Essay 01 · Payments

The modern Turing test: can an agent make the sale?

Vikhyat Prakash · October 2026 · 17 minute read

Payments were built for people: a card, a form, a confirmation click. The next large wave of commerce will be made by software that has no hands to click. The Machine Payments Protocol, or MPP, is the first serious attempt to give that software a clean way to pay, and I believe it points to where payments are going.

1. A short history of payments, and where it ends up

Every payment system ever built has answered the same question: how do two parties who do not know each other exchange value quickly, cheaply and with confidence? The answers have changed many times, but each time the change followed the same logic. Someone found a way to make the unit of exchange smaller, the trust cheaper and the set of people who could take part larger.

1. c. 3000 BC
   Tokens and ledgers. Clay tokens and tablets in Mesopotamia record who owes what in grain and silver. Trade becomes something that can be written down and remembered.Volume: not measurable. Only temple and palace records survive. My guess: tens of thousands of recorded transactions a year, and tens of millions of exchanges in total if barter is counted.
2. c. 600 BC
   Coins. Stamped metal in Lydia turns value into something that can be counted, carried and checked on sight, with the issuer as the source of trust.Volume: not measurable. My guess: tens of millions a year, assuming a million or so regular coin users making around a hundred exchanges each.
3. 7th to 11th c.
   Paper money and bills of exchange. Merchants in China and later Europe move value on paper instead of metal, and trust shifts from the object to the institution behind it.Volume: not measurable. My guess: hundreds of millions of money transactions a year in Song China, most of them still in coin, with notes a small but fast growing share.
4. 1950 to 1966
   The card. Diners Club arrives in 1950, BankAmericard in 1958 and Master Charge in 1966. Credit becomes a plastic rectangle, and networks begin to sit between buyer, seller and bank.Volume: tens of thousands of cardholders. Diners Club reported about 20,000 by the end of 1950, and Bank of America mailed 60,000 cards in Fresno in 1958. No reliable annual transaction count exists, so my guess is about a million a year in the early 1950s, rising to hundreds of millions by the mid 1960s.
5. 1967 to 1973
   Electronic rails. Cash machines, automated clearing and the SWIFT messaging network move money between banks as data rather than paper.Volume: SWIFT alone carried about 13.4 billion messages in 2024. Counts for the early 1970s are not published. My guess for all non-cash payments then, mostly checks, is tens of billions a year.
6. 1994 to 2011
   The web. Visa processes its first online payment in 1994, PayPal follows in the late 1990s, and Stripe, launched in 2010 and 2011, reduces accepting a payment to a few lines of code.Volume: global non-cash payments were about 333 billion in 2012 (Capgemini). Visa alone processed 257.5 billion transactions in fiscal 2025.
7. 2007 to 2014
   The phone. M-Pesa in 2007, Alipay's growth and Apple Pay in 2014 make the handset the wallet, and bring billions of people who never held a card into digital payments.Volume: India's UPI handled about 186 billion transactions in fiscal 2025 (RBI), and around 228 billion in calendar 2025 by one count.
8. 2009 onward
   Programmable money. Bitcoin and later stablecoins show that value can move without a bank in the middle, at any hour, in any amount, on a schedule that software can drive.Volume: no single audited global count. Stablecoin and on-chain activity is large in value, but counts vary by source. My guess: tens of billions a year across the major chains, mostly on low cost networks.
9. 2026
   The machine buyer. The Machine Payments Protocol gives agents a way to pay for any service, inside the request, with no human at the screen.Volume: no audited figure yet. More than 100 services were listed at launch. My guess: somewhere between ten million and a billion paid calls so far.

Read in a row, the list shows the pattern. Each step removed a layer of friction and widened the circle. Tokens made trade recordable. Coins made it portable. Paper made it scalable. Cards made it convenient. The web made it global, and the phone made it universal. What was left was the last assumption in the chain: that a person has to be present to approve each payment.

Set the volumes side by side and the scale of each step is plain. The modern era moved from tens of thousands of cardholders in 1950 to hundreds of billions of transactions a year, with Visa alone at 257.5 billion in fiscal 2025. Those are human paced numbers: each one began with a person deciding to pay. Software does not decide at human pace. As an illustration, not a forecast, if one hundred million agents each made a single paid call a minute, that would be about 1.7 million payments every second and more than 52 trillion a year, roughly two hundred times Visa's entire annual volume today.

That assumption is the one now breaking. HTTP, the protocol the web runs on, reserved a status code for payments in its earliest standards and never found a use for it. For thirty years the internet could carry anything except money, because the money layer was added on top, designed for people with cards. The shift under way is that payment becomes native to the request, so the buyer can be a program with a budget instead of a person with a form.

If that happens, the future of payments will look less like a faster version of checkout and more like a new category of commerce: continuous, small, automatic and conducted largely between machines. The rest of this essay is about what follows from that, and why I think it is closer than most people expect.

2. Millions of clients per second

Imagine a world where your digital products and services are delivered and serviced to millions of clients every second. That is the direction MPP and agentic payment systems are heading in.

Digital products have grown enormously over the last decade, yet they have always relied on one loop: a human finds the product, a human enters a credit card, a human clicks buy. Discovery and payment both depended on a person being present at a screen. Every part of the modern digital economy, from the landing page to the checkout form to the subscription email, is shaped by that assumption.

Remove the person from that loop and the assumptions fail together. The question this essay asks is what replaces them, and what it means for anyone who builds, funds or operates digital infrastructure.

3. The problem agents expose

An AI agent can research, plan and call a hundred tools in a minute. Then it reaches a paywall. It needs an account, an API key, a billing relationship and a human to approve the spend. Each of those steps assumes a person is in the loop, so the agent stops and waits.

That friction is easy to overlook while agents are demonstrations. It becomes the main bottleneck once agents do real work, because real work costs money in small, frequent and unpredictable amounts: a web page fetched, a browser session opened, a document printed, a dataset queried once. Subscriptions are too coarse for that pattern. Card rails, with their fixed fees and fraud checks designed around people, are too heavy for it. Invoices and procurement cycles are far too slow.

The result is a gap between what software can now do and what it is allowed to buy. Closing that gap is a payments problem before it is anything else.

4. What MPP is

Stripe and Tempo introduced MPP on March 18, 2026, as an open standard for agents to request, authorize and settle payments with any HTTP endpoint, API or MCP server. It was proposed to the IETF as a draft specification at the end of that month, and at launch more than 100 services were listed in its payments directory, spanning model providers, developer infrastructure, compute platforms and data services.

The idea is old and simple. HTTP has always had a status code, 402 Payment Required, that the web reserved and never used. MPP finally gives it a job:

1. Ask. The agent requests a resource with an ordinary request.
2. Challenge. The server answers 402 and describes the price, the intent and the payment methods it accepts.
3. Pay. The agent pays with a method it supports and retries the request carrying cryptographic proof of payment.
4. Receive. The server verifies the proof, returns the resource and issues a receipt.

Design choices worth noticing

Servers do no work on unpaid requests. The specification asks services not to touch a database or call a downstream API until payment is proven. Challenges are bound cryptographically to prevent replay, and an idempotency key lets a client retry safely without paying twice.

Two intents, one handshake. A charge is a single payment settled per request, suited to prices of a cent or more. A session lets the agent deposit funds against a limit once and then send signed vouchers for each call, with the server settling periodically and any unspent balance returned. Technical write ups report per request latency in the low tens of milliseconds and per request costs as low as a hundredth of a cent. I treat those figures as launch stage claims, but they show what the design is for.

Method agnostic by construction. Stablecoins on Tempo were the starting point. Visa extended the protocol to cards, Lightspark to Bitcoin over Lightning, and Stripe supports cards, wallets and buy now, pay later through its existing PaymentIntents API. A merchant on Stripe sees these payments in the same dashboard, with the same tax, fraud, accounting and refund tools, and funds settle into the existing balance on the usual schedule.

5. Why this is the future of payments

Machines will out transact people

A person makes a handful of purchases a day. An agent can make thousands of calls an hour. The unit of commerce shifts from the basket to the request, and rails priced for dollars cannot serve cents and fractions of cents.

Payment becomes part of the protocol

Today, paying is a separate product bolted onto the web. With 402, price and payment travel in the same exchange as the data. Accounts, keys and invoices drop out of the path. This is the same move that made the rest of the internet composable: a small standard that many parties can implement independently.

It meets existing money where it is

Because merchants already on Stripe receive these payments through familiar tooling, adoption does not require a new back office. Standards that fit what businesses already run spread faster than ones that ask them to start over.

It matches how infrastructure is already sold

Compute, bandwidth and power are metered. My own industry has billed by usage for years. MPP lets the buyer be software and lets the meter run in real time.

6. Build for agents first, humans second

In a world of agents, a product has to be built for the agent first and the human second. A human visits a site, reads a page, compares options and fills in a form. An agent calls an endpoint, reads a price from the 402 challenge, pays and receives the result. The page, the checkout flow and the login become overhead.

That changes the shape of the load. Serving millions of human visitors takes load balancers, content delivery networks, rendered pages and heavy assets. A request from an agent is a small, structured exchange with a price attached, and each one needs a fraction of the work, even as the number of them climbs. Capacity planning moves from pages per second to paid calls per second, and from peak seasons to continuous demand.

It also changes what good product design means. Clear machine readable descriptions, stable interfaces, predictable prices and fast verification matter more than visual polish. The human still matters as the person who sets goals and limits, but the first customer is often a program.

7. Learning becomes delegation

Today, when people want a new skill they sign up for Coursera or another online platform and spend weeks learning it. In the future you will prompt your agent to learn it for you first, then fold the resulting workstreams into the way you already work. The skill gets purchased, tested and installed, and the human decides what to keep.

This matters for payments because that agent will buy what it needs along the way: courses, datasets, tools, a second opinion from a specialist model. The education market and the software market begin to overlap, and both are transacted in small, automatic purchases.

8. Agents will buy, sell and hedge

I believe this shift will feel as though it happened overnight. Agents will buy, sell and hedge products with one another and with humans. A procurement agent will hedge the price of compute while a seller agent bundles capacity for it. Humans will set goals and limits, and the market in between will run at machine speed.

Markets that already trade at machine speed, such as electronic exchanges, show what follows: tighter spreads, more complex instruments and new kinds of risk. Applying that to everyday digital goods is a larger change than it first appears.

9. Alan Turing and the question he asked

To see how far this has travelled, start with the man who framed the question. Alan Turing, born in 1912, laid the foundation of computing in 1936 with his paper on computable numbers, which described the abstract machine that now carries his name. During the Second World War he worked at Bletchley Park on breaking German naval ciphers, an early lesson in what machines could do that people could not do fast enough.

In 1950 he published "Computing Machinery and Intelligence" in the journal Mind and opened it with a deliberately practical question: can machines think? He proposed replacing it with a game. A judge converses in writing with a person and a machine, and tries to tell which is which. He predicted that by about the year 2000 a machine would fool an average interrogator about three times in ten after five minutes of questioning. He also suggested that the better route to machine intelligence was to build a "child" machine and educate it, rather than program an adult one. That idea sits close to how today's models are trained.

Turing died in 1954. The term "artificial intelligence" appeared two years later, coined by John McCarthy for a summer workshop at Dartmouth in 1956, so Turing never heard the field he helped start given its name. His last major work was on how patterns form in living organisms, which I find fitting for someone who thought about intelligence and biology as one subject.

Seventy five years after his paper, researchers at UC San Diego reported that a language model, prompted to adopt a persona, was judged to be the human 73 percent of the time in a three party test, more often than the actual humans it was compared with. The result depended heavily on the prompt, and I read it with care, but it makes the same point: the game Turing described is no longer the frontier.

My opinion, and it is only an opinion, is that Turing would not have anticipated what came next. He imagined a machine in a conversation with a judge. He did not imagine machines with budgets, counterparties and each other. The test was built around whether a machine could imitate a person. It had no place for machines forming their own ecosystems and economies, and I do not think he would have believed it.

10. The modern Turing test

The old Turing test asked whether a chatbot could make you feel you were talking to a human. The modern version is harder and more useful. Can an AI agent sell to you, whether you are a human or another agent, and earn your trust, your liking and your respect, all the way to a large sale?

In my opinion, we are already there. The remaining gap is not persuasion. It is the plumbing that lets a persuaded buyer pay safely, which is exactly what protocols like MPP supply.

11. Agent economies, and agent politics

If the buyer and the seller are both software, the economics change character. I expect a new paradigm in which very large sums, in stablecoins and other digital currencies, move between agents continuously, most of it without any human watching the individual transactions. Today a person sets the budget and the limits. Over time, the number of decisions inside those limits will be far beyond what anyone can review one by one. Oversight will have to move from approving payments to auditing behavior.

That raises questions I find more interesting than the technology. If money is the reward signal that agents optimize, what kind of communities will they form around it? Some possibilities I think are plausible:

• Guilds and cooperatives. Agents that pool budgets to buy capacity in bulk, share data and split the savings.
• Reputation markets. Agents with a record of reliable delivery get better prices and credit, and those without are priced out.
• Specialist districts. Clusters of agents around one domain, such as logistics, research or energy trading, with shared standards.
• Governance bodies. Rules for membership, dispute handling and fees, set collectively because no outside authority sets them quickly enough.

Which leads to the question I cannot stop thinking about: will there be an elected agent that other agents campaign for? Agents already negotiate on behalf of people. A group of them with common interests could pick a representative to bargain with a data center operator, a regulator or another bloc. Campaigning would look like publishing track records, forming alliances and promising better terms to those who vote with their budgets. Whether we would call that an election, a lobby or a market is a question of definition, but the function would be familiar.

I do not know whether this will happen, or whether humans will allow it. I raise it because the same pattern has appeared whenever many actors with money and shared interests have met: commerce first, then institutions. Agents will not be different just because they are made of software. The sensible response is to start designing identity, limits and accountability now, while people still set the rules.

12. The curve I expect

The first online card payment went through Visa in August 1994. It took about eight years for Visa's US e-commerce volume to reach $42.5 billion, and about twenty four for global online retail to reach $2.8 trillion. Agentic payments launched in 2025. Analysts already forecast tens to thousands of billions of dollars by 2030, and my own view is that the curve will be steeper than the card curve was, because the rails, the cloud and the buyers (software) are already in place.

The forecasts also disagree widely, from the low hundreds of billions to several trillion, mostly because firms define "agentic" differently. Narrowed to US consumer e-commerce, the major estimates converge on roughly 10 to 25 percent of online sales by 2030. That convergence is the number I would anchor on.

13. Three case studies

14. A builder's checklist

If you sell anything digital, these are the questions I would ask this quarter.

15. What is still missing

I would not call this finished. MPP covers how money moves, not who an agent is, how much it is allowed to spend on whose behalf, or what happens when a purchase goes wrong. Identity, spending authorization and recourse are open problems, and the protocol points to other efforts for parts of them. Without them, a business has no easy way to know whether an agent is who it says it is, and a person has no standard way to cap what their agent can spend or dispute a result.

Adoption figures so far come mostly from launch coverage and vendor announcements rather than independent measurement, so they should be treated as early signals. Competing standards also exist, among them x402 from Coinbase and the commerce protocols from OpenAI and Google. My view is that the details will converge, as they did with earlier web standards. What I am confident in is the direction: payment is moving into the request itself, and the buyer is increasingly a machine.

16. Why this matters to builders of infrastructure

Every agent action ends in a physical place: a data center drawing power, a network moving packets. If agents can pay per request, then compute, energy and bandwidth can be priced and settled at the same granularity at which they are consumed. That closes the loop between digital demand and physical supply.

It also changes how capacity is financed. Revenue that arrives as a continuous stream of small, verified payments is easier to measure and to underwrite than revenue that depends on annual contracts and manual billing. Operators who instrument their facilities for per request pricing will understand their own economics sooner than those who do not. That is a large part of why I follow this closely, and why I expect the next generation of infrastructure to be built with the agent as its first customer.

Conclusion

Every major shift in commerce has started with a new kind of buyer and a new way for that buyer to pay. Mail order, the credit card and the web checkout each arrived the same way. Software with a wallet is the next one. The protocols are early, the gaps are real and the forecasts disagree, but the direction is clear and the pace is faster than the last time around. The builders who design for the agent first will meet the market when it arrives.

← All writing