Which layer are you on when the machines start buying?
Autonomous purchasing is being built in layers, and the layers are already owned. For a manufacturer or a distributor the useful question is not whether agents will buy. It is which layer your business sits on when they do, because that decides whether you are quoted or skipped.
In one paragraph, no jargon. Buying software is starting to ask suppliers questions by itself: what do you have, how many, how fast, what does it cost me. Four pieces of that conversation have already been standardised by the big platforms, and three of the four are theirs. The fourth one, the piece that lets anyone ask your catalogue a question, has to sit inside your own systems.
Why that matters to a distributor. When the assistant cannot get an answer from you, it does not wait. It takes the next supplier on the list, and you never see the request that you lost.
Four layers, four owners, one of them yours
Read the picture first, then the detail below it. One of these four bands belongs to you. The other three belong to Google, OpenAI, Stripe, Coinbase and the foundations they handed their standards to.
Layer one is the door
The door an assistant knocks on when it wants to know something about your catalogue.
Layer two is the counter
How the order actually gets placed once the assistant has decided to buy.
Layer three is the permission
Proof that a real buyer, with real limits, allowed this purchase to happen.
Layer four is the change
The payment, measured in cents per question rather than invoices per month.
Access
The catalogue as a callable tool
An agent does not browse your site, it calls functions. It asks for a part, a price, a lead time and a document, and it acts on the answer it receives. This is the only layer of the four that has to live inside your infrastructure.
Without this layer the other three cannot help you: an agent cannot complete a purchase from a catalogue it cannot read.
Donated to the Linux Foundation, Dec 2025
About 97M SDK downloads a month
Commerce
From discovery to order
The cart, the checkout and the order management described so that any agent can buy at any merchant without a bespoke integration per pair. Behind the standard sits a council that already includes Amazon, Meta, Microsoft, Salesforce, Stripe, Mastercard and Visa.
This layer decides the route to the order. It does not decide whether your catalogue can answer the question the agent asks on the way there.
profile at /.well-known/ucp
ACP, OpenAI with Stripe, 2025
Mandate
Proof a real buyer authorised the spend
The question that blocks every enterprise from letting an agent transact: who mandated this purchase, within which limits, and can it be proven afterwards. Google's protocol answers it with signed mandates and moved to the FIDO Alliance in April 2026 with more than sixty organisations behind it.
For a distributor this is the layer that makes agent orders acceptable to your own finance and legal teams, not only to the buyer.
FIDO Alliance, Apr 2026, 60+ organisations
v0.2 added human-not-present
Verifiable Intent, Mastercard, compatible
Settlement
Paying in fractions of a cent
Machine purchases are micropayments rather than invoices. A call costs a fraction of a cent, which sets the economics of the whole stack: nobody gets rich on settlement, and everyone who wants margin has to own discovery or the contract.
This is why the commercial value sits one layer up from where most people expect to find it.
138M+ payments by Q3 2026, 87% under ten cents
MPP, Stripe with Tempo
Eighteen months, and the first deadline is already set
This is not a prediction about the future. Every marker below is a standard that shipped, or a compliance date that a regulator has already published.
The places an agent checks before it trusts you
None of these existed in commercial use three years ago, and all of them are free to publish. Of the 189 we measured, 26 publish a real llms.txt and 10 declare an authorisation server; none publishes the commerce profile.
| Path | What it declares | Across 22 distributors |
|---|---|---|
| /.well-known/ucp | The commerce profile: what can be bought, how, and under which terms | 0 published |
| /.well-known/agent-card.json | Who the agent is talking to, and what it may ask for | 0 published |
| /.well-known/oauth-protected-resource | How an agent is authorised to act on a buyer's behalf | 1 published (dxpe.com) |
| /llms.txt, /agents.txt | A plain description of the catalogue for a machine, in the convention models look for | 26 real file (mscdirect.com) |
| robots.txt | Whether named AI agents are permitted or refused | 0 name an AI agent |
A robots.txt that never mentions an AI agent is not a neutral default. It means the decision has not been taken, and the crawler takes it for you. 94 of the 189 refuse a machine outright or hand it a bot challenge, and from the outside a traffic rule against scrapers is indistinguishable from a policy against agents, which is exactly why we repeat the run and hand over the transcript rather than a score alone.
Four conclusions for an industrial business
Discovery and pricing are the toll booth
Settlement earns fractions of a cent per call. Being the catalogue an agent reads is a contract, and the contract is worth thousands of times more than the clearing fee.
Contract pricing is the moat
A list price is public and commoditised within a quarter. The price for this buyer, inside their own agreement, is not, and no platform can publish it for you. That is why the layer sits in your systems.
The missing piece is the connector
Payments companies own the rails, platforms own the storefront. Neither built the layer that makes an ERP-resident catalogue answerable in the buyer's own context, with an audit trail, without the data moving.
Readiness pays before volume
The work that makes a catalogue answerable to a machine also fixes what your human buyers complain about: two prices for one part, a lead time that contradicts the warehouse, a datasheet three revisions old.
Why this may be early
From our own research rather than from a vendor deck.
0.2% of sessions
Agent-initiated volume is small, and most production settlement is still test traffic.
over 40% by 2027
Gartner expects more than 40% of agentic projects to be cancelled before the end of 2027.
Two of four still contested
Commerce and settlement each have more than one live standard. We say in writing when the timing is wrong for you.
Where the claims above come from
Each line is checkable. Where a fact is contested among sources we say so rather than pick the flattering version.
MCP governance
Anthropic open-sourced the protocol in 2024 and donated it to the Linux Foundation in December 2025. Download figures are Anthropic's own published numbers.
UCP and ACP
Universal Commerce Protocol was launched by Google with Shopify at NRF in January 2026 and publishes its profile at a well-known path on your own domain. Agentic Commerce Protocol is the OpenAI and Stripe route. Both shipped code in 2026, and the commerce profile consolidated on UCP's side.
AP2 and mandates
Google announced the Agent Payments Protocol in 2025 and moved it to the FIDO Alliance in April 2026 with more than sixty organisations. Version 0.2 added the human-not-present flow. Mastercard's Verifiable Intent is designed to be compatible rather than competing.
Machine settlement
x402 was published by Coinbase and now sits under the Linux Foundation, with more than 138 million payments by the third quarter of 2026 and 87% under ten cents. Stripe and Tempo published MPP.
Agent share of traffic
About 0.2% of sessions on enterprise sites in the figures we could verify; 68% of US searches end without a click; 11% of consumers say they would let an AI buy for them.
Find your position on the stack
Send a domain. We will show you which layer your catalogue currently answers on, and what it takes to answer on the one that matters.
hq@arliwork.com · Shenzhen Capernaum Technology Co., Ltd · PIAKHOUSKI DZMITRY