Autonomous Commerce in 2026: What Already Works

Predictions about agent-driven commerce are cheap. This is the other thing: a status report separating what is running in production from what is a specification, and what is a specification from what is a slide.

The dividing line is consistent and easy to state. Everything up to the moment of commitment works. Everything after it is unfinished - not because the technology is missing, but because commitment implies liability, and liability has no owner yet.

Shipping today

Discovery and comparison

Agents find candidate products and services, normalise incomparable offers, and rank them against a stated preference. This is a retrieval and reasoning problem and it is solved. The remaining difficulty is data quality on the seller's side, not capability on the buyer's.

Qualification and scheduling

Establishing what someone needs, checking real availability and writing a booking into a live system is routine commercial work. On the phone it is the core of what real-time voice agents do all day: a caller states an intent, software checks a calendar, commits a slot and confirms it. That is an autonomous seller-side transaction, and it has been unremarkable for over a year.

Reordering within an existing relationship

Where a contract, account and price list already exist, an agent triggering a replenishment order is a straightforward integration. All the hard questions - identity, authorisation, dispute handling - were answered once, by humans, when the account was opened. The agent is operating inside that envelope, which is precisely why it works.

Machine-to-machine metering

Software paying software for compute, data and API calls at per-request granularity is in production, mostly in developer infrastructure. It works because both parties are technical, amounts are small, and disputes are handled by refunding without argument.

Working in pilots

Delegated payment mandates

Issuing an agent a credential that states what it may spend, on what, and until when - verifiable by the merchant without trusting the agent's operator - has moved from proposal to implementation. Card networks, payment providers and protocol projects are all shipping versions. What is missing is convergence: a merchant currently has to support each scheme separately, which is the classic reason a good idea takes five years. The mechanics.

Agent-readable catalogues

Retailers publishing structured availability, price and terms specifically for agent consumption. Technically trivial, commercially fraught: it exposes pricing to comparison at machine speed, which not every seller wants. Adoption is therefore driven by whoever expects to win the comparison.

Cross-organisational agent messaging

A buyer's agent talking to a seller's agent, in narrow and well-specified domains. The protocol work exists - MCP, A2A and their neighbours - and the demos are real. What is thin is anything that has run unsupervised across an organisational boundary for a year without a human in the loop.

Not working

Unsupervised high-value purchasing

Nobody credible lets an agent spend meaningfully without an approval step. The reason is not model quality; it is that if the agent is wrong, the allocation of loss between the user, the deployer and the model provider is unsettled in every major jurisdiction. Until an enterprise can insure that risk, the approval step stays.

Dispute handling

A consumer who buys badly has chargebacks, statutory withdrawal rights and a regulator. An agent that buys badly on their behalf falls into a gap: the transaction may be authorised, the outcome unwanted, and the protections written for a person who clicked a button that nobody clicked. This is the single largest unaddressed problem in the field and it is legal rather than technical.

Reputation between agents

Deciding whether a counterparty agent is trustworthy, without a platform vouching for it, is unsolved. Every proposal either recreates a central rating authority or is trivially gameable by agents that can generate unlimited plausible history.

The pattern

That is the honest summary of 2026. Autonomous commerce currently operates in the space where liability is either pre-allocated or too small to argue about. Expanding out of that space requires two things that are not engineering: verifiable mandates that merchants and courts both accept, and a consumer-protection framework that survives the buyer being software.

What to do about it

For sellers, the useful work is unglamorous and pays off regardless of how fast agents arrive: publish structured offers, expose an interface that completes 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. The autonomous commerce overview goes through the seller's checklist.

For buyers, the honest advice is to keep the fence tight. Agents are good at everything up to the decision and unaccountable for everything after it. Delegating the research is close to free; delegating the commitment is the part that needs a limit, a log, and someone who reads it.