Pillar
Autonomous commerce: buying without a buyer present
Software that searches, compares, negotiates, books and pays - while the person who wanted the outcome is doing something else. Parts of this ship today. The parts that do not are held back by liability, not by capability.
The definition, and the useful distinction inside it
Autonomous commerce is a transaction in which at least one side is represented by software that selects and commits without step-by-step human approval. The important word is commits. A recommendation engine is not autonomous commerce; a price comparison is not autonomous commerce. The line is crossed when the software creates an obligation - a booking, an order, a payment.
It helps to separate two very different cases that get discussed as one. Agent-to-service commerce is an agent buying from a normal business through normal interfaces; it is happening now. Agent-to-agent commerce is two pieces of software negotiating terms with each other; it exists in pilots and in narrow machine-to-machine settings, and almost nowhere else.
The seven steps, and how far each one is
- 08RecourseDisputes, refunds and liability when the agent was wrong. Unallocated in every major jurisdiction.Unsolved
- 07CommitmentWhether an agent’s agreement binds the person behind it. Contract law was not written for this.Unsolved
- 06SettlementActually moving the money. Rails exist - stablecoins, mandates, HTTP-native payment schemes.Emerging
- 05AuthorisationProving the agent was allowed to spend this much, on this, now. Standards drafted, adoption early.Emerging
- 04NegotiationAgreeing terms, slots or price with the other side. Works in narrow, well-specified domains.Emerging
- 03SelectionWeighing options against a stated preference and choosing. Reliable.In production
- 02ComparisonNormalising incomparable offers into a decision. Reliable.In production
- 01DiscoveryFinding candidate products, services and providers. Solved.In production
The gradient is unmistakable. Everything up to and including choosing is solved - that is a retrieval and reasoning problem, and retrieval and reasoning work. Everything after choosing runs into questions that are institutional rather than technical: whose money, whose fault, whose court. A status report on the parts that ship goes through this in more detail.
Why payments are the hinge
Card networks assume a cardholder who can be shown a total and asked to confirm. Take away the confirmation step and the entire fraud and chargeback apparatus loses its anchor. This is why agent payments are not a matter of plugging in an existing API: the authorisation model has to change, not the transport.
Two approaches are converging. Delegated mandates issue an agent a credential describing exactly what it may spend, on what, within what window - verifiable by the merchant without trusting the agent's operator. Stablecoin and HTTP-native payment schemes handle amounts and frequencies that card rails price out of existence, which matters when an agent pays a tenth of a cent for a structured answer several thousand times a day. Can AI agents make payments? works through both.
An agent books a non-refundable flight to the wrong city. Who carries the loss - the user who delegated, the company that deployed the agent, or the provider of the model that made the error? No jurisdiction has a settled answer, and until one does, every serious deployment keeps a human approval step in front of anything expensive.
What it looks like from the seller's side
Most writing about this takes the buyer's perspective. The seller's is more immediately practical, because businesses will be selling to agents long before they are buying with them.
- Be readable by machines. Availability, price, terms and constraints in structured form. An agent that cannot parse your offer treats you as absent.
- Be callable. An interface that completes a booking without a session cookie and a rendered form. Agents that must drive a browser will do it, badly, and blame you for the failures.
- Be answerable. Decide whether an agent may hold a slot, and for how long. Inventory contention between agents is a real operational problem long before it is an interesting one.
- Know who is asking. "Which agent, acting for whom" turns into a pricing and abuse question the moment agent traffic is non-trivial. See agent identity.
One rung of this is already ordinary. Businesses that answer inbound calls with a real-time agent are, without framing it that way, already operating an autonomous seller-side channel: the caller states an intent, software checks live availability, commits a slot and confirms it. Products such as bitpull.ai do this over the phone network today. When the caller is itself an agent - which will not require any new invention - both halves of an autonomous transaction are in place, and the only missing piece is a way for each side to prove what it was authorised to agree to.
Autonomous commerce will not arrive as a launch. It will arrive as the moment nobody notices that neither party on a call was human.
The machine economy
Extend the same logic past retail and it produces something genuinely new: services priced and consumed by software at a granularity no human purchasing process could administer. Compute by the second, data by the record, an API answer by the answer, settled continuously between parties with no contract and no invoice. This is what people mean by the machine economy, and it is the reason settlement infrastructure keeps reappearing in a conversation that started about language models.
Whether that economy is large or a curiosity depends on one thing: whether an agent's commitments can be made binding in a way a court recognises. Everything else is engineering.
Frequently asked
What is autonomous commerce?
A transaction where at least one side is represented by software that selects and commits without a person approving each step - an agent that searches, compares, books and pays on someone’s behalf. The defining moment is commitment, not recommendation.
Can an AI agent legally buy something for me?
It can technically, and in low-value contexts it already does. Whether the resulting contract binds you is unsettled: it depends on how the mandate was granted, whether the merchant could verify it, and which jurisdiction applies. This is why serious deployments cap what an agent may commit to.
What has to exist before agent-to-agent commerce is normal?
Three things: verifiable identity for non-human actors, machine-readable authorisation that states spending limits and scope, and a dispute path that does not require a human to reconstruct what happened. The models are already good enough; these three are not in place.
How should a business prepare to sell to agents?
Publish offers in structured form, expose an interface that can complete a transaction without a browser session, decide explicitly whether agents may hold inventory, and be able to tell which agent is asking and for whom. That work is useful even if agent traffic stays small.