricity made motive power abundant by making it divisible and capable of being located in any quantity, at any point. The shipping container collapsed the cost of moving goods so that geography was no longer a constraint on where things are made and where they could be consumed.² Computing made calculation abundant, and the internet did the same for communication.

The benefits of these transformations lay not in how they made things cheaper and more effective, but in what became possible once society had reorganised itself around this new abundance.³ Every significant advance in modern civilisation has occurred downstream of such a removal of scarcity.⁴

Intelligence has, until now, existed only in human heads.⁵ It has been expensive to produce (20 years to make a doctor or a lawyer), indivisible (a great engineer cannot be in multiple places at once), and incapable of being transferred (a master’s expertise dies with them). As a result, it has been clustered in cities, institutions, and rich countries where it could be nurtured, compounded, and scaled.⁶

AI will make intelligence abundant. In doing so, it will transform civilisation just as other powerful technologies have before it. The precise effects of this are, as yet, unknowable; just as few could predict that computers and digital technologies would eventually follow from the discovery of electricity. One thing, however, is certain: the benefits of AI will only materialise once society is able to reorganise itself to take advantage of what this new abundance makes possible.

This report is about what it will take to do that. It describes the changes needed to absorb an abundance of intelligence and what individuals, organisations, industries, and nations have to do to enable it. It describes this in terms of the absorption frontier—the line that separates those who can absorb what AI makes possible from those who cannot. Crossing it requires three things: data that is legible to AI, systems that are actionable by it, and outputs that are verifiable against ground truth. This report argues that crossing that line should be our most important objective.

Absorption of Technology

Azeem Azhar, a technology analyst, has produced the first bottom-up reconstruction of AI spending across the entire stack. He estimates that the generative AI economy generated US$110 billion in revenue over twelve months ending in June 2026—an annual run rate, stripped of double-counting and excluding China, of more than US$175 billion. That is roughly three times the growth rate of the mobile or internet revolutions before it.⁷

Despite this staggering growth, the average person and business are still largely untouched by AI. According to the Massachusetts Institute of Technology’s (MIT) NANDA project, of the US$30–40 billion poured into AI enterprise pilots, only one in 20 produced measurable value.⁸ By the first half of 2026, AI had shifted from “too cheap to meter” to “too expensive to afford.” Microsoft has clamped down on internal AI licences,⁹ Uber reportedly burned through its annual AI coding tools budget in four months,¹⁰ and firms that had cut headcount in anticipation of automation discovered that AI costs more than the people it replaced.¹¹

The Gap Between Invention and Impact

This lag between invention and impact is not unique to AI.¹² Even though the electric light bulb was patented in 1880, at the turn of the century, barely three percent of American homes had electric lighting, and 20 years later, less than half of US industrial capacity had been electrified. In his history of electrification, Warren Devine explained this delay by pointing out that power accounted for only three percent of a manufacturer’s costs.¹³ A factory that swapped its steam engine for an electric motor saved little and transformed nothing. It was not until factories completely redesigned their processes around what electricity enabled that the gains finally arrived.

The real work lies not in the invention of a new technology but in rebuilding the world to absorb it.

Absorption of AI

To bring about a transformation, every new technology has to operate on a specific substrate. In the case of electricity, this was machines and mechanisation. Factories of the late nineteenth century were designed around a central drive shaft—an iron rod powered by a single steam engine—that supplied the energy every machine in the factory used to run. When these factories were electrified, the steam engines were replaced by electric motors. Since all this did was replace one source of power with another, the gains were marginal.

It was not until industrialists realised that energy could be supplied to wherever the machines were located that the factory floor was rearranged around the logic of the manufacturing process rather than the geometry of the drive shaft.¹⁴ It was then that the productivity gains of electrification finally appeared.

New technologies take a while to produce transformative outcomes. Not only do they have to overcome the path dependence of doing things “the way they have always been done,”¹⁵ but real benefits only come after society has realigned itself to absorb them.

The point at which an economy has reorganised itself sufficiently to unlock productivity gains from a new technology is the absorption frontier. Countries behind it will struggle to reap the benefits of that technology, while those that have crossed it will unlock benefits far beyond what they could have foreseen.

So, where does the AI absorption frontier lie?

The Substrate of Intelligence

For data to deliver results, intelligence must be applied to it. To diagnose what the person is suffering from, a doctor is needed to analyse the relevant patient data. To build a bridge or a building, an engineer is needed to process data and generate the structural drawings needed to build it safely. If these results can only emerge from the brains of the experts who currently produce them, the constraint on intelligence will forever be the humans required to deliver it.

This is what AI promises to unlock by making intelligence abundant.

The substrate on which AI operates is data. AI systems that have analysed vast volumes of data can infer patterns and predict future outcomes.¹⁶ AI trained on vast datasets of software code can generate code that is often good enough to go straight into production, just as AI trained on images of skin is able to identify a malignant lesion as accurately as a dermatologist.¹⁷ But to do that, the data substrate needs to be organised so that it is ready for AI. This restructuring needs to take place in three specific ways:

• First, the relevant information needs to be made legible: its data, standards, and accumulated know-how captured in a form that is capable of being ingested and processed by AI models;
• Second, the data systems that house the data substrate must be actionable so that it can be accessed and operated upon by AI agents;
• Third, the intelligence system must be verifiable in such a manner that the outputs that it generates can be checked against the relevant ground truth.

When all three features are present, the organisation, sector, or country in question will have crossed the AI absorption frontier.

Of the three requirements listed above, verifiability is the hardest to achieve. Whether an output can be checked, how quickly, and against what depends on the nature of the data and the domain in which it is deployed. Software is inherently verifiable because code either compiles or does not. This is why AI has been able to quickly learn what code works and what does not. And why AI has had such a significant impact on the software sector.

Since other domains do not function in the same way, verification has to be designed differently for them. In this regard, there are broadly three additional modalities to consider:

• In an empirical domain, verifiability is observable but with a lag. Questions of whether a borrower defaults on a loan, a seed variety yields, or a candidate selected for employment turns out to be a good fit are all predictions that are verifiable but only significantly after the fact.¹⁸
• In a consensus domain, truth is what qualified practitioners in the domain agree that it is. Legal opinion and policy design are neither right nor wrong. However, if accepted universally, or at least by a significant consensus of experts in the field, they are considered true. Verifiability in these domains is conformance with the consensus rather than an unequivocal truth.¹⁹
• In an adversarial domain, success is determined with reference to the response of a counterparty. A negotiating position that concedes too much cannot be identified as wrong on its own terms. Its success is determined in relation to the other party’s strategy.²⁰

These three modalities determine both where human judgment will remain necessary and how far a country can push absorption without frontier-scale models. For AI to be absorbed into each of these domains, an appropriate architecture of verifiability needs to be constructed based on their different characteristics.²¹

What Crossing Changes

While no country has crossed the absorption frontier, at least one sector has. If we can analyse how that happened, we can extrapolate from that experience what needs to be done to help others make the crossing.

For decades, the software industry has used version control systems to track new code and edits to existing code. Since these systems are centralised in cloud-based digital code repositories, the entirety of software knowledge work has been digitally legible for a while. The industry’s use of protocols and other open, interoperable digital interfaces has made the data contained in these repositories actionable. And since no software system ever goes into production without testing, the institutional and organisational frameworks needed to granularly verify the output of code²² have existed for decades.

When AI arrived, it was able to plug directly into this infrastructure of legibility, operability, and verifiability. In a relatively short period of time, this has radically transformed the way software is produced²³ to the point where, in various domains, close to a third of new code is written by AI.²⁴ AI was able to transform the software industry because the industry, unbeknownst to itself, had been preparing for this moment for decades.

What the software sector achieved by accident, others can intentionally implement.²⁵ They need to organise their data so that it is better understood by AI models and ensure that the systems that manage it can be operated by agentic systems.

Most importantly, however, they will need to build reliable verification systems that can autonomously validate the output of artificial intelligence systems against ground truth. Done right and at scale, any sector can reach the absorption frontier and unlock the abundance of intelligence within it. To better appreciate what it will take to reach the absorption frontier, the annexes to this report describe the readiness of three sectors of varying data maturity—software, finance, and law.

What applies to organisations and industries can apply to countries as well. Governance data can serve as the substrate for national-level intelligence, enabling them to leverage AI in their own contexts to advance their developmental objectives. In this, some countries are already better placed than others. Those that have established digital public infrastructure (DPI) have, in effect, already laid down the first two layers:²⁶ they have structured their data to be machine-legible, and have put in place standardised interfaces through which it can be accessed.²⁷

The Role of Humans

It is also important to reflect on the role of humans in this future. As much as one of the requirements of safe and trusted AI has always been keeping a human in the loop, if every AI output has to be reviewed by a person before it can be relied upon, all we will have done is shifted the cost of intelligence from producing the answer to confirming it.²⁸ The abundance that AI promises will remain theoretical until its outputs can be acted on without being checked. When verification can be automated to the point where we no longer require a person to inspect each instance, what will be the role of humans?

The likely effect of AI absorption will be to shift human contribution to the questions that verification cannot easily settle. This, with specific reference to the modalities described above, would be in the direction of the consensus and adversarial domains—from work that can be checked to work that must be judged.²⁹ When lawyers no longer need to check drafts, they will have more bandwidth to decide on the optimum negotiating strategy. When doctors no longer need to read scans, they will be able to better evaluate what the treatment strategy should be. When engineers no longer need to verify every calculation, they will have more time to exercise their creative judgment on how they build.³⁰ None of this suggests that jobs will not disappear. They will, particularly those roles that have been built entirely around verifiable output. But what will also happen is new jobs will get created as human judgment shifts up the value chain offering new opportunities because of what this makes possible.

Achieving AI Supremacy

What impact does all of this have on the race to AI supremacy?

We have been made to believe that those who build the most advanced AI models will reign over those who do not.³¹ Be that as it may, the value of a new technology is only ever realised when it is deployed into the world. This has been true of every technological transformation so far³² and will be true of AI as well. As a result, it is not until AI is absorbed into our lives that its impact will be felt.

This, based on our experience with past technologies, is far from trivial. However, once an organisation, sector, industry, or country has crossed the absorption frontier, the cost of applying intelligence to every task diminishes to the point where it can be easily distributed to every instance. When that happens, the benefits that get unlocked downstream of AI will more than justify the cost of producing and consuming it.

A country that has crossed the absorption frontier will be able to convert its ability to distribute AI into leverage. It will be able to engage with AI laboratories from a stronger position than countries that merely host consumers for the intelligence they provide. Because such a country has organised itself to most efficiently absorb the intelligence they provide, AI laboratories will have every incentive to make investments in data centres and inference compute in the country to serve its market most effectively. Coupled with carefully designed policy initiatives, competing laboratories can be encouraged to vie with each other for a dominant position in the market. This will not only benefit customers, it will also reduce the risk of over-dependence on any one provider.

A second source of leverage will be data—more specifically, the verified inference data that is only possible at the absorption frontier. Every AI system generates a record of what was asked and the answer that was returned, which the model uses to improve its performance. Where verification architectures exist, it is easy to determine whether the inference output was correct.³³ AI laboratories will benefit significantly from this as models that learn from verified datasets will perform better than models that do not.³⁴

It is critical to secure this data advantage or else risk what Satya Nadella, the CEO of Microsoft, calls the reverse information paradox.³⁵ Making knowledge legible to AI risks transmitting the embedded process knowledge to those who control the model. To guard against this, the protocols that render data legible to AI should also be able to determine which model reads it and under what conditions. It is only if we can control the protocols in this manner, that we will have any influence over the intelligence generated through them. Those who build this protocol layer will be able to resist being held captive to any single provider of intelligence and, as advanced users of technology, will no longer approach the frontier laboratories as supplicants but as states that must be courted.

None of this is sufficient on its own. Absorption readiness is of little use without the capacity to run inference at the scale necessary to diffuse intelligence throughout the economy. As Anton Leicht, a well-regarded commentator on AI in relation to the middle powers, has argued, middle powers that are unable to build at the frontier will be tempted into protectionism as a substitute, and will find themselves in a stable but steadily diminishing position on the “permanent periphery.”³⁶

Absorption does not offer immunity against this future but the resilience to withstand its consequences. A mature absorption stack lowers the capability threshold of the model required to run it, because the environment supplies the structure and the verification that the model would otherwise have to supply for itself.³⁷ A country that has crossed the absorption frontier will be able to extract enough value from smaller, older, or open-weight models to substitute for not having access to frontier models.

Conclusion

The value of any technology is in its application. It is those who are able to use it most effectively who will be able to generate the greatest value from it. AI supremacy will come not from the creation of technology but from the extent to which it can be consumed. That is best determined in relation to the absorption frontier.

Inasmuch as the steps described in this report will help unlock intelligence abundance, the real gains will come when we are able to tear up the existing workflows, organisational structures, and institutional arrangements that were built around the old constraints of human intelligence. It is in rearranging the entirety of the knowledge work around those constructs that the value of distributed intelligence will be realised.³⁸ It is difficult to assess all that will go into this radical overhaul or predict what lies downstream of it.

What is clear is that, notwithstanding all the hype around what AI can deliver, the real benefits lie in what is yet to come.

In order to practice what this article preached, the author made every attempt to use AI in as many aspects of the writing of this piece as possible. That included brainstorming ideas, sharpening arguments, and structuring the flow of explanations. AI was also used to surface arguments for and against the propositions as well as to come up with outlines, generate drafts, and in the copyediting and proofreading of the final text. For all this, the author relied exclusively on Claude Code via the Command-Line Interface. Despite how extensively it was used, AI was never anything more than a sophisticated tool in the hands of the author — that he used to better convey original ideas constructed from the experience of a lifetime at the confluence of law, technology and society.

Annexes

Software

Software was the first industry to cross the absorption frontier.

For decades, code was written using version control systems that recorded every line and all changes made to it, along with comments that explain the reasoning behind each line and why it was changed. As a result, the sector has a complete, machine-readable history of its work product. Software systems are designed to run using machine-readable interfaces, package registries, and cloud platforms. Long before anyone imagined they would be operated by an AI agent, every system was capable of being reached and operated by every other.

The compiler was the first test of verifiability—code either compiles or it does not—and the industry spent 30 years extending it through automated test suites, continuous integration pipelines, and the discipline that every bug, once found, becomes a permanent regression test. Ground truth in software is nearly free, and feedback arrives instantly.

AI simply plugged into this pre-existing framework. What remained to be done was largely cosmetic: databases were made to expose their contents through model context protocols (MCP);³⁹ browsers, chat interfaces, and commerce platforms redesigned themselves so that agents could operate them; and harnesses⁴⁰ were trained and tested against real-world interfaces until their output could be relied upon unsupervised.⁴¹

The results are measurable. A study in Science found that the share of new software in the United States relying on AI climbed from around five percent in 2022 to nearly 30 percent by early 2025. Beyond the productivity of professionals, anyone with an idea was suddenly able to build working software⁴² regardless of prior training—a democratisation that has produced a proliferation of niche applications that would otherwise never have been built.

Software demonstrates what can happen when the absorption frontier is crossed. It shows that when the harness is model-agnostic, the deployment survives the replacement of the model. Since capability lives in repositories, pipelines, and verifiability, a mature organisation retains its value regardless of the underlying model.

This is what the far side of the absorption frontier looks like.

Finance

What finance has always lacked—at least in countries like India—was legibility. This is what digital public infrastructure enabled, deliberately and at population scale. The Jan Dhan-Aadhaar-Mobile (JAM) trinity and Unified Payments Interface (UPI) enabled a cash economy to be transformed into structured, machine-readable transaction flows.⁴³ The Account Aggregator framework extended this by making financial data portable with consent while the Unified Lending Interface extended legibility to non-financial information that has intrinsic value.

India built legibility in finance as a public good. And since the framework used for this construction was open, interoperable protocols,⁴⁴ the architecture of finance in India is inherently operable by agentic systems should they be allowed to do so.

The third requirement is real-time verification against ground truth that can be carried out without the need to rely on human validation. Double-entry book-keeping—now five centuries old—is the oldest verification mechanism in commerce. It makes it possible for records to check themselves, because books that are created using this framework either balance or they do not. Everything since the invention of double-entry book-keeping has merely been an extension of this foundational idea: reconciliation is mechanical verification, audit is verification institutionalised, and marking to market is continuous verification against the world.

If we can build a digital harness on the architecture of double-entry accounting, we can enable real-time verification of financial transactions. This will allow AI systems with access to legible financial data to operate the systems that manage them and ensure that the financial transactions they carry out using digital financial systems are always aligned with the ground truth through automated verification systems.

Once this happens, the financial sector will have achieved the absorption frontier.

Law

The Indian legal sector is probably further from the absorption frontier than any other. The entire corpus of law is largely illegible. Statutes, subordinate legislation, and the sprawling penumbra of edicts (such as circulars, office memoranda, and guidelines) that constitute the full gamut of compliance obligations are scattered across dozens of government sites, with many not even formally notified.⁴⁵ In addition, laws are often further specified in decisions of the court that fill in the gaps that the strict word of law cannot encompass.

This illegibility is the State’s to cure. It needs to establish a single source of truth for all laws—one canonical, digitally accessible repository of every statute, rule, and instrument, coupled with a guarantee that no one is obliged to comply with any legal obligation that is not contained in that repository.⁴⁶ Once the canonical repository of laws has been established, the same can be done for judgments and all other forms of judicial pronouncements that layer context onto the text of the bare statute.

But the law is not contained solely in the statutes and the judgments that describe it. What makes a good lawyer, general counsel or judge, is the experience and expertise that comes from years of practice, and learned wisdom of what will work in a given context and what will not. Almost all this information resides in that individual’s head, and there is no incentive, given the way the profession works, for any of this to be written down. However, the only way that the legal industry will even begin to approach the absorption frontier is if this is made legible. There is no shortcut to this. Just as software engineers got into the habit of annotating code they produce with comments that provide an explanation as to why it is there, lawyers will need to consciously document their tacit expertise in order to make their intuition explicit.

Once the entire corpus of legal knowledge is legible, it needs to be made machine-readable and operable. In an increasingly digital world, tokenised assets and smart contracts offer new ways to achieve this in respect of commercial transactions. In addition, as legal compliance processes shift from the portal-based forms to real-time digital interfaces such as Application Programming Interfaces (API) and MCP servers, these new frameworks will be the portals through which AI agents can operate.

While the hardest feature to implement will be verification, it is worth noting that the profession already verifies its work—it just does so at human speed. The senior lawyer’s markup of a junior’s work, the precedent system that ensures legislative consistency, and the process of appellate review are all different verification mechanisms that operate on the different aspects of the legal corpora. The challenge is to convert this existing human-speed, sampling-based system of review into one that is continuous, automated, and complete.

What would this look like in practice?

It may be possible to generate a firm’s playbook and build systems to check each new draft for deviations against the firm’s standard positions. For this to work, each partner’s markup must be captured as a training signal so that every correction makes the next review that much stronger. This will not apply to legal work that essentially can only be validated in a courtroom or based on the outcome of a negotiation—they will still need to be verified by humans—but everything else, the bulk of what lawyers actually do, can be reviewed by machines, against standards that the profession has articulated.

The original article can be accessed from this link.