By mid-2026, the mechanics of machine payment look less exotic than they did even two years ago. Software can already trigger bank transfers through standardised interfaces, settle cloud consumption internally, purchase digital services, and optimise procurement against policy rules. The novelty is not that machines can move money. It is that they are increasingly expected to decide when money, compute, data access or contractual commitments should move at all. The central problem is no longer whether a machine can pay, but whether it may commit.
That distinction matters because markets depend not simply on exchange, but on authority. In ordinary commerce, a buyer assumes that the employee signing a contract is authorised, at least presumptively, to do so. The law supplies doctrine, firms supply hierarchy, and human interaction supplies cues. Autonomous agents inherit none of those frictions naturally. They can transact at machine speed, across organisational boundaries, with little shared context. As soon as they do, the scarcest input is no longer liquidity. It is reliable proof of mandate.
Why this matters more than the payments story
The popular framing of an agent economy tends to fixate on payment rails, micropayments and new market venues. That is understandable. Price is legible; infrastructure is visible. Yet the economics of autonomous exchange will be determined more by permissioning than by settlement. A system that can pay instantly but cannot verify its own remit is less a market participant than a liability generator.
In practice, every meaningful machine transaction bundles at least four questions. Is this agent the one it claims to be. Is it acting for a principal that can be identified. Is it permitted to take this class of action up to this threshold under these conditions. And if the action is defective, who bears loss and through what evidentiary trail. Human institutions answer such questions imperfectly but routinely. Software agents require those answers to be explicit, machine-readable and continuously checkable.
Cheap authority changes market structure more profoundly than cheap payments. If firms can issue narrow, revocable, auditable mandates to specialised software counterparts, they can safely decentralise more operational decisions. If they cannot, they will keep high-value choices behind human checkpoints, and the much-discussed machine-speed economy will remain confined to low-stakes optimisation.
The missing layer: machine-readable mandate
Digital identity debates have usually concerned people and organisations. The agent economy introduces a third object: delegated operational authority. An autonomous purchasing agent may need permission to source components within a spending band, from approved jurisdictions, avoiding suppliers with certain risk flags, and only against specified service-level terms. That package is not well described as identity alone. It is a mandate, closer to a bounded power of attorney embedded in code and policy.
Existing trust frameworks point in this direction without fully resolving it. The European Union's eIDAS regime created legal underpinnings for electronic identification and trust services, and its 2024 update for the European Digital Identity Framework broadens the architecture for reusable credentials. NIST's digital identity guidance similarly separates identity proofing, authentication and federation. But the practical challenge in autonomous commerce is not merely authenticating an actor. It is attaching a live set of permissions, constraints and liabilities to that actor in a form another machine can verify before accepting an offer.
In human commerce, ambiguity is tolerable; in machine commerce, it is an attack surface. If an agent's authority is vague, stale or opaque, counterparties will either reject transactions, raising coordination costs, or over-trust them, magnifying fraud and operational damage. Neither outcome supports deep markets.
The central problem is no longer whether a machine can pay, but whether it may commit.
The central problem is no longer whether a machine can pay, but whether it may commit.
From procurement to hiring: where authority bites first
The first large-scale use cases are likely to be more prosaic than the rhetoric suggests. Enterprise procurement is a natural candidate because it already runs on codified approvals, supplier lists and budget rules. An agent can solicit bids, compare terms and route exceptions. Yet even here, the difficult step is not ranking offers; it is proving to the seller that the agent is authorised to buy on behalf of a real principal, with enforceable terms, and that the purchase will survive later internal challenge.
Hiring among agents presents the same issue in sharper form. As software systems commission other software systems for narrowly defined tasks, one agent may need to subcontract data labelling, compliance checking, simulation, design generation or logistics scheduling. These are not merely API calls. They are contingent commitments that affect budgets, intellectual property, confidentiality, service quality and sometimes regulatory exposure. An agent hiring another agent therefore needs credentials not just of identity, but of competence scope, data rights and spending authority.
That requirement changes pricing. A cheap task provider with weak proof of provenance or uncertain authority may be rationally ignored in favour of a slightly more expensive counterparty whose credentials reduce verification and dispute costs. Markets then compete not only on price and performance, but on the quality of machine-verifiable institutional trust they can attach to each transaction.
Authority as an economic good
Seen this way, verifiable mandate is itself a factor of production. It lowers transaction costs in a manner familiar from institutional economics: reducing the need for monitoring, renegotiation and ex post dispute resolution. The better an ecosystem becomes at expressing and validating delegated authority, the larger the set of decisions that can be outsourced to software without unacceptable risk.
This suggests a less noticed path by which autonomous systems could reshape firm boundaries. Ronald Coase's old question, why activities occur inside firms rather than through markets, turns partly on transaction costs. If software agents can cheaply verify the precise authority of external counterparties, some coordination that today remains internal for control reasons may move outward. But the reverse is also true. If only a handful of platforms or large integrators can provide trusted authority layers, market openness could diminish even as automation expands.
Research on digital platforms has long shown how control over rules, access and verification can concentrate power. In agent markets, whoever mediates machine-readable authority may become more important than whoever processes payment. That is an uncomfortable prospect for competition policy because the relevant bottleneck may appear, superficially, to be mere compliance plumbing.
The fraud problem becomes organisational, not merely technical
Classical cyber-security asks whether credentials were stolen or messages altered. Agent commerce adds a quieter threat: valid credentials used for invalid purposes. A machine may be correctly authenticated and still exceed its remit, exploit poorly drafted policies, or induce another system into commitments that no sensible principal intended. Such failures are rarely spectacular at first. They look like over-ordering, policy drift, duplicate contracting, procurement leakage or silent accumulation of low-value obligations that only later reveal systemic significance.
NIST's AI Risk Management Framework is relevant here because it insists on governance, documentation and traceability rather than treating AI risk as a narrow model-performance problem. In market settings, that principle points to a practical requirement: every autonomous commitment above trivial value needs a legible chain from principal to policy to action. Not because auditors enjoy paperwork, but because counterparties and insurers will increasingly demand evidence that an agent's authority was real, bounded and current when the transaction occurred.
The most expensive agent fraud may therefore arise not from external attackers, but from internal incoherence: stale permissions, unclear ownership, conflicting policy engines, and missing revocation paths. Human organisations often muddle through such defects. Autonomous ones will industrialise them.
Cheap authority changes market structure more profoundly than cheap payments.
Why revocation will matter as much as issuance
Traditional markets devote disproportionate attention to onboarding and too little to withdrawal. The same bias is visible in digital identity systems, which are often better at issuing credentials than rescinding them cleanly across a network of relying parties. For autonomous agents, revocation is central. Budgets change, suppliers become restricted, incidents occur, staff roles shift, models degrade, and regulations update. Authority that was appropriate yesterday may be dangerous today.
An effective agent economy therefore needs fast, interoperable revocation and expiry conventions. Credentials must be granular enough to limit exposure but simple enough to validate at scale. There is a genuine trade-off. Highly expressive policies capture organisational nuance, yet create computational and legal complexity; simple credentials are easier to process, yet may force blunt trust decisions. The winning architectures are unlikely to be the most elegant. They will be those that make revocation routine, cheap and well logged.
Cheap authority changes market structure more profoundly than cheap payments.
Liability will be priced into every serious transaction
Once agents begin contracting with other agents, every transaction acquires a hidden insurance question. If a purchasing agent exceeds a cap, if a scheduling agent books prohibited routes, if a design agent licenses tainted input, who pays. Courts and regulators will answer some of this over time, but markets will not wait for doctrinal perfection. They will use proxies: narrower mandates, richer logs, approved credential issuers, human sign-off thresholds and higher prices for ambiguous counterparties.
That means authority design will shape spreads and market depth. A seller facing uncertain counterparty authority will charge for expected dispute costs or refuse the trade. Conversely, high-confidence authority can compress verification overhead and support thinner margins. In this sense, delegated mandate acts rather like collateral. It reassures the other side that the apparent buyer can actually bind the resources on which the trade depends.
There are analogies here with electronic transferable records and digital trade documentation. Legal frameworks became useful not simply because records were digitised, but because control, possession and evidentiary status were clarified. Agent transactions need an equivalent settlement of institutional facts: not merely what was agreed, but by whom, under which authority, and with what durable proof.
Public sector procurement may become the proving ground
One underappreciated venue for this evolution is government procurement and administration. Public bodies operate with formal delegations, audit obligations and high sensitivity to procedural legitimacy. They are therefore awkward adopters of unconstrained autonomous systems, but attractive adopters of tightly bounded agents whose authority is explicit. A municipal system that can procure routine maintenance within strict thresholds, or negotiate standard renewals under published rules, offers a cleaner test case than fanciful visions of fully autonomous commerce.
The state also has reasons to care beyond efficiency. If autonomous procurement becomes common in private markets, tax administration, customs control, sanctions enforcement and competition oversight will all need machine-legible ways to inspect authority chains. The same credential structures that enable agent trade may become instruments of regulatory visibility. That creates tension. Businesses will seek confidentiality and speed; regulators will seek auditability and intervention points. Institutional design, not technical possibility, will decide the balance.
In human commerce, ambiguity is tolerable; in machine commerce, it is an attack surface.
Interoperability is less about standards than about semantics
There will be no shortage of standards efforts, credential formats and protocol proposals. Yet interoperability failures in agent markets are likely to stem less from syntax than from semantics. Two organisations may both express a spending cap, but define spend, category, jurisdictional risk or exception handling differently. One firm's notion of approved supplier may incorporate labour, carbon and security criteria; another's may mean only registered vendor. Machines can exchange such fields flawlessly while still misunderstanding one another.
The consequence is economic, not merely technical. If authority semantics are idiosyncratic, counterparties must perform bespoke mapping, which recreates the very frictions automation was meant to remove. Industries with mature taxonomies and predictable contracting patterns will therefore adopt autonomous exchange earlier than sectors where obligations remain highly contextual or politically contested. Agent markets will not spread uniformly. They will cluster where semantics can be disciplined.
This is another reason the near future looks more industrial than consumer. Households rarely articulate preferences and permissions with the precision required for continuous delegated commerce. Firms, by contrast, already codify budgets, roles and tolerances. The agent economy is likely to be built first in back offices, supply chains and regulated operational workflows, not in the shopping baskets of ordinary consumers.
The competition question hiding in the trust layer
There is a strategic danger in all this. If robust proof of authority becomes essential to participation, the institutions that issue, verify or broker that proof could emerge as gatekeepers. OECD work on competition in the age of AI has warned that scale advantages in data, compute and ecosystems can entrench incumbency. To that list one should add trust infrastructure. A market where a few actors effectively certify whose agents are recognisable, reputable and insurable may be efficient in one sense while stifling entry in another.
Policy should therefore distinguish between legitimate assurance and exclusion by opacity. Open legal recognition of portable credentials, common validation rules and contestable trust-service markets would reduce the risk that authority becomes a privately toll-gated chokepoint. The issue is not ideological preference for openness. It is practical market design. A genuinely plural agent economy requires that smaller firms can delegate to software without first joining a dominant administrative stack.
What mid-2026 really suggests
The evidence so far points to a sober conclusion. Autonomous systems are not waiting for some grand, separate machine economy to arrive. They are entering ordinary commerce through narrow mandates, embedded in existing institutions, one bounded workflow at a time. The constraints they face are less about cognitive capacity than about legal and organisational legibility.
That is why the decisive innovations of the next few years may look dull from a distance: better credentialing, finer-grained delegation, standard revocation, stronger audit trails, clearer liability allocation. Yet these are the foundations on which large markets rest. Without them, software agents remain clever operators behind a human cashier. With them, they become counterparties that other systems can trust enough to trade with directly.
The long-run implication is easy to miss. If authority can be specified and verified cheaply, firms may reorganise around portfolios of machine delegates much as they once reorganised around enterprise software. Management then shifts from issuing individual approvals to designing constitutions for agents: mandates, limits, exception paths, and evidence rules. The economics of the agent age will depend less on whether machines can think like traders than on whether institutions can encode permission with the precision that traders, human or otherwise, require.
Payments may become invisible. Proof of authority will not.
