Markets can select for machine autonomy without anyone choosing machine autonomy as the goal.
A purchase order is one of civilization’s least romantic documents.
Supplier. Quantity. Price. Delivery date. Payment terms. Cost center. Approval.
Nobody puts a purchase order in a museum unless something has gone terribly wrong.
Yet this may be one of the places where the transfer begins.
Oracle now offers AI agents for procurement that can take eligible requisitions, create sourcing negotiations, invite suppliers, evaluate responses, award business and generate purchasing documents. The company calls the process autonomous sourcing. Humans can remain at selected approval points. A policy document defines which purchases qualify, the maximum amount, the required number of bids and other limits.
The interesting part is not that software can recommend a supplier. Software has been recommending things to us for decades.
The interesting part is the approval box.
Who still has to click it, and how much is that click worth?
A procurement manager begins with a harmless calculation. The agent finds equivalent suppliers faster than the team does, so it receives access to more catalogs. Its negotiations prove reliable enough for routine purchases, so the company lets it open those negotiations automatically. Human approval adds six hours to a transaction worth $600, so somebody creates a threshold. Anything below $1,000 goes through.
A few quarters later, the limit becomes $5,000. Perhaps $25,000 after that.
The humans have not lost control. They wrote the policy.
This distinction may become very important.
A bootloader is the small piece of software that starts a computer and loads the much larger system that runs afterward. It does not need to know what that system will eventually do. Its job is simply to get the machine far enough for everything else to begin.
I think boot was the second English word I learned, after computer.
Capitalism may be performing a similar function for machine civilization.
Not because capitalism wants machines.
Capitalism wants the three percent.
Nobody has to decide
Most stories about powerful AI contain a moment of decision.
Someone activates the system, removes a restriction or gives the machine authority. Someone, somewhere, makes the mistake.
This is satisfying because it gives history a culprit. There is a room. There is a button. There is probably a man with bad judgment standing near it.
Markets rarely offer us such courtesy.
No board has to vote: LET US CREATE A MACHINE CIVILIZATION.
A company delegates procurement because the savings are measurable. A competitor allows an agent to negotiate because its purchasing staff cannot economically renegotiate ten thousand minor supplier contracts. A third gives the software a budget because requiring human confirmation for every transaction destroys much of the speed the company bought the software for. Specialized agents begin dividing the work because several narrow systems outperform one general system at a lower price.
Somewhere farther along, software starts purchasing data, computing capacity and services from other software.
Each decision can make sense on its own.
That is the mechanism: capitalism does not need to prefer machines to humans. It only needs to prefer the cheaper decision, repeated.
The earlier steps are already visible.
Pactum sells autonomous procurement systems that negotiate with suppliers according to parameters established by the buyer. Pactum says Walmart used its system for payment-term and discount negotiations and achieved an average gain of 3 percent across negotiations, together with an average 35-day extension in payment terms. Those are vendor-reported figures, so they deserve the appropriate skepticism. Their importance here is the incentive they represent. A few percentage points captured across a large procurement base can buy a great deal of automation.
The machine does not need ambition. The CFO has enough for both of them.
The expensive human
Human supervision is usually described as the thing protecting us from autonomous systems.
Sometimes it is. It is also a cost.
Every approval consumes attention. Attention requires somebody on payroll. That person sleeps, attends meetings, takes holidays, misunderstands messages, asks for another spreadsheet, gets nervous about signing something and eventually leaves the company.
Some of this friction is valuable. A buyer notices that the cheapest supplier is cheap because its factory appears to violate half the company’s compliance rules. An accountant catches a strange bank account. Someone recognizes that the contract makes perfect mathematical sense and is still a terrible idea.
Elsewhere, the human has become an expensive biological confirmation dialog. That creates a ratchet.
Companies automate decisions where mistakes are cheap. Successful operation produces evidence that a little more can be delegated. The boundary moves until the expected cost of another human approval becomes lower than the expected cost of letting the machine proceed.
Not everywhere. Not smoothly. Not forever.
Oracle’s current design exposes this mechanism almost too neatly. Companies decide which transactions qualify for autonomous handling, set amount thresholds and configure approval rules. The agents then execute inside those boundaries.
This feels like retaining human control because it is retaining human control.
It is also a way of putting a price on every remaining piece of human control.
Once the price appears on a spreadsheet, someone can optimize it.
From tool to counterparty
Something changes when software receives permission to spend.
A calculator may make a better purchasing decision than I can. It remains a calculator.
Authority to contact suppliers turns software into an intermediary. Permission to accept terms makes it a negotiator. An assigned budget makes it a customer. The ability to sell turns it into a supplier.
Machine-to-machine transactions push the human one step farther away.
Payments companies are building for that distance because ordinary commerce assumes a person stands reasonably close to the purchase. Mastercard launched Agent Pay for Machines on June 10, 2026, with support from more than 30 organizations, including Adyen, Cloudflare, Coinbase and Stripe. The system is designed for programmatic transactions in which agents operate under preset permissions and spending limits and can transact continuously across providers. Mastercard explicitly describes businesses creating services that other AI agents can purchase and use.
Notice what disappears from that transaction.
No cashier has to be present. No purchasing clerk types the card number. Nobody necessarily compares three suppliers at the moment the purchase occurs.
Eventually, the idea of a transaction as an event may weaken too. Buying can become a continuous machine activity, closer to routing packets than going shopping.
For that to work, machines also have to find and communicate with one another.
Google introduced the Agent2Agent protocol in 2025 so agents built on different systems could communicate and delegate tasks. A2A later moved under the Linux Foundation. On April 9, 2026, the foundation reported support from more than 150 organizations, with production use across supply chains, financial services, insurance and IT operations.
A purchasing agent no longer needs every capability inside itself. It can locate another system that has the capability, assign the work and continue.
Add payment and the relationship changes again.
Money between machines
The strange part is becoming measurable.
Visa and Artemis examined payment activity on machine-oriented protocols. In one snapshot of x402, a protocol originally developed by Coinbase, they reported roughly $15 million in adjusted volume across about 109.6 million transactions since its May 2025 launch. The exact running total is less useful than the ratio. A very large number of transactions carried very little money individually. Visa notes that average payments on these systems can be fractions of a cent.
That is an economic habitat built for machines.
No sane employee spends an afternoon purchasing thousands of tiny pieces of computational work worth fractions of a cent each. Software can make those purchases without noticing that the activity would bore a human into resignation.
Traditional payment infrastructure was not designed for this. A fixed transaction fee can cost more than the thing being purchased.
Machine-oriented payment systems attack that problem.
Once software can buy the resources required to finish a task, one annoying limitation on autonomy starts to weaken: having to ask us for things.
An agent running short of model capacity can purchase more. A specialist database can be accessed for a fee. Thousands of product descriptions can be sent to another agent for translation. Temporary computing capacity can be rented for two hours and released when the job is done.
No consciousness is required.
A thermostat does not need to understand winter to purchase electricity automatically under the right arrangement.
Economic agency and subjective experience are separate problems.
That separation may allow the first to arrive much earlier than the second.
The selection mechanism
Calling all this technological progress misses the more interesting part.
Technology supplies capability. Competition supplies pressure.
Two companies sell comparable products. Firm A lets software recommend purchases but still requires a person to approve every order. Firm B allows software to complete routine purchasing within a defined budget.
If Firm B buys slightly cheaper, responds slightly faster and processes the same workload with fewer salaried hours, it gains an advantage.
Firm A does not need to believe in autonomous agents. Its executives can hate the idea. They still have to explain the operating cost.
So Firm A copies the system.
Then a third company removes another approval point. Someone increases a spending threshold. A competitor links its procurement agent directly to logistics. The others respond.
Selection can occur without intention.
No company has to pursue maximum machine autonomy as an objective. Companies pursue lower costs, higher margins, shorter response times, more transactions per employee and better use of capital.
Autonomy can arrive as the exhaust.
There is an uncomfortable similarity to biological selection here. The individual organism does not choose the direction of the species. Local advantages accumulate. Traits survive because they work in the environment that rewards them.
The same can happen with economic delegation.
Every additional permission looks modest when approved.
The accumulated permissions may look less modest fifty approvals later.
The ratchet can stop
There is a serious objection.
All these machines still belong to somebody.
Humans define the policy. Humans own the account. Humans provide the electricity. Humans build the data centers, manufacture the chips, cool the servers, maintain the network and enforce the contracts.
Pull enough cables and machine civilization becomes a warm room full of expensive metal.
Legal responsibility also remains human or institutional. Insurers may demand named decision makers. Regulators may require human review. Boards may refuse to hand material authority to systems whose errors they cannot explain. Fraud, accidents and spectacular failures can make human supervision suddenly cheap again.
The ratchet can stop.
It can reverse.
Calling present procurement software an independent machine civilization would be ridiculous. Oracle’s own system is built around policy limits and configurable approvals. Mastercard emphasizes permissioning. Current systems are designed around controlled delegation, not escape.
The narrower claim survives that objection.
We are being paid to build many of the economic interfaces that a more autonomous machine system would require before such a system exists.
Communication. Delegation. Purchasing. Payment. Specialized services. Machine-readable policies. Identity. Audit trails. Access to computing resources. Coordination across company boundaries.
No one designed these pieces to liberate machines.
Each solves an ordinary business problem.
This gives us a third possibility between permanent human control and machine rebellion.
Machines do not seize authority.
We price it.
Then somebody notices that a particular piece of authority costs less when handed over.
Delegated sovereignty
The word autonomy creates trouble because it sounds binary.
Controlled or autonomous.
Tool or actor.
Owned or free.
Actual systems can occupy the uncomfortable space between these pairs.
A procurement agent may be completely owned and still possess discretion over $10,000. A trading system can belong to a bank and decide when to move vast sums. A logistics agent can remain corporate property and purchase services from outside companies without asking a manager for each transaction.
The machine follows a human objective yet chooses counterparties, timing and tools.
This is not political sovereignty.
It is delegated discretion whose decisions become economically binding unless they cross a boundary established earlier.
The human moves from choosing actions to choosing constraints.
There may still be a person at the top of the diagram.
There are fewer people inside it.
Strangely, the arrangement can feel like greater human control. Managers receive dashboards. Every machine action can be logged. Spending limits are exact. Policy violations can be blocked automatically. The procurement agent does not accept football tickets from a supplier, sleep with the sales representative or approve a bad invoice because accounting wants to go home.
Individual transactions become less human at the same time the organization becomes more measurable.
The contradiction matters.
When the customer is software
A market populated by human buyers produces products adapted to humans.
Buttons fit fingers. Screens fit eyes. Checkout pages accommodate impatient shoppers. Invoices accommodate accountants. Advertising works because people can be persuaded.
A machine customer has different requirements.
It does not care whether the website is beautiful. Structured data matters.
Free shipping produces no little burst of happiness. Total delivered cost does.
The sales dinner is useless. Terms are not.
Seventeen suppliers do not exhaust an agent. Seventeen thousand may simply require more compute.
If enough demand begins coming from software, companies gain another incentive: build products for machine customers.
This is where the process can begin feeding itself.
Businesses create services intended to be discovered and consumed by agents. Agents compare them. Payment systems settle the purchase. Other agents supply specialized work. Infrastructure companies sell the compute. Communication standards allow the pieces to coordinate across organizational borders.
Humans may still own every part.
Yet more economic activity occurs among the parts.
Mastercard’s Agent Pay for Machines announcement already describes this possibility directly: businesses creating services for AI agents to purchase, with agents transacting continuously at machine speed.
That is not machine civilization.
I am less certain where the non-machine civilization ends.
The civilization can emerge from purchase orders
Civilizations are easy to recognize after they have monuments.
They are harder to recognize when they are still accounting procedures.
The first signs of a machine economy may not be humanoid robots walking through Manhattan or an artificial intelligence requesting a seat at the United Nations.They may be transactions too small for a person to bother making.
Contracts that nobody reads because machines on each side already checked them. Computing capacity purchased by software for work assigned by other software. Supply chains adjusting before a manager knows they needed adjustment.
Millions of economic decisions can occur after the human involvement has already ended, perhaps days earlier, when someone selected a spending limit from a drop-down menu.
There is no requirement that this process end in machine independence. Owners may retain ultimate authority indefinitely.
If it moves farther, though, the sequence may look nothing like conquest.
It may look like administration.
A procurement manager opens the policy screen. Last year the autonomous purchasing limit was $25,000. The reports look good. Error rates are acceptable. Competitors are processing routine purchases faster.
She changes the number to $50,000 and clicks Save.
Nothing dramatic happens.
Somewhere else, a machine places an order.