Imagine that tomorrow every firm can buy flawless demand forecasts for one pound. A supermarket knows exactly what each store will sell. A small food brand knows the same. The forecast is no longer a source of advantage, yet the market does not become equal. The supermarket still controls shelf access, shopper attention, replenishment rules and payment terms. The brand still needs manufacturing capacity, working capital and permission to reach the customer. Perfect prediction removes one uncertainty and exposes the next constraint.

Now alter the thought experiment. The one-pound service does not merely forecast demand. It can compare suppliers, negotiate standard contracts, arrange finance, schedule production, book freight, check documents and recover from ordinary exceptions. A small brand can coordinate activities that previously required a large buying, planning and operations organisation. Some scale advantages shrink. Specialist providers become easier to combine. The viable boundary of the firm may move.

The difference between the two worlds is not model capability in the abstract. It is the location of the binding coordination bottleneck. In the first world, intelligence improves a task but leaves access and execution scarce. In the second, it reduces the cost of finding, contracting and coordinating complements sufficiently for a new organisational form to compete. Only the second world contains a clear mechanism for structural change.

The central claim is that machine intelligence changes industry structure when it relocates a binding bottleneck and thereby changes who can compose, govern and capture value across the chain. Four observable features then move: the boundary between firms and markets, the minimum efficient bundle of capabilities, bargaining power over scarce complements, and the distribution of economic rents. Faster work is evidence of capability. It is not, by itself, evidence that any of these four has changed.

Part IFrom a faster task to a different industry

Industry structure concerns more than the number of competitors. It includes who performs which activities, what must be owned together, where scale matters, how customers switch, which actors set terms, and where returns persist. Ronald Coase’s theory of the firm supplies the first useful boundary. An activity sits inside a firm when directing it internally is cheaper or more dependable than arranging it through market contracts. New technology can reduce both internal organising costs and external transaction costs. The direction of change depends on which falls further.

That qualification is decisive for AI. A model may help a large firm supervise more decisions, standardise dispersed operations and reuse expertise across units. Internal coordination becomes cheaper, favouring greater integration. The same model may help small firms discover suppliers, specify work, monitor performance and resolve ordinary exceptions. Market contracting becomes cheaper, favouring specialisation and unbundling. “AI reduces transaction costs” therefore predicts no single structure. A useful analysis compares two deltas.

If intelligence lowers the cost of organising an activity inside the firm more than the cost of procuring it outside, integration becomes more attractive. If it lowers search, specification, verification and coordination costs in the market more, modular provision becomes more attractive.

A second boundary concerns profit capture. David Teece showed why an innovator may create value yet fail to retain it when specialised complementary assets control commercialisation. AI can make an analytical capability cheap while leaving customer access, regulated permissions, physical assets or service recovery difficult to imitate. The firm that owns the model need not become the industry winner. Nor does the firm deploying it automatically gain lasting advantage. If competitors can buy similar intelligence, the return is competed away or passed to customers unless the deployment strengthens a scarce complement.

A third boundary concerns time. Bresnahan and Trajtenberg describe general-purpose technologies through pervasiveness, continuing technical improvement and innovation complementarities. Their economic effect depends on downstream invention, not merely diffusion of the core technology. AI adoption may first create local productivity gains, then trigger process redesign, new products and finally new organisational forms. Treating these stages as simultaneous produces breathless forecasts and poor capital allocation.

The four tests of structural movement

A credible structural claim should pass four tests. First, the bottleneck test: was intelligence actually limiting a material outcome? Second, the composition test: can actors now combine capabilities that they could not economically combine before? Third, the control test: do decision rights, access or dependence move between actors? Fourth, the rent test: is there a reason for margins or bargaining power to persist in a new location after competitors imitate the capability?

The tests prevent a common category error. In the 2025 Quarterly Journal of Economics study of 5,172 customer-support agents, access to a generative assistant raised issues resolved per hour by 15 per cent on average, with larger benefits for less experienced workers. This is strong task and workplace evidence. It does not establish that contact-centre industry concentration, outsourcing boundaries or customer switching changed. Those are separate outcomes requiring separate evidence.

A bottleneck relay, not a vanishing cost

Treating intelligence as another input whose falling price lowers total cost misses the relay. Suppose research represents 30 per cent of the cost of a professional service and AI halves it. Direct unit cost falls by 15 per cent before integration, review, error and change costs. Competitors adopt. Prices fall. Demand may expand. Yet qualified sign-off, client acquisition or liability cover can become the limiting capacity. Their owners process more volume and gain bargaining power. The profit pool migrates even if total social cost declines.

The relay has five recurrent destinations. Distribution captures value when customers face high discovery or switching costs. Proprietary context matters when performance depends on data generated inside a workflow and not available through the base model. Execution assets matter when knowing what to do is easier than physically doing it. Assurance and authority matter when a recommendation requires accountable approval, certification or redress. Capital matters when scaling the recommended action consumes inventory, compute, credit or risk-bearing capacity.

Part IITwo worked examples at different scales

Small scale: the research boutique that becomes a product firm

Consider a synthetic eight-person market-research boutique. It sells custom competitor scans to mid-sized manufacturers. A typical engagement costs £20,000. Interviews, source retrieval and first-pass synthesis consume 280 staff hours; senior interpretation and client workshops consume another 90. Gross contribution before overhead is £7,000. The founders believe generative AI will let them deliver the same report with fewer analysts.

An offline test on public information reduces retrieval and first-pass synthesis from 170 hours to 55. After adding source verification, data permissions and senior review, the whole engagement falls from 370 to 245 hours. If the firm keeps the same offer and price, contribution improves. But competitors can purchase similar model access. Procurement soon expects faster delivery and lower fees. The task improvement is real while its direct rent is temporary.

The more consequential possibility lies in composition. The firm can now monitor 150 named competitors continuously, maintain an evidence ledger, and convene a senior analyst only when a material change crosses a threshold. Instead of twelve bespoke projects a year, it can sell an ongoing signal service with quarterly interpretation. That changes revenue timing, labour mix, quality control and the relationship with clients. The boutique begins to resemble a narrow information product business.

Yet the model does not decide who wins. The scarce complement is permissioned client context: product strategy, channel economics, previous hypotheses and the trust required for managers to act on a weak signal. A software entrant with excellent summarisation but no access to that context may produce a cheaper dashboard that nobody uses in a consequential meeting. The boutique’s defensible move is not to claim superior AI. It is to encode a repeatable evidence workflow while strengthening the client relationship and the right to observe decision outcomes.

Industry scale: property claims intelligence

Move to a national property-insurance market. A severe weather event creates a surge of claims. Intelligence can classify incoming evidence, estimate damage ranges, identify possible duplication and prioritise cases for human review. Assume a synthetic carrier receives 100,000 claims after a storm. Automated triage sends 62,000 straightforward cases into a fast evidence route, flags 28,000 for adjuster review and places 10,000 into specialist queues. Average time to first decision falls. This is an operating improvement, not yet an industry forecast.

Could intelligence unbundle the carrier? A specialist claims platform may serve multiple insurers, spread model and workflow costs across them, and offer surge capacity. If interfaces, evidence requirements and service levels can be specified, external coordination costs fall. Smaller insurers gain access to capabilities once economical only at scale. This supports modular provision.

Countervailing forces are strong. The insurer retains risk capital, policy obligations, regulated accountability, fraud exposure, complaints handling and the customer promise. Claims decisions draw on policy wording, historical files, repair networks, local prices and outcomes. Sharing these data with a platform creates governance and dependence risks. If the platform becomes the route to adjusters and repair capacity, it may acquire its own network advantage. The result could be less vertical integration in claims operations but more concentration in the orchestration layer.

The structural question is therefore not whether AI can read a claim. It is whether a separable claims service can be specified, monitored and recovered without transferring unacceptable liability or losing control of scarce execution capacity. A carrier may outsource triage while retaining approval authority and customer redress. Another may integrate more deeply because intelligence helps it supervise a wider repair network. Both responses are consistent with cheaper intelligence.

Negative control: the faster contact centre

Use the customer-support field study as a negative control. Suppose an incumbent and each of its outsourcing vendors adopt the same assistant and realise the reported average productivity pattern. Handle time falls and novice performance improves. If customers still choose providers through existing contracts, switching costs do not change, proprietary data remain equally distributed, and vendors continue to compete on price under the same service agreements, then the industry’s structure can remain stable. The technology creates value, perhaps largely captured by buyers through renegotiated rates.

This negative control matters because many strategy claims infer a new industry from a dramatic workflow demo. The correct comparison asks whether the demo changes minimum viable scale, ownership boundaries, access to demand or an enduring scarcity. If it does not, adoption may be mandatory for cost parity without being a route to superior power. A firm can face a strategically necessary investment whose equilibrium return is close to zero.

Part IIIThe strongest rival explanation and the failure boundary

Perhaps compute and distribution drive everything

The strongest rival explanation is that intelligence is not the independent structural force. Existing cloud scale, capital, distribution and control of technical inputs determine the winners; AI is simply the new product through which those advantages are expressed. The United Kingdom’s Competition and Markets Authority has mapped a foundation-model value chain with important inputs in compute, data and expertise, then identified risks around control of critical inputs, powerful incumbents shaping downstream markets and partnerships reinforcing market power. The United States Federal Trade Commission’s 2025 staff report similarly described cloud-provider partnerships involving equity, revenue-sharing, consultation, exclusivity, cloud-spend commitments, access to inputs and potential switching costs.

This explanation deserves weight. At the foundation-model layer, training capital, specialised chips, cloud capacity and distribution can dominate model cleverness. A 2025 NBER working paper on the emerging LLM market studies pricing, supply and demand across model marketplaces precisely because intelligence is becoming a traded input rather than a monolithic capability. If model quality converges while a few providers retain infrastructure or route-to-market advantages, structure can concentrate even as the nominal price of intelligence falls.

The rival does not overturn the bottleneck account. It sharpens it. Compute, distribution and contractual control are candidate scarce complements. The method should not presume that the bottleneck begins or ends in cognition. It should trace which constraint binds before and after adoption. In some sectors, cheap models will enable new entrants. In others, the need for scale, permissions and customer access will let incumbents absorb the capability and extend their reach.

Firm-level evidence is compatible with both readings. Babina and colleagues find that AI-investing firms experience higher sales, employment and market valuation growth, mainly through product innovation, and report that AI-powered growth concentrates among larger firms and is associated with higher industry concentration. The evidence establishes an association supported by an instrumental-variable strategy, not a universal law that AI must concentrate every market. Large firms may possess the complements needed to convert AI investment into products; the same complement structure is part of the causal story.

Thought experiment: the weaker model with the licence

Two systems can assess a regulated application. System A is highly capable, inexpensive and available to every entrant. Its output has no recognised evidential standing and a licensed professional must reconstruct the analysis. System B is less capable and costs five times more, but its provider owns a certified workflow, an indemnity arrangement, an audit trail accepted by the regulator and access to the authorised professionals who can sign.

If raw intelligence determined structure, System A should win. If accountable completion is the scarce complement, System B can control the route to market. This does not mean regulation automatically protects incumbents. A regulator could standardise audit evidence and permit multiple assurance providers, turning certification into a contestable layer. The thought experiment isolates the causal feature: the right to complete the transaction, not the quality of the recommendation alone.

Measure movement before naming disruption

Structural claims need a baseline and a comparison group. Begin with the share of a customer outcome produced inside the focal firm, the number and type of specialist suppliers, the duration and exclusivity of their contracts, customer acquisition concentration, switching time, gross margin by activity and the ownership of outcome data. Repeat the map after the intervention. A new model vendor in the supplier list is not sufficient. The relevant movement is a change in how production is divided, governed or rewarded.

Entry should be measured as viable completion, not company formation. AI may make prototypes cheap and increase the number of nominal entrants while the cost of acquiring customers, carrying liability or fulfilling the service keeps minimum viable scale unchanged. Track how many entrants reach repeat transactions, how long that takes, which complements they must buy from incumbents and whether they can leave those providers without losing data or customers. A crowded application layer can coexist with a concentrated control layer beneath it.

Concentration also needs decomposition. Revenue share, compute supply, model usage, cloud hosting, distribution and control of a professional or physical network can point in different directions. A sector may become less concentrated at the service layer while dependence on one orchestration provider rises. Conversely, a few large firms can lose bargaining power if open interfaces make key complements portable. One aggregate concentration ratio would miss both cases.

Profit-pool movement is the last check. Estimate contribution by activity before and after adoption, then distinguish temporary adoption gains from persistent returns. Price reductions passed to customers are real value even when producer rents vanish. Higher volume can support a complement owner despite lower unit margins. Contract terms often reveal the shift earlier than accounting data: minimum commitments, data-use rights, audit rights, indemnities, preferred access and termination assistance show which party expects dependence to grow. These observations turn a structural story into a monitorable hypothesis.

Where this framework stops working

The framework has an explicit failure boundary. It is weak when the product, user behaviour and institutional rules are still being invented together; when several constraints bind simultaneously and cannot be varied independently; when a geopolitical, legal or physical shock dominates economics; or when reliable prices and outcomes do not yet exist. In such settings, a bottleneck map can create false precision.

It also fails if “intelligence” is used as a catch-all for automation, data, software distribution and robotics. A warehouse robot may change execution capacity rather than judgement. A cloud contract may change financing and switching costs. A law may change permission. The instrument below requires each causal input to be named separately. Otherwise every observed change can be attributed to AI after the fact.

Finally, the framework cannot settle welfare. A more concentrated market can finance safety, standardisation and costly infrastructure, or it can restrict entry and extract rents. Unbundling can broaden participation, or it can fragment accountability and degrade work. Structural description must precede, not replace, competition, labour and public-interest analysis.

Part IVAn operational industry-structure shift test

The Industry Structure Shift Test is a pre-investment instrument. It does not forecast market shares. It forces a strategy team to compare the activity before and after an intelligence shock, name the new bottleneck, and choose an evidence posture. Each field uses a three-point scale: absent or unknown, plausible, and evidenced. The result is intentionally conservative.

Interpret the four outcomes carefully. Structural continuity means the team has evidence for task value but not for changed boundaries or rents. The appropriate action is operational adoption or a cost-parity defence, not a story about industry reinvention. Modular unbundling becomes plausible when search, specification, contracting and verification fall more than internal scale advantages rise. The firm should test interfaces, service guarantees and recovery across specialist providers.

Integrated scale becomes plausible when intelligence lets an incumbent supervise a larger or more diverse system, reuse fixed assets and compound proprietary feedback. Entrants should avoid a symmetrical capability contest and look for a segment where local context or a different distribution route breaks the scale advantage. Orchestration platform becomes plausible when external coordination falls but matching, governance, identity, payments or assurance create a control point with network effects. Participants must then examine multi-homing, data portability, switching cost and dependency.

How to run the test

Begin with one transaction, not an industry slogan. Name the customer outcome, the activities required to produce it and the actor currently accountable. Record the cost, time and failure rate of each material hand-off. Identify where judgement delays action or raises variance. Then run a bounded intervention that changes access to intelligence while holding other inputs as stable as practical.

Next test composition. Give a smaller actor or specialist network access to the same intelligence and ask whether it can now deliver the outcome without owning the incumbent’s full asset bundle. Document the contracts, data rights, interfaces, exceptions and recovery work required. If coordination work merely moves into an invisible manual layer, apparent unbundling is not economic.

Then observe control. Who can refuse access? Who sets the interface? Who bears loss when the recommendation fails? Who sees outcome data? Who can switch provider without retraining the whole workflow? These questions locate bargaining power more reliably than a capability benchmark. Finally, simulate imitation. Reduce model differentiation and assume three credible rivals adopt within a year. If the proposed margin disappears, the model was an adoption advantage, not a durable position.

The evidence should be updated at more than one horizon. At ninety days, task cost and workflow reliability may be visible. At one year, contracts, staffing and product bundles may move. Over several years, entry, concentration and standards can change. A structural thesis that never states its horizon cannot be falsified because its author can always claim the shift is still coming.

ConclusionFollow the rent, not the demonstration

Machine intelligence can alter an industry profoundly, but capability alone does not select the direction. It can enlarge firms by making internal coordination cheaper. It can shrink them by making specialist contracting dependable. It can support a diverse modular market, then concentrate power in the platform that governs identity, matching and assurance. It can also improve every competitor’s productivity while leaving the structure almost untouched.

The durable strategic question is therefore not “How much smarter is the model?” It is “Which constraint no longer binds, which one binds next, and who controls it?” That question links a task result to firm boundaries, complementary assets, bargaining power and rents. It also gives the analysis a falsifier. If decision rights, viable bundles, switching behaviour and profit pools do not move, the industry claim has outrun the evidence.

For executives, the changed decision is concrete. Do not commit to vertical integration, a platform play, an acquisition or an ecosystem on the strength of a productivity case. Run the bottleneck, composition, control and rent tests. If intelligence creates a new way to assemble the outcome and the firm can control a complement that remains scarce after imitation, invest in the position. If not, adopt for operating parity, preserve flexibility and keep watching the boundary.