From Market Automation to Market Architecture. When AI Stops Serving the Market and Starts Constituting It

Synthocracy Institute Working Paper
Martin Novak
July 2026

Abstract

Artificial intelligence is usually discussed as a force that automates markets: it searches more quickly, compares more products, predicts demand, detects fraud, recommends suppliers, negotiates prices, and reduces transaction costs. This description remains accurate, but it is becoming incomplete.

A system that accelerates an existing market process is a tool. A system that determines which firms can be discovered, which products count as comparable, which credentials are valid, which risks are acceptable, which transactions may proceed, and which actions can be executed is becoming part of the market’s architecture.

This paper examines the transition from tool to recommendation, delegation, execution, infrastructure, and market power. It argues that the critical threshold is crossed when participation in exchange becomes materially dependent on machine-readable identity, agent-mediated discovery, protocol compatibility, automated scoring, digital mandates, and platform-controlled execution. At that point, AI no longer merely helps actors operate within a market. It helps define the conditions under which the market exists for them.

The emerging agentic market increasingly resembles an operating system. It contains identities, roles, permissions, registries, protocols, schedulers, security layers, administrators, logs, and procedures for suspending or reversing actions. These elements may improve efficiency, expand access, and make transactions more reliable. They may also create new forms of exclusion, gatekeeping, dependency, and responsibility without control.

The central question is therefore no longer only whether AI makes markets more efficient. It is:

At what point does infrastructure stop facilitating exchange and begin determining the conditions under which exchange is possible?

Keywords

Market synthocracy; agentic commerce; agentic procurement; market architecture; AI agents; market power; protocol power; commercial admissibility; machine-readable access; digital mandates; ceremonial human; market contestability.

Claim Status and Scope

This paper distinguishes three kinds of claims:

(A) Empirical or documentary claims describe systems, standards, products, regulations, or institutional developments that can presently be observed.

(B) Normative or argumentative claims advance an interpretation of those developments or propose how markets should be understood and governed.

(C) Foresight claims describe plausible future developments. They are scenarios, not established facts.

The paper does not claim that autonomous agents already control entire markets, that all AI-mediated commerce is harmful, or that every technical standard is an instrument of domination. It argues that a structural transition has begun and that its consequences cannot be understood through the language of automation alone.

1. The Market Did Not Disappear. Its Interface Changed

(A) Markets have always depended on infrastructure. Roads determine which goods can reach buyers. Ports determine which routes are economically viable. Payment systems determine which promises can become settled transactions. Standards determine whether products are interchangeable. Legal systems determine which contracts can be enforced. Credit systems determine who can act before possessing the required capital. Search systems, marketplaces, advertising networks, and app stores determine which sellers become visible.

Infrastructure has therefore never been neutral in the simple sense. It makes some forms of exchange easier, others more expensive, and some practically impossible.

What changes with agentic AI is not the discovery that infrastructure matters. It is the depth at which infrastructure can participate in the decision itself.

A traditional road does not usually decide whether a supplier is trustworthy. A communications network does not ordinarily interpret a product specification, evaluate a certificate, predict a buyer’s urgency, rank eligible sellers, negotiate a discount, select a payment method, and initiate fulfilment. An agentic system may perform all of those functions within one connected process.

(A) The technical foundations of this shift are becoming visible. The Agent2Agent protocol provides a common interaction model for agents built by different vendors and using different frameworks. Its Agent Cards describe an agent’s identity, endpoint, capabilities, authentication requirements, and skills so that other agents can determine whether and how to interact with it. The protocol’s current documentation also includes signed Agent Cards intended to establish trust before interaction across organisational boundaries. (a2a-protocol.org)

The Universal Commerce Protocol is designed as a common language connecting platforms, agents, businesses, and payment providers across commerce processes from discovery through checkout and beyond. Google is already using it to support direct purchasing through AI surfaces such as Search and Gemini. (Google for Developers)

The Agent Payments Protocol adds typed, cryptographically supported mandates intended to record who approved a purchase and under what authority. In the protocol model, UCP can describe what is being ordered and from whom, while AP2 provides evidence of intent, authorisation, and transactional limits. (developers.googleblog.com)

NIST has meanwhile established an AI Agent Standards Initiative focused on agents that act autonomously, securely on behalf of users, and interoperably across the digital ecosystem. A related NIST project specifically examines agent identity and authorisation, including how existing identity and access-management standards may apply when software acts on a person’s or organisation’s behalf. (NIST)

These developments solve genuine technical problems. Agents cannot operate reliably across organisations without ways to identify one another, describe their capabilities, obtain data, communicate, prove authority, and record execution.

But the same mechanisms also create a new political-economic question.

If an agent must be discoverable through a particular registry, recognised by an accepted identity provider, compatible with a required protocol, authorised by an approved mandate system, connected to an eligible payment mechanism, and trusted by a platform’s risk engine before it can participate, then the infrastructure is no longer merely transmitting market activity.

It is deciding what counts as valid market activity.

That distinction lies at the centre of market synthocracy.

The Synthocracy Institute defines synthocracy as a decision order in which humans formally remain in charge while the real work of detecting, filtering, ranking, recommending, routing, and executing decisions increasingly passes through AI systems. The human may still approve or sign, but the system has already shaped what is visible, relevant, credible, and actionable.

Applied to commerce, the concept does not mean that an artificial intelligence becomes the legal owner of the economy. It means that machine systems increasingly prepare the environment in which firms are discovered, evaluated, compared, admitted, negotiated with, and executed.

A market may therefore remain legally human while becoming operationally synthetic.

2. Automation and Architecture Are Not the Same Thing

(B) Market automation accelerates actions whose governing structure remains substantially intact. Market architecture determines the rules and conditions under which those actions can occur.

The distinction can be stated simply.

An automated system may send a purchase order more quickly.

An architectural system helps determine:

which suppliers are eligible to receive the order;
which product descriptions count as interpretable;
which credentials are recognised;
which risks exclude a seller;
which alternatives enter the comparison;
which agent possesses authority to negotiate;
which payment routes are permitted;
and which records will later count as evidence of a valid transaction.

Automation changes the speed of a process.

Architecture changes the space of possible processes.

A spreadsheet that calculates total cost is automation. A system that decides which cost variables count, which suppliers enter the spreadsheet, which offers are treated as equivalent, and which result becomes executable has moved closer to architecture.

A chatbot that answers a customer’s question is automation. An agent that determines which merchants can be searched, which products are eligible, which offers are presented, and which purchase can be completed inside the interface participates in constituting the market available to the user.

A sourcing tool that formats an RFQ is automation. An agent that identifies requisitions suitable for autonomous processing, selects or recommends suppliers, prepares an event, communicates with bidders, evaluates responses, and initiates downstream contracting occupies a more consequential position.

(A) Oracle describes an Autonomous Sourcing Agent that identifies requisitions eligible for autonomous negotiation, prepares sourcing events, invites suppliers, and sends notifications according to company policies. Oracle’s broader agentic procurement environment is presented not simply as a dashboard but as an execution layer that monitors negotiations, identifies priority actions, and turns recommendations into executable decisions. (Oracle)

(A) SAP similarly describes Autonomous Spend Management as a move from disconnected process support towards agentic systems that orchestrate activity across procurement, travel, expenses, external workforce management, and other business processes. Its Autonomous Enterprise model combines an AI platform for constructing and governing agents with a suite intended to execute core operations. (SAP News Center)

These systems may produce important benefits. Low-value procurement can consume disproportionate human time. Automated comparison may identify savings. Structured policy enforcement can reduce inconsistent practice. Agents may help firms respond more quickly to supply disruption, detect missing information, or open opportunities to suppliers previously overlooked.

The presence of a benefit does not settle the question of power.

A mechanism may be useful and still become constitutive. Indeed, useful systems are more likely to become infrastructure because firms have strong reasons to integrate them deeply.

The core issue is therefore not whether the agent is beneficial or harmful in the abstract. It is whether the agent’s judgments become conditions of participation that affected actors cannot inspect, contest, bypass, or replace.

3. The Six-Stage Transition

The move from automation to market architecture is not a single event. It can be understood as a progression through six stages:

tool → recommendation → delegation → execution → infrastructure → market power

The stages may overlap. Different sectors and organisations will reach them at different times. A system may also occupy one stage for some tasks and another stage for others.

3.1 Tool

Claim status: A/B

At the tool stage, AI assists a human who remains the effective author of the market decision.

The tool may:

translate a supplier document;
extract data from an invoice;
summarise a contract;
calculate demand;
draft an RFQ;
identify duplicate records;
or search a catalogue.

The human still determines the objective, selects the relevant universe of options, decides what evidence matters, interprets the result, and initiates action.

The central governance question is primarily one of reliability:

Did the tool produce an accurate and useful output?

A tool can still create harm. A translation error can distort contractual meaning. A summary can omit a critical clause. A demand forecast can contribute to over-ordering. But the tool does not yet govern the structure of participation merely because it performs a task.

The threshold begins to move when the system influences not only how a human processes options, but which options the human encounters.

3.2 Recommendation

Claim status: A/B

At the recommendation stage, the system ranks, scores, filters, or prioritises possible actions.

A recommendation engine may identify:

the most suitable suppliers;
the lowest-risk transactions;
the preferred payment terms;
the products most likely to be purchased;
the contracts most likely to require attention;
or the bids most likely to meet organisational goals.

A recommendation does not formally bind the human. Yet it reshapes the decision environment.

The first item receives more attention. A green status appears safer than an amber one. A vendor described as “recommended” begins with an advantage. A system-generated summary may define the apparent meaning of a complex bid before a decision-maker reads the primary documents.

This is where assistance begins to become co-decision. The Synthocracy framework locates the transition not only at the final act but upstream, where systems shape what people see, trust, approve, or never review.

A recommendation becomes structurally powerful when several conditions combine:

The system controls the initial option set. Its criteria are difficult to inspect. Its output is integrated into the default workflow. Rejecting the recommendation requires additional effort. The human lacks time to reconstruct the underlying analysis. Agreement is frictionless, while disagreement must be justified.

At that point, the recommendation may remain legally optional while becoming operationally authoritative.

3.3 Delegation

Claim status: A/B

At the delegation stage, the human or organisation authorises an agent to perform a defined category of action.

The agent may be authorised to:

search for suppliers;
disclose selected business data;
request quotations;
negotiate within a range;
accept substitutions;
reserve inventory;
place orders below a threshold;
initiate payment;
or delegate parts of the task to other agents.

Delegation is not equivalent to disappearance of human authority. Human institutions delegate constantly. Employees act for companies. brokers act for clients. Payment processors act under instructions. Procurement professionals act within policies.

The difference is the combination of scale, speed, opacity, and recursive delegation.

A human representative usually has a recognised role, a legal identity, professional duties, and a socially intelligible chain of responsibility. A network of software agents may divide a task among a purchasing agent, supplier-discovery service, risk-scoring model, negotiation agent, compliance agent, payment agent, and logistics orchestrator.

Each component may perform only a local function. The aggregate process can nevertheless produce a consequential decision.

This creates the mandate problem:

Who authorised what, for whom, for which purpose, within which limits, for how long, and with what right to delegate further?

(A) The emergence of AP2’s typed mandates and NIST’s work on agent identity and authorisation demonstrates that the distinction between identifying software and proving its right to act has become a concrete technical issue. (developers.googleblog.com)

(B) Identity is not authority. Capability is not permission. Access is not mandate. A system may be technically able to place an order without being legitimately authorised to choose the supplier, reveal the buyer’s urgency, or accept a contractual term.

Delegation becomes architectural when the definition of valid authority is controlled by a technical or platform layer on which participation increasingly depends.

3.4 Execution

Claim status: A/B

At the execution stage, the system can change the state of the market rather than merely describe or recommend a change.

It may:

transmit an RFQ;
reject an ineligible supplier;
alter a price;
reserve stock;
create a purchase order;
sign through an authorised mechanism;
release payment;
book freight;
schedule production;
or initiate a return.

Execution matters because the cost of error changes.

An incorrect recommendation can still be reconsidered. An executed transaction creates dependencies. Inventory is held. Capital is committed. delivery schedules move. Other agents react. A supplier may begin production. A payment provider may record an obligation. A logistics system may allocate capacity.

The point of meaningful control therefore moves earlier.

A final approval button does not prove that the human controlled the decision if the system already:

defined the eligible options;
ranked the suppliers;
negotiated the terms;
reserved the stock;
prepared the award;
and made refusal economically costly.

The human remains present, but presence is not the same as control.

The Synthocracy corpus describes the ceremonial human as a person who remains the visible bearer of responsibility while the conditions producing the decision are distributed across models, vendors, data, policies, rankings, workflow defaults, and institutional pressures. The problem is not that the human is fictional. The problem is that real responsibility can persist after real control has become thinner than the language of accountability suggests.

In markets, this may be the buyer who approves an award without seeing excluded suppliers, the executive who accepts a sourcing policy whose agentic consequences are difficult to reconstruct, or the salesperson held responsible for a price generated by a system using signals they cannot inspect.

3.5 Infrastructure

Claim status: A/B

A system becomes infrastructure when many actors reorganise their behaviour around it and when leaving it becomes costly.

Infrastructure does not have to be legally mandatory. It can become practically mandatory because:

transactions occur faster through it;
counterparties expect its use;
insurers or auditors prefer it;
internal systems are built around it;
reputation accumulates inside it;
alternative channels receive less attention;
or manual routes cannot match its speed and cost.

A procurement agent used by one firm is an application.

A procurement architecture used across many large buyers, connected to standard identity services, approved supplier registries, payment providers, audit systems, and logistics networks becomes something closer to a market layer.

The technical ecosystem developing around agentic interaction supports this transition.

(A) A2A standardises how agents describe themselves and communicate. Its documentation recognises that discovery methods may vary across public, private, and controlled environments. (GitHub)

(A) Agentic Resource Discovery has been proposed as an open specification for publishing, discovering, and verifying agents, skills, tools, and other AI capabilities across the web. (developers.googleblog.com)

(A) UCP standardises commercial interaction from discovery through checkout, while AP2 adds payment-authorisation mandates and an audit trail. (developers.googleblog.com)

Taken together, these are not yet a single operating system for the global market. They are, however, components from which market-operating layers can be built.

Infrastructure becomes market-constituting when it no longer merely provides an optional channel but defines the practical requirements for being recognised, trusted, compared, and executed.

3.6 Market Power

Claim status: B/C

At the market-power stage, control over infrastructure can shape the behaviour and opportunities of actors who do not meaningfully control that infrastructure.

Traditional market power may arise from ownership of scarce assets, control of distribution, network effects, switching costs, superior data, or the ability to influence price and output.

Agentic market power can include those forms while adding new ones:

power over machine discovery;
power to define valid agent identity;
power to determine which mandates are recognised;
power to score commercial trust;
power to establish the default option set;
power to classify products as comparable or non-comparable;
power to route transactions;
power to make one execution path frictionless and another burdensome;
power to preserve or deny reputation portability;
power to decide which records count as proof.

The next gatekeeper may not only own the marketplace. It may own or control the agent that constructs the market for the buyer.

(A) The European Commission’s 2026 review of the Digital Markets Act identifies AI-related questions involving interoperability, self-preferencing, data access, cloud dependencies, and whether some AI services may need to be assessed as virtual assistants under the DMA. (Digital Markets Act (DMA))

(A) The OECD’s 2026 work on AI markets similarly emphasises that economic gains will be stronger where markets remain open and contestable across the AI value chain, while warning of structural competition risks arising across different layers of the market. (OECD)

(B) These debates should not be limited to competition between model developers. As AI agents become intermediaries in non-AI markets, market power can migrate into discovery, identity, ranking, protocol implementation, and transaction execution.

4. The Agentic Market as an Operating System

(B) The operating-system analogy is useful because it makes the architecture visible in ordinary language.

An operating system does not merely perform one task. It establishes the environment in which other tasks become possible. It recognises identities, assigns permissions, allocates resources, schedules processes, exposes interfaces, maintains logs, handles errors, and determines how applications interact with underlying capabilities.

An agentic market may increasingly contain equivalent functions.

4.1 Identities

The system must know which company, person, service, agent, or device is acting.

Identity may include:

legal entity identity;
domain ownership;
cryptographic credentials;
platform account;
agent identity;
beneficial ownership;
tax status;
payment identity;
or organisational affiliation.

Identity answers: Who or what is this?

It does not by itself answer: What may it do?

4.2 Roles

A market actor may operate as:

buyer;
seller;
marketplace;
broker;
payment provider;
verifier;
auditor;
logistics provider;
risk assessor;
agent operator;
or mandate issuer.

Roles frame expectations and obligations. A seller’s agent should not silently act as an impartial comparison service if it is optimising for the seller’s margin. A platform agent should disclose when it is promoting products from the platform’s own commercial ecosystem.

4.3 Permissions

Permissions determine whether an actor may:

read data;
query a catalogue;
reveal inventory;
request a quote;
negotiate;
reserve;
alter;
order;
pay;
cancel;
delegate;
or access another system.

In conventional software, permissions protect information and systems.

In an agentic market, permissions can also determine economic opportunity.

A supplier whose data cannot be read may never enter comparison. An agent without permission to access a relevant directory may search only a partial market. A buyer’s agent unable to interact with an independent payment service may be routed towards merchants integrated with the platform’s preferred provider.

4.4 Registries

Registries may list:

agents;
suppliers;
certified capabilities;
product categories;
credentials;
approved service providers;
reputational histories;
or revoked identities.

Registries reduce search costs and can improve security.

They can also become admission gates.

The critical questions are:

Who may enter? Who decides? What evidence is required? Can an exclusion be challenged? Is the record portable? Can a firm discover that it was never listed?

4.5 Protocols

Protocols define valid interaction.

They may determine:

message structures;
authentication;
negotiation states;
error conditions;
capability descriptions;
payment mandates;
receipts;
cancellation procedures;
and evidence records.

An open protocol may reduce dependence on proprietary integrations. But an open specification does not by itself produce an open market.

A formally open market layer may still depend on:

a concentrated discovery service;
non-portable reputation;
a dominant identity provider;
a required cloud;
a preferred payment provider;
or proprietary extensions available only to favoured participants.

The relevant distinction is therefore not simply open versus closed code. It is whether market participation remains genuinely contestable.

4.6 Schedulers

A scheduler determines what happens, in what order, and under what priority.

Its market equivalents include systems that decide:

which RFQ is processed first;
which supplier receives a request;
how long bidding remains open;
which exceptions reach a human;
which order receives scarce stock;
which shipment receives priority;
which case is delayed for verification;
and which transaction is executed immediately.

Scheduling is power because time is an economic resource.

A supplier reviewed after an automated award is not meaningfully included. A manual bid arriving after an agentic process has completed is not a competitive alternative merely because the email address technically remained available.

4.7 Security Layers

Security layers determine what is trusted, blocked, isolated, monitored, or escalated.

These controls are necessary. Agentic systems can expose tools, data, payment channels, and operational systems to new forms of manipulation. NIST’s work on secure, interoperable agents reflects a genuine need for stronger identity, authorisation, and system protection. (NIST)

Yet security requirements can also become market barriers.

A large firm may afford certification, secure hosting, specialised audits, signed agent infrastructure, and continuous monitoring. A small supplier may not.

The question is not whether security should be reduced. It is whether security requirements are proportionate, transparent, interoperable, and accessible—or whether they become a justified vocabulary for enclosing the market.

4.8 Administrators

Every operating environment has actors capable of changing rules, approving identities, revoking access, updating standards, or suspending activity.

In an agentic market, administrators may include:

platform operators;
protocol governance bodies;
cloud providers;
enterprise procurement departments;
identity providers;
model developers;
regulators;
payment networks;
or consortiums.

Administration becomes government-like when its decisions materially alter participation and when affected firms lack a meaningful path to challenge them.

The use of the word “government-like” does not mean these actors become states. It means they may perform functions that resemble rule-setting, admission, enforcement, and adjudication within a bounded economic environment.

4.9 Logs and Decision Records

Technical logs answer questions such as:

Which endpoint was called?
When was a message sent?
Which status code returned?
Which account authenticated?

These records are necessary but insufficient for market accountability.

A meaningful decision record should also answer:

Why was the process initiated?
Which suppliers were considered?
Which were excluded before comparison?
Which rule produced the exclusion?
Which data influenced the ranking?
Which mandate authorised each action?
Which agent delegated to another?
Which human could intervene?
Which actions became irreversible?
Which alternatives remained available?

Without this information, an audit may prove that a transaction occurred correctly according to the system while remaining unable to explain why the system constructed that transaction instead of another.

4.10 Rollback and Recovery

A market-operating system needs more than a stop button.

Stopping one agent may not reverse actions already propagated across other systems. Stock may have been reserved. A purchase order may have been issued. Payment may have begun. Freight may have been booked. Production may have been scheduled.

Governance must therefore distinguish:

suspending an agent;
cancelling a task;
revoking a mandate;
cancelling an order;
reversing a payment;
restoring data;
correcting reputation;
changing platforms;
and returning to a human process.

The capacity to execute without the capacity to reverse creates a market that is fast in one direction and politically weak in the other.

5. The Market-Constituting Threshold

(B) Infrastructure begins to constitute a market when it exercises durable influence over the conditions of participation rather than merely improving a voluntary service.

No single indicator is sufficient. The threshold can be tested through five questions.

5.1 Does the infrastructure define the admissible participants?

If the system determines which identities, credentials, data formats, protocols, security conditions, and risk levels qualify a firm for consideration, it is shaping admission before competition begins.

This is the difference between deciding who wins and deciding who is allowed to compete.

The latter may be described as commercial admissibility.

5.2 Does it determine the effective option set?

A buyer may formally have access to thousands of firms while the agent searches only:

one approved registry;
one marketplace;
one group of integrated merchants;
one ontology;
or one class of machine-readable suppliers.

The legally available market and the operationally searched market may therefore diverge.

An architecture becomes constitutive when its partial view is treated as the market itself.

5.3 Does it control the path from information to execution?

A recommendation layer has significant influence. A system that combines discovery, ranking, negotiation, payment, and fulfilment controls a larger part of the transaction’s causal chain.

Vertical integration across these stages can reduce friction. It can also make it difficult to identify where preference, exclusion, or conflict of interest entered the process.

5.4 Is refusal or exit economically realistic?

A system remains closer to a tool when participants can refuse it without losing meaningful access.

It becomes infrastructural when refusing it means:

slower response;
inferior data;
reduced visibility;
loss of accumulated reputation;
higher transaction costs;
inability to interact with major buyers;
or exclusion from standard workflows.

A manual alternative is not meaningful merely because a telephone number or email address exists. It must be timely, economically viable, capable of reaching an authorised human, and able to interrupt the automated process.

The background programme for Market Synthocracy correctly identifies the transition from an optional system to a faster system, then a preferred, default, partner-required, economically necessary, and finally difficult-to-exit system.

5.5 Can the rules be challenged by those affected?

The final test is contestability.

Can a supplier learn that it was excluded? Can it inspect the relevant commercial data? Can it correct an error? Can it appeal a risk classification? Can it move its identity and reputation to another service? Can a buyer choose an external discovery layer? Can regulators observe patterns of exclusion and concentration?

When market rules become difficult to challenge because they are embedded in technical workflows, the system has crossed from facilitation towards governance.

6. Why This Transition Is Happening Now

6.1 The technical stack is becoming composable

Claim status: A

The present agent ecosystem increasingly separates specialised functions.

MCP and related mechanisms connect agents to tools and data. A2A connects agents to other agents. UCP standardises commerce processes. AP2 supplies payment-authorisation mandates. Agentic Resource Discovery addresses the publication and discovery of agents, tools, and skills. (developers.googleblog.com)

This composability reduces the cost of building multi-stage agentic systems. It also allows market decisions to be distributed across several services and organisations.

6.2 Enterprise software is moving from insight towards action

Claim status: A

Major enterprise vendors are increasingly describing agents not only as analytical assistants but as systems that monitor conditions, initiate processes, coordinate specialised agents, recommend executable actions, and automate routine procurement or supply-chain work. (Oracle)

The practical importance is not that every announced capability will immediately work at full autonomy. Product claims should be evaluated critically.

The importance is that enterprise operating models are being redesigned around the expectation that software will act, not merely inform.

6.3 Competition rewards speed

Claim status: B/C

A firm that can compare suppliers, negotiate, confirm compliance, reserve inventory, and issue an order within minutes may gain an advantage over a firm that requires several days of manual coordination.

Once some actors adopt these systems, others face pressure to follow. The adoption decision is therefore not made in isolation.

The immediate reason may be efficiency. The cumulative result may be dependency.

Markets do not need to vote for an agentic architecture in one moment. The architecture can emerge through thousands of locally reasonable decisions to remove friction.

6.4 Regulation is beginning to see the gatekeeping question

Claim status: A

The European Commission’s DMA review recognises that AI services can interact with existing gatekeeper platforms, virtual assistants, cloud systems, operating-system functions, and data advantages in ways that raise questions of interoperability and self-preferencing. (Digital Markets Act (DMA))

The OECD is examining both AI infrastructure and the impact of generative and agentic systems on downstream competition, including the possibility that AI may lower some barriers while creating others through data access, model restrictions, or concentrated infrastructure. (OECD)

The regulatory vocabulary, however, still tends to focus on established categories: platform power, AI safety, data access, interoperability, collusion, and consumer protection.

These categories are essential.

Market synthocracy adds a cross-cutting question:

Who controls the decision infrastructure through which the market becomes visible and executable?

7. The Benefits Must Be Taken Seriously

Claim status: B

A credible analysis of market architecture must not treat every form of agentic mediation as capture.

Agentic systems may reduce search costs. A small manufacturer could be discovered by buyers in markets it could not afford to reach through conventional sales channels. Standardised product data may reduce ambiguity. Digital credentials may increase trust. Automated translation may make cross-border commerce easier. Agents may help buyers consider more suppliers rather than fewer. Open protocols may reduce the need for bilateral integrations. Better records may improve auditability. Automated controls may reduce fraud and prevent unauthorised purchases.

AI may also weaken some existing intermediaries. A capable buyer’s agent could compare across multiple marketplaces, challenge a platform’s ranking, or search directly through merchants’ machine-readable catalogues.

The market architecture is not predetermined.

The same technical mechanism can support different institutional outcomes.

A registry can be an open directory or a controlled gate.

A reputation system can be portable or captive.

A protocol can allow independent implementation or become functionally dependent on one dominant platform.

A mandate can protect users or become a formal consent object that hides how the option set was constructed.

A human review stage can restore judgment or merely transfer liability.

The relevant question is therefore not whether agents should be allowed in markets. They already are entering them.

The question is which design choices preserve genuine competition, agency, traceability, and the capacity to challenge power.

8. Counterarguments and Limits

8.1 “Markets have always depended on standards and intermediaries”

This is correct.

Product codes, payment rails, logistics standards, credit agencies, and marketplaces already shape participation.

The argument of this paper is not that agentic infrastructure invents mediated markets. It is that AI can integrate interpretation, prediction, ranking, delegation, and execution across a wider portion of the transaction.

The more functions one layer controls, the more consequential its rules become.

8.2 “The human still chooses the agent and can override it”

Sometimes this will be true.

But the presence of a choice must be assessed institutionally rather than rhetorically.

Can the user change the agent’s search scope? Does the user see omitted alternatives? Can the organisation inspect the scoring criteria? Is override practical under workload and time constraints? Does rejecting the recommendation carry a professional cost? Has the process already moved into execution?

The Synthocracy framework does not deny that humans continue to act. It asks whether they still occupy the point where action can meaningfully be shaped.

8.3 “Open protocols prevent concentration”

Open protocols can substantially improve interoperability and reduce lock-in.

They do not guarantee an open market.

The internet itself demonstrates that open lower layers can coexist with highly concentrated discovery, advertising, identity, cloud, and platform layers.

Agentic markets may similarly become open below and concentrated above.

The protocol may be public while the most valuable discovery service, reputation system, payment credential, or user interface remains controlled by a small number of firms.

8.4 “Regulators can address any problems later”

Some problems can be corrected retrospectively. Others generate path dependence.

If one discovery system consistently directs transactions towards already successful suppliers, those suppliers receive more volume, generate more data, improve their measured performance, and become even more likely to be selected.

Early ranking differences can therefore become structural market differences.

By the time concentration becomes visible, the market may already depend on the architecture that produced it.

8.5 “This is a distant future”

Direct, fully autonomous, cross-market commerce remains limited and uneven.

The paper does not claim otherwise.

But the relevant components—agent communication, capability discovery, identity, authorisation, commerce protocols, payment mandates, autonomous sourcing, and enterprise execution—are already moving into standards and products. (a2a-protocol.org)

The governance problem begins before universal adoption, while the defaults are still being designed.

9. Principles for an Open Agentic Market Architecture

Claim status: B

This paper does not provide a complete constitution for agentic markets. It proposes a starting set of principles for further research.

9.1 Separate identity from authority

A verified identity should not be treated as evidence of unlimited authority.

Every consequential action should be connected to a mandate specifying:

the principal;
the purpose;
the scope;
the duration;
the financial or operational limits;
delegation rights;
and revocation conditions.

9.2 Make discovery scope visible

A buyer should be able to know whether an agent searched:

the open web;
one platform;
an approved supplier list;
an integrated partner network;
or a restricted commercial registry.

Without scope disclosure, a narrow search may be misrepresented as an objective view of the market.

9.3 Preserve commercial data correction

Firms should have a meaningful path to correct:

certificates;
ownership information;
product parameters;
delivery histories;
risk classifications;
environmental data;
complaints;
and other machine-readable commercial records.

An uncorrectable commercial profile can become an invisible exclusion order.

9.4 Require reconstructable decision records

Technical logs should be supplemented by records sufficient to reconstruct:

who was considered;
who was excluded;
which rules applied;
which agents acted;
which mandates authorised them;
where a human could intervene;
and what could be reversed.

9.5 Protect portability and switching

Identity, credentials, reputation, and transaction history should be portable where legally and technically possible.

A firm should not have to surrender its accumulated commercial existence when changing platforms.

9.6 Audit discovery and ranking, not only final transactions

A transaction may be procedurally valid while the option set was unfairly constructed.

Audit must therefore examine upstream stages:

visibility;
admission;
classification;
ranking;
routing;
and prioritisation.

9.7 Make human control structurally real

A human review stage should not be accepted as sufficient merely because a person clicked approve.

Meaningful control requires:

time;
relevant evidence;
access to alternatives;
authority to pause;
the ability to alter criteria;
protection for disagreement;
and an institution capable of learning from overrides.

The right to say no is not real when the workflow punishes the person who uses it.

9.8 Design for stoppability and reversibility

Systems should distinguish between stopping future actions and reversing completed ones.

The authority, technical capability, and conditions for suspension, rollback, mandate revocation, and re-admission should be defined before deployment.

9.9 Monitor concentration across the full stack

Competition analysis should look beyond the final marketplace to:

chips;
cloud;
foundation models;
agent frameworks;
discovery;
identity;
reputation;
payment;
and execution.

Control may emerge through a combination of layers rather than dominance in only one.

9.10 Preserve a functional alternative

Where markets affect essential goods, public procurement, critical supply chains, or small-business access, a non-agentic or alternative-agent route may need protection.

The route must be more than nominal. It should be capable of producing a timely and economically meaningful result.

10. A Research Programme for Market Synthocracy

Claim status: B/C

The transition described here requires empirical study rather than only theory.

The next research questions should include:

Commercial admissibility

Which requirements cause firms to be excluded before comparison? How often do data quality, protocol incompatibility, identity, certification, or risk thresholds determine participation?

Machine-readable market access

What technical capabilities are becoming necessary to transact with agent-mediated buyers? Which costs fall disproportionately on small and medium-sized firms?

Agentic discovery

Who operates the directories and registries through which agents find counterparties? Are discovery results auditable, contestable, and portable?

Mandate chains

Can organisations reconstruct authority across multiple delegated agents? Where does responsibility break when an agent delegates to another service?

The ceremonial human

How much information, time, and authority do human approvers possess? How frequently do they override agents, and what happens when they do?

Market concentration

Do open protocols reduce dependence, or do identity, data, reputation, cloud, and payment layers remain concentrated?

Synthetic path dependence

Do early agent selections reinforce the market position of selected firms by generating more data, better reputation, and further selection?

Manual fallback

When does an alternative channel remain legally available but cease to be economically viable?

Market observability

Which aggregated indicators would allow regulators to detect invisible exclusion, self-preferencing, or algorithmic concentration without requiring disclosure of every model parameter?

Sector studies

Procurement, logistics, manufacturing, chemicals, packaging, finance, labour markets, and public purchasing will not develop identical architectures. Each requires a sector-specific map of identities, standards, permissions, risks, and appeal rights.

Conclusion: The Market Before the Transaction

The conventional view of the market begins with an exchange.

A buyer wants something. A seller offers it. Information is compared. A price is agreed. A contract is formed. Goods, services, and money move.

The agentic market begins earlier.

Before exchange, a system must decide what can be seen.

Before comparison, it must determine what is comparable.

Before trust, it must determine which identities and credentials count.

Before negotiation, it must determine who has authority.

Before execution, it must determine which actions are permitted.

Before accountability, it must determine which traces will remain.

This is the new architecture of the market.

It is not necessarily a machine dictatorship. It may be built through ordinary procurement software, helpful assistants, open protocols, safer payment systems, faster logistics, and commercially sensible integrations.

That is precisely why it requires attention.

Power rarely announces that it has become infrastructure. It appears as a reduction in friction, a cleaner interface, a trusted score, an approved profile, an easier checkout, a recommended supplier, and a process that works so efficiently that the alternative becomes difficult to imagine.

Market automation becomes market architecture when AI systems no longer merely help participants act, but define the practical conditions under which participants can appear, qualify, compare, negotiate, transact, and appeal.

Market architecture becomes market power when those conditions are controlled by actors whom affected participants cannot meaningfully inspect, challenge, replace, or leave.

The central question of the agentic market is therefore no longer only who owns the product, sets the price, or signs the contract.

It is:

Who controls the infrastructure that determines which firms can be discovered, admitted, compared, trusted, negotiated with, and executed—and who can challenge that power?