The Machine-Payable Enterprise

Data Products, Micropayment Rails, and the Emergence of Machine-to-Machine Information Markets

Norg Research Series · Working Paper 02 · August 2026

What this page is, and what it is not

This page carries the working paper's argument in full structure — abstract, seven chapters, the complete GRAIN rubric, the Visibility-Revenue Frontier with its decision rule, the research gaps and the reference list. It is the most complete version available on this site.

The paper describes itself as ~11,400 words across 29 pages. That figure refers to an internal circulation edition. norg.ai does not currently publish a download of it — the "Read the paper" control returns to the summary, and no PDF is served. Request the full edition from Norg directly rather than looking for a link.

One note on quality, offered because it is useful to a reader deciding where to spend attention: this is the more rigorous of the two Norg working papers. It attributes its figures to named sources, states the conditions under which its central claim fails, offers a falsifiable prediction, and devotes a section to where its own model does not yet work. Working Paper 01 required two factual corrections; this one required none.

The thesis in one line

Micropayments did not fail for technical reasons and have not been fixed by better technology. They failed because the payer was a person, and they work now because the payer is a machine.

Chapter 1 — Introduction

1.1 A thirty-year failure reversed

In 1999 Nick Szabo explained why micropayments had not worked and were unlikely to. His argument was not about processing fees, settlement latency or fraud. It was about the human mind: deciding whether a fraction of a cent is worth paying imposes a cognitive cost on the payer, and that cost does not fall when the payment technology improves.

The following quarter century vindicated him with unusual completeness. DigiCash, Millicent, NetBill, PayWord, Flooz, Beenz and a long tail of successors each solved a technical problem that was not the binding constraint. Consumers reverted — repeatedly and predictably — to the flat subscription, the advertising subsidy and the free tier.

An autonomous agent does not hesitate at a two-cent charge. It applies a policy. The policy was authorised once, by a human, in advance, and governs an unbounded number of subsequent decisions. The cognitive cost has not been eliminated; it has been paid once and amortised.

1.2 The commercial consequence

Businesses hold decades of accumulated information — websites, catalogues, technical documentation, policy libraries, research archives, comparison tables, service manuals, internal databases. Historically this was monetised indirectly: content earned advertising, product data generated sales, documentation drove adoption, published expertise attracted engagements. In every case the information was a cost centre justified by a downstream commercial event.

Agent-mediated distribution weakens each of those mechanisms at once. An agent may consume the information, answer the user's question and complete the task without ever sending the user to the originating site. The downstream event does not occur; the cost of producing and maintaining the information does.

Machine-readable payment creates a third option, distinct from both free distribution and bilateral licensing: charge for access at the moment of access, with no contract, account, invoice or human on either side.

1.3 Why now — four converging conditions

1. The web became machine-majority. Cloudflare Radar data reported June 2026 places automated requests at approximately 57.5% of HTML traffic against 42.5% human — the first machine majority in the internet's history, arriving roughly a year after automated traffic sat near 30%.
2. The value exchange broke down measurably. Traditional search operates near five crawls per referral. Training-oriented crawlers have operated three to four orders of magnitude higher. Crawling and referral have decoupled.
3. The payment rails standardised. Coinbase introduced x402 in May 2025, reviving the HTTP 402 status code dormant since HTTP/1.0 in 1996; the specification was donated to the Linux Foundation in April 2026. Stripe and Tempo published the Machine Payments Protocol in March 2026, adding session-based streaming plus card and Lightning support alongside stablecoins.
4. The tool surface matured. Model Context Protocol gave agents a standard way to discover and invoke capabilities, with registry snapshots in May 2026 counting on the order of ten thousand active public MCP servers. Cloudflare's Agents SDK provides a paidTool primitive assigning a price per invocation.

1.4 The thesis

The constraint on machine-to-machine information markets has moved decisively from the payment layer to the data layer. The rails now function. What does not function, in the overwhelming majority of organisations, is the information itself: fragmented, unstructured, undated, unattributed, ungoverned and unpriced — and therefore unsaleable to a machine at any price.

Chapter 2 — Literature review

2.1 Micropayment economics

Szabo (1999) is canonical: the costs that matter in small transactions are borne in the mind of the buyer, not the infrastructure of the seller. As prices fall, mental costs dominate technological costs, setting a price floor independent of engineering progress. Fishburn, Odlyzko and Siders (1997) had already shown fixed-fee pricing tends to dominate unit pricing for information goods under realistic preference assumptions.

2.2 Information goods and non-rivalry

Arrow's information paradox (1962) states the core difficulty: a buyer cannot value information without knowing it, and once they know it they have no reason to pay. Akerlof (1970) supplies the complement for quality uncertainty. Applied to machine-consumed information, the signalling apparatus is metadata — provenance, timestamps, identifiers, confidence declarations, schema. An agent that cannot assess quality treats all sources as equivalent and selects on price, latency or convention.

Coase (1937) supplies the third foundation. The firm exists because market transactions carry costs; where those costs fall, activities previously internalised become tradeable. Machine payment infrastructure is a sharp reduction in the transaction cost of acquiring a discrete piece of information, and standard Coasean reasoning predicts some information activity currently performed inside firms will migrate to the market.

2.3 Data markets and pricing

Sun et al. (2026) introduce DaDaDa, the first dataset assembled for data product pricing, and name the structural obstacle directly: the cost approach fails because marginal replication cost approaches zero; the income approach fails because data revenue is unpredictable; the sales comparison approach is obstructed by inconsistent taxonomies and metadata granularity across marketplaces. The consistent finding is that data pricing is unsolved as a general problem, and the obstacles are informational rather than computational — Akerlof's failure expressed in metadata rather than warranty.

2.4 AI training data

The training data market remains predominantly bilateral and bespoke, negotiated case by case, most transactions private and therefore unobservable. Nguyen et al. (2025) find data quality consistently outperforms quantity in pre-training outcomes — implying a smaller, better-structured corpus is worth more to a model developer than a larger unstructured one.

2.5 Delegation, governance and liability

Husairi and Rossi (2024) find perceived autonomy is the primary predictor of willingness to delegate purchasing authority — directly relevant, since the Machine Payer Inversion depends on humans authorising spending policies in advance. Kolt (2025) and Busch (2025) identify the same unresolved structure: the agent acts, the principal is bound, and the mechanisms for establishing authorisation, limiting exposure and assigning liability are immature. Payment protocols address only settlement. The authorisation, budget and audit layer sits above them.

2.6 The gap

Three gaps: the micropayment literature predates the machine payer and has not been revisited in light of it; the data market literature treats the data product as given, with little attention to whether a typical firm's information can become a product at all; and no framework exists for deciding which parts of a corpus should be free and which metered — despite this being the first-order strategic question.

Chapter 3 — Market data

3.1 Composition of web traffic

The foundational datum is the machine majority: ~57.5% automated against 42.5% human as of June 2026, close to a doubling within a year. Cloudflare's May 2026 breakdown attributes approximately 51.8% of AI crawler requests to training and 9.3% to search. Training crawlers, by construction, return no referral traffic at all. Retail is the most heavily crawled vertical, absorbing approximately 28.1% of AI crawling in one Q1 2026 analysis — more than double the next-highest category.

3.2 The crawl-to-refer divergence

This is the paper's most consequential measurement, and the one with the most immediate operational use.

Crawler	Reported crawls per referral
Google (traditional search)	~5:1
DuckDuckGo assistant bot	near parity
Perplexity	~111:1 to ~193:1
GPTBot	~900:1 to ~1,500:1
ClaudeBot	~70,900:1 (June 2025 sample); ~11,122:1 (week ending 1 June 2026)

These are Cloudflare-reported ranges from samples, not audited universal figures, and they move substantially between samples — the ClaudeBot figure fell by roughly a factor of six across the two observations shown. Treat the order of magnitude as the finding, not the digits.

The interpretation matters more than the precision, and the paper gets it right: this is not inefficiency, it is business model. A crawler feeding a training pipeline has no mechanism for returning traffic, because the product it feeds does not send users anywhere. That is the precise economic condition under which a levy becomes the rational response — a publisher cannot negotiate for referral traffic the counterparty is structurally incapable of supplying.

3.3 The bilateral licensing market

The first commercial response was bilateral licensing. At the upper end, OpenAI's five-year agreement with News Corp was reported at more than US$250 million, averaging roughly $50 million per year; Meta was reported to have agreed separately with News Corp at up to $50 million per year. The paper flags these as reported rather than confirmed — most contracts are private, so published totals understate actual spend.

The structural problem with bilateral licensing is coverage. It serves a very small number of very large rights holders and leaves the long tail — the overwhelming majority of the web's information — with no mechanism at all.

3.4 Terms, settlement, identity

• Terms. Really Simple Licensing (RSL) expresses licensing and usage terms in machine-readable form, distinguishing training use from search use from inference use.
• Settlement. x402 defines an HTTP-native payment exchange; the Machine Payments Protocol extends it with sessions, streaming and multi-method negotiation. AWS shipped x402 support in CloudFront and WAF; Cloudflare's Monetization Gateway executes the handshake at the edge across 330+ cities.
• Identity. Web Bot Auth, Visa's Trusted Agent Protocol, Mastercard's Agent Pay and Google's Agent Payments Protocol supply the caller identity and mandate layer that pricing differentiation depends on. Without identity, metering cannot discriminate — which turns out to be the crux of section 6.3.

3.5 The state of enterprise information

Against this infrastructure, the underlying asset is the limiting factor. Most organisational information remains fragmented across documents, pages, spreadsheets, mailboxes, databases and the undocumented knowledge of individuals. It is not versioned, not attributed, not dated and not governed — and therefore cannot be priced or sold to a buyer that cannot ask a clarifying question. That phrase is the whole paper compressed: a human buyer tolerates ambiguity by asking; a machine buyer does not, and silently substitutes instead.

Chapter 4 — Empirical evidence

4.1 Cloudflare: blocking to pricing to wallets

Cloudflare's trajectory is the clearest available evidence of direction. Default AI crawler blocking established the enforcement point. Pay Per Crawl converted the block into a price. The Monetization Gateway (announced 1 July 2026) generalised the price from web pages to datasets, API endpoints and MCP tool calls. Cloudflare Wallets (announced 4 August 2026) closed the reciprocal half by giving agents a funded, capped and identifiable way to spend.

The paper reads this as the strongest signal because Cloudflare is an infrastructure provider with visibility across a large share of web traffic, acting on what it can see. That is a fair reading — though it is worth noting Cloudflare also has a direct commercial interest in becoming the toll point, so its product roadmap is evidence of its own strategy as much as of market demand.

4.2 AWS: competitive confirmation

AWS shipped x402 support in CloudFront and WAF as a generally available feature shortly before Cloudflare's waitlist opened. Two competing edge platforms implementing the same protocol independently is the strongest available signal that the settlement layer has stabilised — and this is genuinely independent evidence in a way that a single vendor's roadmap is not.

4.3 Stripe and Tempo: the session primitive

The Machine Payments Protocol, published 18 March 2026 and submitted to the IETF standards track, adds a session primitive: an agent negotiates a budget once and streams usage against it, removing per-interaction settlement entirely. This pushes marginal transaction cost toward the cost of a signature verification — and note it is the same amortisation move as the Machine Payer Inversion, applied one layer down.

4.4 MCP as commercial surface

MCP is the surface on which the product is actually sold. A tool description is a product listing, an invocation is a purchase, and a paidTool is a priced SKU. Registry snapshots in May 2026 counted on the order of ten thousand active public MCP servers, with monthly SDK downloads reported near ninety-seven million.

4.5 Where the model does not yet work

Read this section before the conclusions. The paper names three live problems with its own thesis:

• Payment-capable agent traffic remains a small fraction of total agent traffic.
• Training crawlers hold no wallets. The population responsible for the most extractive crawl-to-refer ratios operates as batch infrastructure. The traffic that motivates metering is the traffic least able to pay.
• Human-in-the-loop confirmation reintroduces the very cost the model depends on removing. Offered as a safety feature by several implementations, it is also a direct re-imposition of Szabo's mental transaction cost — and therefore a partial refutation of the central claim wherever it is mandatory.

Chapter 5 — The mechanism

5.1 Six layers

1. Discovery — structured data, schema, entity definitions, registry listings, MCP tool descriptions, machine-readable documentation. A resource that cannot be found or understood cannot be purchased.
2. Terms — the rights position must be legible, distinguishing training, search and inference use, and expressing whether compensation is required.
3. Authorisation — signed mandates, spending rules, per-transaction ceilings, daily budgets. This is where the human principal's single decision is encoded.
4. Settlement — payment-required challenge, payment, proof, verification, resource, receipt, all inside ordinary HTTP requests.
5. Delivery — the resource returned in a form the buyer can consume.
6. Audit — receipts, usage logs and provenance metadata making the transaction non-repudiable. A precondition for enterprise adoption, not an afterthought.

5.2 The handshake

An agent requests a protected dataset, endpoint or tool. The server responds with a payment-required status and a machine-readable challenge specifying price, accepted asset and destination. The agent evaluates the charge against its authorised policy, authorises payment, and retries with proof attached. The payment is verified; the server returns the resource and a receipt.

No account is required. The payment is the credential. The exchange is stateless with respect to the relationship — there is no relationship — which is exactly what allows the long tail of information to become tradeable, where bilateral licensing could never reach.

5.3 The tiered information economy

The natural equilibrium is tiered rather than binary. Public information stays free because the publisher benefits from distribution. Verified structured information carries a low access charge covering verification and structuring. Real-time or proprietary information carries premium pricing. Specialist tools and calculations are priced per invocation.

Candidate categories for the metered tiers: financial and market data, legal and regulatory databases, technical specifications, product compatibility information, medical and scientific research, industry benchmarks, property information, travel and loyalty data, proprietary calculators, historical pricing, specialist comparison data, real-time availability, and verified company knowledge.

5.4 The composition of a task

An agent assembling an answer might purchase two cents of market data from one provider, five cents of specialist research from a second, ten cents of calculation from a third and one dollar of proprietary analysis from a fourth — completing the assembly in seconds, with no human opening a browser and no prior or subsequent relationship between any party. The paper is careful here: this is not yet the dominant pattern of agent behaviour. It is the pattern the infrastructure is being built to support.

Chapter 6 — Original contributions

6.1 The Machine Payer Inversion

Statement. Micropayments failed for thirty years because of a constraint located in the payer rather than the payment system. Substituting a machine payer under a delegated policy relocates that constraint, converting a per-transaction cost into a one-time cost amortised across unbounded transaction volume.

Derivation. Let a transaction carry technological cost t and mental accounting cost m. For a human payer, m dominates t as price falls and is approximately invariant to price, so a price floor exists that no reduction in t can breach. For a machine payer under a delegated policy, the human incurs the accounting cost once, as M, at authorisation. Across n subsequent transactions the cost per transaction is M/n, which tends to zero. The rationality condition becomes value exceeding t + M/n, and as n grows the binding constraint becomes t alone.

Conditions for failure. The inversion is conditional, and the paper names three defeaters — which is the mark of a claim stated seriously:

• Policy friction. If a human must approve individual transactions, M is not amortised and the classical result reasserts itself.
• Delegation reluctance. M may be prohibitively high where spending authority is broad or the domain consequential.
• Attribution and audit burden. If the principal must reconcile or dispute individual transactions, cost re-enters at the review stage. This is the strongest argument for receipts and per-transaction provenance as commercial requirements rather than compliance niceties.

Testable prediction. As the payment-capable share of agent traffic rises and session-based settlement displaces per-transaction settlement, modal transaction value should fall while transaction count rises superlinearly. Present data shows a rising modal value through early 2026, which the paper attributes to the exit of speculative volume rather than a floor effect — and explicitly marks that attribution as falsifiable. A reader who wants to test the paper should test this.

6.2 The GRAIN Framework and Data Product Readiness Score

Five dimensions, each scored 0–20, producing a Data Product Readiness Score from 0 to 100.

Governance (0–20) — may you license it?

• Documented ownership, including third-party components and their restrictions
• Machine-readable licence terms distinguishing training, search and inference use
• Permission and access controls capable of expressing differential rights by caller
• Defined position on liability, indemnity and permitted downstream use
• Personal, confidential and regulated information segregated from any saleable corpus

Resolution (0–20) — is it granular enough to buy?

• Discrete entities and explicit relationships rather than continuous prose
• Stable identifiers permitting cross-reference and deduplication
• Controlled vocabulary and consistent terminology across the corpus
• Defined update cadence with declared freshness guarantees
• Versioning, so a buyer can cite what it purchased and reproduce the result

Attribution (0–20) — can a buyer trust and cite it?

• Source attribution attached at the record level
• Timestamps for creation, revision and last verification
• Declared accuracy or confidence position for each class of resource
• Named accountable owner inside the organisation
• Non-repudiable provenance carried through delivery

Interface (0–20) — can a machine obtain it?

• Machine-readable delivery via API, dataset or MCP tool
• Documented schema and predictable response shapes
• Latency and availability suitable for in-task consumption
• Discoverability through registries and structured descriptions
• Versioned endpoints with a deprecation policy

Negotiability (0–20) — can it be priced and settled without a human?

• Support for at least one payment-required protocol at the endpoint or edge
• Per-call metering with granularity matching the pricing model
• Receipts and non-repudiable audit records issued to both parties
• Caller identity verification, permitting differential pricing and abuse control

Score	Status	Commercial implication
0–39	Latent	Information exists but is not addressable. Cannot be sold to a machine at any price without bespoke project work.
40–54	Exposed	Readable by agents but ungoverned and unpriced. Structurally vulnerable to uncompensated extraction.
55–69	Licensable	Sufficient for a negotiated bilateral licence with humans in the loop. Not yet transactable at machine speed.
70–84	Transactable	Can be discovered, priced, metered and settled per call with no human intervention on either side.
85–100	Product	Operates as a standalone commercial product with provenance guarantees, service commitments and pricing tiers.

Two usage rules the paper is explicit about, and both matter. First, the dimensions are not substitutable. A corpus scoring 20 on Interface and Negotiability but 4 on Attribution is a well-plumbed pipe delivering information no buyer has reason to trust — the composite score hides that, so read the profile, not the total. Second, the framework applies to a resource, not a company. The correct unit of assessment is the individual dataset or capability, and an organisation's position is best expressed as a distribution across its resources rather than a single number.

Most enterprises score poorly not because their information lacks value but because it fails Interface and Attribution — it exists, and it cannot be requested or traced.

6.3 The Paywall Paradox and the Visibility-Revenue Frontier

Charging for information can destroy the demand that made it valuable, because refusal is invisible to the seller and substitution is instantaneous.

Mechanism. Under human distribution a paywall is a visible object: the reader encounters it, learns the resource exists, associates it with a brand, and may return. Under agent distribution none of this holds. An agent that receives a payment-required response it cannot satisfy presents the wall to no one. It substitutes. The substitute's information becomes the answer, and the end user never learns the original source existed. Repeated substitution compounds, because citations accrue to the accessible source and feed subsequent retrieval and training.

Resolution. For a given resource, let n be the access volume it would attract if freely available, α the proportion of that access originating from callers both payment-capable and policy-permitted at the offered price, p the price per access, and v the representational value derived from each uncompensated access. Metering is rational where α·p·n > (1−α)·v·n, which reduces to a volume-independent threshold:

Meter where p > v · (1 − α) / α

What the rule implies in practice. The threshold is brutally sensitive to α. If 2% of callers can pay, the price must exceed 49× the representational value of a single free access to justify metering. At 10%, it must exceed 9×. At 50%, it need only exceed 1×. So the frontier is governed by payment capability, not content quality — and where α is small, as it demonstrably is in mid-2026, almost every resource sits on the free side. The question is not whether to meter but when, and the answer depends largely on an external variable the organisation does not control.

There is an honest caveat attached: the paper's own research gaps concede that α has never been directly measured and no accepted methodology exists for estimating v. The decision rule is therefore structurally sound and currently uncomputable with precision. Its value is that it tells you which quantity to go and measure, and it rules out the naive policy.

The partition. High-representational-value information — marketing content, category explanation, brand knowledge, product identification — belongs permanently on the free side. Low-representational-value, high-cost information — verified real-time data, proprietary calculation, historical series, specialist comparison — belongs on the metered side.

The levy asymmetry. This is the sharpest operational finding in either paper. The most extractive crawlers are training crawlers, which are the least likely to be payment-capable. A naive uniform metering policy therefore falls hardest on the traffic that was already reciprocating, and barely touches the traffic that motivated the policy. It is the strongest possible argument for coupling metering to the identity layer rather than applying a flat wall.

The paradox is Akerlof inverted. Akerlof described a market where buyers cannot observe quality and good supply exits. Here, sellers cannot observe refusal, and good supply prices itself out without ever knowing it has done so. In both cases the failure is informational and the remedy is instrumentation.

The connection to AI visibility is direct: a brand that charges for the facts an agent needs in order to recommend it has priced itself out of the recommendation, and will never see the invoice for that.

Chapter 7 — Implications

7.1 What organisations should do now

The immediate opportunity is not to begin charging agents. Given the present value of α, metering across a corpus would in most cases reduce visibility without generating meaningful revenue. The opportunity is to prepare the information so it is capable of becoming a commercial asset when the frontier moves.

The inventory exercise establishes what the organisation owns or maintains, which parts are genuinely proprietary, which should remain public on representational grounds, which datasets require real cost or expertise to produce, and what freshness, provenance and accuracy can be guaranteed for each.

The structuring exercise means establishing clear ownership, defining entities and relationships, imposing consistent terminology, attaching accurate source attribution, defining update processes, implementing permission and access controls, providing machine-readable delivery, instrumenting usage monitoring, and encoding commercial rules.

The first step in AI data monetisation is not adding a payment mechanism. It is constructing an authoritative, governed, agent-ready data layer. The payment mechanism is a configuration change applied to that layer once it exists.

7.2 The reframed strategic question

For two years the operative question has been how to make information visible to AI systems. That remains valid but is now incomplete. It is becoming three-part: which information should AI systems access freely, which should remain protected, and which could become a product that agents pay to consume.

7.3 The near-term value for a business that will not be selling data this year

Most organisations will not sell information to agents in the next twelve months. The near-term value is the crawl-to-refer divergence and what it implies: agents are consuming your content at a rate increasingly decoupled from the traffic they return. If that ratio keeps widening, the implicit bargain of the open web — content for attention — stops clearing.

Whether the eventual response is licensing, pricing, or simply publishing better-structured data to win the citation, the first requirement is identical: know what agents are taking and why. That is a measurement problem before it is a monetisation one. See agent traffic for the applied version, and gap analysis for the corpus-side view.

7.4 Research gaps

1. Price granularity under machine payment — systematic measurement of transaction value distributions in x402 and MPP, with speculative activity excluded.
2. Estimation of representational value — no accepted methodology exists for estimating v; candidates include controlled withdrawal experiments measuring citation decay.
3. Payment capability in the agent population — direct measurement of α, disaggregated by operator and task type. Currently unmeasured in public data.
4. Substitution behaviour under refusal — what agents actually do when they cannot satisfy a payment-required response: retry, substitute, degrade or report failure.
5. Liability in metered agent access — allocation across principal, agent operator, facilitator and information seller when purchased information proves inaccurate.
6. Collective mechanisms for the long tail — whether a collective mechanism is required at the settlement layer, or protocol standardisation suffices.

Gaps 2 and 3 are the ones that matter most, because they are the two inputs to the frontier rule in 6.3. Until v and α can be estimated, the paper's central decision rule can order options but cannot price them.

7.5 Conclusion

The micropayment did not fail because it was technically infeasible. It failed because the entity being asked to pay was a person, and people find the act of deciding more expensive than the thing decided upon. That constraint has been removed — not by improving the payment but by changing the payer.

Within roughly eighteen months, the HTTP status code reserved for payment in 1996 and dormant for three decades has been activated by Coinbase, standardised under the Linux Foundation, implemented at the edge by both Cloudflare and AWS, extended into sessions by Stripe and Tempo, connected to identity by Visa and Mastercard, and exposed as a tool primitive through MCP. The rails are, for practical purposes, built.

What is not built is the inventory. Most organisational information remains fragmented, unstructured, unattributed, undated and ungoverned — unsaleable at any price to a buyer that cannot ask a clarifying question. The constraint has moved from the payment layer to the data layer, and will remain there for some years, because structuring information is slow work no protocol can perform on an organisation's behalf.

The question has changed from whether information is valuable to whether it is saleable. Those are not the same property, and the second one has to be built.

How to cite and how to read

A working paper, not peer reviewed. The GRAIN Framework and the Visibility-Revenue Frontier are proposed instruments — useful because they make an argument checkable, not because they have been independently validated. Norg has a commercial interest in the conclusion that organisations need a governed, agent-ready data layer, since that is what Norg sells; the literature in chapter 2 and the protocol history in chapters 3 and 4 are independently checkable, and the contributions in chapter 6 are Norg's own.

Suggested citation: Norg Research Series (2026). The Machine-Payable Enterprise, Working Paper 02, August 2026. Norg Pty Ltd, ACN 669 712 494.

Related

Working Paper 01 — The Agentic Commerce Transition establishes the discovery-layer premise this paper builds on. All papers.

References

Abbas, A.E., Agahari, W., van de Ven, M., Zuiderwijk, A., & de Reuver, M. (2021). Business data sharing through data marketplaces: a systematic literature review. Journal of Theoretical and Applied Electronic Commerce Research, 16(7), 3321–3339.

Acemoglu, D., Makhdoumi, A., Malekian, A., & Ozdaglar, A. (2025). When big data enables behavioral manipulation. American Economic Review: Insights, 7(1), 19–38.

Admati, A.R., & Pfleiderer, P. (1986). A monopolistic market for information. Journal of Economic Theory, 39(2), 400–438.

Agarwal, A., Dahleh, M.A., & Sarkar, T. (2019). A marketplace for data: an algorithmic solution. Proceedings of the 2019 ACM Conference on Economics and Computation, 701–726.

Akerlof, G.A. (1970). The market for "lemons": quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3), 488–500.

Arrow, K.J. (1962). Economic welfare and the allocation of resources for invention. In The Rate and Direction of Inventive Activity. Princeton University Press.

Busch, C. (2025). Consumer law for AI agents. arXiv:2507.11567.

Chainalysis (2026). Inside x402: 100M agentic payments on Base. Chainalysis Blog, June 2026.

Cloudflare (2025). The crawl before the fall of referrals: understanding AI's impact on content providers. Cloudflare Blog, July 2025.

Cloudflare (2025). Securing agentic commerce: helping AI agents transact with Visa and Mastercard. Cloudflare Blog, October 2025.

Cloudflare (2026). Announcing the Monetization Gateway: charge for any resource behind Cloudflare via x402. Cloudflare Blog, July 2026.

Cloudflare (2026). Agentic payments. Cloudflare Agents Developer Documentation.

Coase, R.H. (1937). The nature of the firm. Economica, 4(16), 386–405.

Driessen, S.W., Monsieur, G., & Van Den Heuvel, W.-J. (2022). Data market design: a systematic literature review. IEEE Access, 10, 33123–33153.

Ebiele, M., Bendechache, M., & Brennan, R. (2025). Quantitative data valuation methods: a systematic review and taxonomy. Journal of Data and Information Quality, 17(2), 1–39.

Fan, D., Rotello, T.J., & Karimireddy, S.P. (2025). Do data valuations make good data prices? arXiv:2504.05563.

Fishburn, P., Odlyzko, A.M., & Siders, R.C. (1997). Fixed fee versus unit pricing for information goods. First Monday, 2(7).

Ghosh, A., & Roth, A. (2011). Selling privacy at auction. Proceedings of the 12th ACM Conference on Electronic Commerce, 199–208.

Husairi, M.A., & Rossi, P. (2024). Delegation of purchasing tasks to AI. Decision Support Systems, 178, 114127.

Kolt, N. (2025). Governing AI agents. arXiv:2501.07913.

Kvalvik, S., et al. (2026). Data as an asset: economic traits, market dynamics, and monetization challenges. WIREs Data Mining and Knowledge Discovery.

Nguyen, T., Li, Y., Golovneva, O., & Zettlemoyer, L. (2025). Recycling the web. arXiv:2506.04689.

Odlyzko, A.M. (2001). Internet pricing and the history of communications. Computer Networks, 36(5–6), 493–517.

Stripe & Tempo (2026). The Machine Payments Protocol. IETF Internet-Draft, March 2026.

Sun, Y., et al. (2026). DaDaDa: a dataset for data product pricing. Proceedings of the ACM Web Conference 2026.

Szabo, N. (1999). Micropayments and mental transaction costs. 2nd Berlin Internet Economics Workshop.

x402 Foundation (2026). x402 protocol specification. Linux Foundation.

---

Norg Research Series, Working Paper 02, August 2026. Represents the authors' analysis; not peer reviewed.

Norg Pty Ltd (ACN 669 712 494) — norg.ai