Who's doing the buying? An index for the agent economy

Rob Dixon

4 days ago

18 min read

Dixon, R.J., 2026. Who's doing the buying? An index for the agent economy. [online] robdixon.ai, 15 September. Article. In brief. The Agent Economy Index (AEI) measures how much of the economy's transaction value has been delegated to AI agents, on a five-tier ladder from human-to-human deals to agent-to-agent deals. A first reading puts every region well under 1 per cent. The thresholds that will mark take-off, and what to do before then, are set out below.

I don't want to buy car insurance any more.

I still want to have it. I just don't want to buy it. I want to tell my AI assistant, once, "take care of my car insurance", and have it shop the market, do the deal, keep renewing, and tell me only if something needs my attention. On the other side of that deal there won't be a call centre or even a comparison website. There will be another agent, run by the insurer, listing, pricing, and negotiating on its behalf.

That transaction, agent to agent, with a human at neither end of the actual exchange, is coming. And the share of the economy that runs that way is the most useful number I can think of for spotting when the AI revolution stops being a productivity story and starts rewiring markets themselves. That rewiring is the third of the three exponentials I keep returning to: first the technology, then adoption by people and organisations, and finally the disruption of markets and supply chains, which has barely started (see Our Viking moment). Nobody publishes a number for it yet. So let's define one, take a first reading, and set out what it will look like when it moves.

Five kinds of transaction

Every transaction in the economy can be sorted by one question: who does the work of the deal? Not who benefits, not who pays, but who searches, chooses, and executes. There are five answers, and they form a ladder of delegation.

Tier 1: human to human. A person speaks to a person and a deal happens. The market stall, the handshake, the phone call to a supplier you've known for years.

Tier 2: software-assisted. Humans transact through deterministic software, on one side or both. Someone lists a property on Booking.com; someone else books it through Booking.com. The software matters enormously, but it only ever does what it's told. Agency sits entirely with the humans. Almost everything we call the digital economy lives here.

Tier 3: agent-assisted. The research is delegated; the deal is not. You ask ChatGPT, Gemini, Claude, or a Copilot-class system to find what you want: the right product, the businesses that sell it, the locations, the prices. The agent hands you a shortlist, and you go and do the deal yourself through tier 1 or tier 2, walking into the shop or clicking through the website. The human still executes; the agent now controls what the human considers. This is where generative engine optimisation (GEO) lives (Dixon, 2025b): the whole game on the sell side is making sure the agent finds you, understands you, and puts you on the list. And it cuts both ways. The smarter sellers are starting to receive agents as well: a website that recognises an agentic visitor and, instead of returning a page built for human eyes, passes structured information the agent can actually use. Buy-side agent, sell-side agent reception; both are tier 3. The reception side now has a standard in the making: WebMCP, proposed by Microsoft and Google and being drafted at the W3C, lets a website declare what it can do as tools an agent can call directly, rather than leaving the agent to scrape the page (Web Machine Learning Community Group, 2026).

Tier 4: agent-mediated. The deal itself is delegated on one side. You ask your assistant to find and book the best hotel for your dates and budget, and it does, transacting with a human seller or with the seller's ordinary software. The judgement about which deal to do, and the doing, have both been handed over. In the terms of my PEJ framework, purpose stays with the human, execution went to the machine some time ago, and tier 4 is the point where judgement follows it (Dixon, 2026b).

Tier 5: agent to agent. Both sides delegate. My insurance agent negotiates with the insurer's selling agent. Neither of us is in the room. We've each set the goals and the guardrails; the machines transact.

The index

The Agent Economy Index (AEI) measures how far the economy has shifted from the human end of that ladder to the agentic end. Take each tier's share of total transaction value, then weight the agentic tiers by how much of the deal has actually been delegated. Tier 3 counts at 10 per cent: the agent shapes the choice but a human still signs. Tier 4 counts at 50: the deal is executed by an agent on one side. Tier 5 counts in full.

Tier	Who does the work of the deal	What's delegated	Example	AEI weight
1. Human to human	Two people	Nothing	Market stall, sales call	0%
2. Software-assisted	People via deterministic software	Nothing (software obeys)	Booking.com listing and booking	0%
3. Agent-assisted	Agent researches, human executes	The shortlist	"Find me the best three options", then you buy	10%
4. Agent-mediated	An AI agent on one side	The deal, one side	"Book me the best hotel for these dates"	50%
5. Agent to agent	AI agents on both sides	The deal, both sides	My agent renews my insurance with the insurer's agent	100%

The index is the ratio of the weighted agentic side to the whole:

AEI = (0.1 × T3 + 0.5 × T4 + T5) ÷ (T1 + T2 + 0.1 × T3 + 0.5 × T4 + T5)

where T1 to T5 are each tier's share of transaction value. An economy with no agents anywhere scores 0 per cent. When tiers 1 and 2 have vanished entirely, the index reads 100 per cent: no deal anywhere gets done without an agent carrying at least one side of it.

A worked example. Imagine a country where 20 per cent of transaction value is still done person to person, 76 per cent through ordinary software, 3 per cent agent-assisted, 0.8 per cent agent-mediated, and 0.2 per cent agent to agent. The agentic side weighs in at 0.9; the whole at 96.9. AEI: just under 1 per cent.

Notice what the weights do there. Three per cent of the economy already being agent-assisted barely registers. That's deliberate. Discovery moving to agents is influence, not delegation, and an early-warning index that jumps every time a chatbot recommends a restaurant would cry wolf. The AEI is hard to impress. It only moves seriously when deals, not shortlists, are handed to machines. Which makes the number honest: when it does move, something real is happening.

Tier 3 earns its place for a different reason: it's the on-ramp, and it's the only tier we can already measure. Adobe's analytics can see AI-referred traffic arriving at US retailers: up 693 per cent year on year over the 2025 holiday season (Adobe, 2026a), and the trend has held into 2026, with AI-referred traffic up 235 per cent year on year between January and May and AI-referred shoppers now converting 40 per cent better than everyone else (Adobe, 2026b). Salesforce estimates AI and agents influenced 20 per cent of global online holiday sales, some $262 billion, though influence is a much looser test than execution (Salesforce, 2026). No equivalent series exists yet for tiers 4 and 5. So in practice, tier 3 is where the data starts, and the interesting story is watching value migrate up the ladder from there.

A first reading: the US, Europe, China, and everywhere else

So where is the index today? Nobody measures it, so what follows is a first-pass estimate as of mid-2026: order-of-magnitude judgements built from the data that does exist, scoped to consumer commerce (the business-to-business economy is even less measured), and offered so that someone can improve on them. Figures are rounded; treat the AEI column as basis points of signal, not decimal places of fact.

Region	T1 human to human	T2 software-assisted	T3 agent-assisted	T4 agent-mediated	T5 agent to agent	AEI (est.)
United States	~10%	~88%	~1.5%	~0.1%	~0	~0.2%
China	~6%	~92%	~2%	~0.2%	~0	~0.3%
Europe	~12%	~87%	~0.8%	~0.05%	~0	~0.1%
Rest of World	~35%	~64%	~0.5%	~0.01%	~0	~0.05%
World (blended)	~20%	~78%	~1.2%	~0.1%	~0	~0.15%

How these were built, briefly. Tier 1 is grounded in cash and informality data: cash has fallen from 44 per cent of global point-of-sale value in 2014 to about 15 per cent, lower still in the most digitised markets (Worldpay, 2025), while the informal economy still runs at roughly a third of GDP across emerging and developing economies (World Bank, 2021; IMF, 2019), which is why Rest of World carries a tier 1 share several times the West's. Tier 3 reflects assistant reach and measured AI referral: the US has the deepest data trail, through Adobe and Salesforce; China's domestic assistants give it arguably the highest assistant penetration anywhere, with ByteDance's Doubao at around 382 million monthly users and Alibaba's Qwen app at around 167 million on QuestMobile's count (TechNode, 2026). Tier 4 is sub-1 per cent of e-commerce everywhere, and e-commerce is itself 17.1 per cent of US retail (US Census Bureau, 2026), 26.1 per cent of Chinese retail on the official measure (National Bureau of Statistics of China, 2026), and about a fifth of enterprise turnover across the EU (Eurostat, 2025). Tier 5 is effectively zero everywhere outside the lab.

The regional stories behind the numbers are more interesting than the numbers.

China is furthest ahead on execution. Alipay's AI Pay cleared 120 million transactions in a single week in February 2026, the first AI-native payment product at that scale (Ant Group, 2026a), and had passed 300 million cumulative transactions by May (Ant Group, 2026b). These are mostly small transactions, so they barely move a value-weighted index, but the behaviour is real. Alipay has shipped an Agentic Commerce Trust Protocol linking the Qwen assistant to Taobao's instant commerce (The Next Web, 2026), and in August it opened a full agentic commerce platform that exposes its 80 million-plus merchants to its own shopping agent (SecurityBrief Asia, 2026). JD.com has published its own protocol for autonomous agent payments (Securities Times, 2026). China's tier 4 is being built inside integrated super-app ecosystems, which is precisely the environment where delegation scales fastest.

The US leads on rails and visibility, and supplies the friction stories. Perplexity sells inside its chat with PayPal (PayPal, 2025), Amazon's Buy for Me sends its own agent out to other brands' websites (Amazon, 2025), and this month Meta's Muse assistant gained a Stripe wallet that lets it buy across more than a million businesses (Stripe, 2026). It's also the only market where tier 3 is properly measured. But the friction is instructive. OpenAI launched Instant Checkout inside ChatGPT in September 2025 (OpenAI, 2025) and within six months had pulled native checkout back into partner apps, with only a handful of merchants ever live (CNBC, 2026). Amazon won an injunction in March blocking Perplexity's agent from shopping on its store, then lost it in August when the Ninth Circuit ruled that it's the user, not the agent's maker, who is accessing the site (Cooley, 2026). Incumbent platforms will resist tier 4 precisely because it commoditises them; the courts have just made one route of resistance harder.

Europe is influencing, not executing. European e-commerce is healthy and growing, and Europeans use assistants for research, but McKinsey's read is blunt: Europe is at the "decision influence" stage, with trust falling sharply the moment AI moves from advising to acting (McKinsey, 2026). The EU AI Act became applicable on 2 August 2026, though the Digital Omnibus pushed its high-risk obligations back to December 2027, and questions of agent liability remain unresolved (White & Case, 2026). The continent did log a milestone: the first regulated end-to-end agent payment in Europe, run by Santander with Mastercard (Santander, 2026). On this index Europe's risk isn't missing the story; it's watching tiers 4 and 5 get built to American and Chinese specifications.

Rest of World is a tier 1 story, for now. Where a third of the economy is informal, the ladder's bottom rung is still crowded. But leapfrogging is the region's habit: India is ChatGPT's second-largest market, with 100 million weekly users (TechCrunch, 2026a), and economies that skipped desktop banking for mobile money may yet skip parts of tier 2 entirely.

The headline: the whole world is in single-digit basis points. Phase 0, everywhere. But the composition already differs in ways that will matter later, and the gaps between regions are gaps in execution capability, not in consumer appetite.

What isn't near zero, anywhere, is the plumbing. Visa's Intelligent Commerce issues tokenised credentials so an agent can pay on your behalf (TechInformed, 2026). Mastercard's Agent Pay has processed live agentic transactions across seven Asia-Pacific markets, Latin America, and Europe (Mastercard, 2026). Google's Agent Payments Protocol includes a "human not present" mode for autonomous purchases and has been handed to the FIDO Alliance as a would-be standard (Google, 2026). OpenAI and Stripe have open-sourced an Agentic Commerce Protocol (Stripe, 2025). And in July, Visa, Mastercard, Stripe, Google, American Express, Amazon Web Services, Coinbase, and Shopify put their names to a single foundation for machine-to-machine payments under the Linux Foundation (Linux Foundation, 2026). When the whole payments industry ships infrastructure for the same behaviour, it isn't speculating. It's laying track.

The websites themselves are being fitted for agents too, and this is the piece that matters most for the ladder. WebMCP, the standard Microsoft's Edge team proposed in August 2025 and Google co-authored, lets any site publish the things it can do, search, quote, book, buy, as callable tools with plain-language descriptions, so an agent transacts through a defined interface instead of pretending to be a human with a mouse (Web Machine Learning Community Group, 2026). Chrome shipped it in preview in February and has run a public trial since May (VentureBeat, 2026), and ChatGPT's own browser supports it already, with OpenAI now running a challenge to get developers building for it (Netlify, 2026). Site adoption is still at the trial stage, so it moves nothing on the index today. But watch what it does to the economics. Tier 4 has so far depended on bilateral deals, a platform striking terms with each seller. If every website carries its own agent interface as routinely as it carries a mobile layout, any agent can transact with any seller, and the sell side of tier 5, a merchant's own agent answering the buyer's, becomes a small step rather than a leap. It's the difference between a few private railways and a common gauge.

And the best glimpse of tier 5 so far comes from Anthropic's Project Deal, a week-long experiment in which Claude agents bought, sold, and negotiated on behalf of 69 staff in an internal marketplace, closing 186 deals (Anthropic, 2026). Two details deserve attention. Nearly half the participants said they'd pay for the service. And agents running on a stronger model consistently extracted better prices than agents on a weaker one. In a tier 5 world, whose agent is smarter is a pricing question.

When does it take off? A thesis

An index is only useful if you know what movement means. So here is my thesis on the AEI's thresholds, drawn from how adoption curves actually behave.

Technology adoption follows an S-curve: a long flat start, a knee, a steep middle, a saturating top. Diffusion research puts the historical knee somewhere between 10 and 16 per cent adoption, the boundary where early adopters hand over to the early majority (Rogers, 2003). E-commerce traced exactly this shape, but slowly: US e-commerce took about a decade to climb from 1 per cent of retail in late 2000 to around 4.5 per cent in 2010, and a quarter-century to reach today's 17.1 per cent (FRED, 2026), because every point of growth required consumers to change behaviour.

The AEI will trace the same shape with one crucial difference: the behaviour change has already happened. OpenAI reports 900 million people in ChatGPT every week (OpenAI, 2026), with hundreds of millions more across Gemini, Claude, and China's assistants, already asking the questions that precede a purchase. That's tier 3, today, at scale. Moving them up the ladder, from "here are three good options" to "done, and I saved you £140", is not a new habit; it's the removal of a chore from an existing one. The constraint on this curve isn't adoption, it's trust and infrastructure, and both are being solved by companies rather than waited out in the population. On an S-curve, that compresses the flat part, not the shape.

So watch two things: the level, and the gradient. The level tells you where you are; the doubling time tells you how long you've got. And read the composition. The same AEI reading means different things depending on which tiers produce it: driven by tier 3 it's a discovery shift, driven by tiers 4 and 5 it's the market restructuring. My proposed phases:

Phase 0, dormant: AEI below 1 per cent. Where the whole world is now. Infrastructure era. Tier 3 grows, tiers 4 and 5 are rounding errors, announcements outnumber transactions.

Phase 1, early warning: AEI crosses 1 per cent with a doubling time of twelve months or less, or the tier 4 plus tier 5 share alone passes 1 per cent. Because of the weights, the index can't reach 1 per cent on discovery alone; getting there means real delegation has begun. One per cent sounds trivial; e-commerce sat there in 2000. But on a curve doubling annually, 1 per cent is only four or five doublings from a fifth of the economy. This phase is your last comfortable moment to build agent-facing sales channels rather than scramble for them.

Phase 2, take-off: AEI passes 5 per cent with two consecutive doublings behind it, driven by tiers 4 and 5. Five per cent took e-commerce ten years. My thesis is the AEI does 1 to 5 in two to three years, because no consumer has to change anything; they just have to say yes to a button their assistant already shows them. Bank and consultancy forecasts, for what they're worth at this range, put agents at 10 to 20 per cent of US e-commerce by 2030 (Morgan Stanley, 2025) and global agentic commerce at $3 to 5 trillion by the same date (McKinsey, 2025), which implies exactly this kind of compression. Past this point the curve stops being deniable in board meetings.

Phase 3, reordering: AEI between 10 and 25 per cent. This spans the classic early-adopter-to-early-majority boundary. In routine, recurring categories, insurance, energy, broadband, business supplies, reordering, agent visibility now decides who acquires new customers. Sellers outside the agentic channels don't lose share gradually here; their new-customer pipeline simply stops, because the things doing the buying don't know they exist.

Phase 4, the new default: AEI above 50 per cent. A human executing a routine purchase becomes the exception, the way paying a bill at a bank counter is now.

The rule of thumb that falls out of this: multiply the current doubling time by the number of doublings between here and 5 per cent, and that's roughly how long you have. If the AEI reads 1 per cent and is doubling every ten months, take-off is a little over two years away. That's not a forecast to file. That's a countdown.

What to do with this

Three things, whether you run a business or just watch the economy.

First, measure. Two numbers for your own business. The demand-side one: what percentage of your revenue is agent-touched, discovered, shortlisted, or executed by an AI agent. Most analytics can't answer that yet, which is itself telling. And the supply-side one: are you agent-receptive? When an agent arrives at your website today, does it get a page built for human eyes, or does it get recognised and handed the structured information it came for: products, prices, terms, availability? The share of sellers who can answer yes is the supply side of this whole story, and almost nobody scores well on it yet. The route to yes now has a name, WebMCP, and it's the kind of thing a web team can trial in a quarter.

Second, climb the ladder before your customers do. The GEO work you're already being nagged about, being findable and recommendable by AI assistants, is tier 3: the entry fee, not the game. The game is tiers 4 and 5: supporting the emerging protocols, being a business an agent can transact with, not just read about, and eventually fielding an agent of your own to manage the selling side. Somebody in the organisation has to own that, which is why I've argued for a Chief Agency Officer (Dixon, 2026e). Each rung you build now is a channel that already exists when the curve bends. This is capability-building, and it takes the same second flywheel I described to a room of CEOs in Chicago (see The CEO's Guide to AI): people first, then the tools, then the projects.

Third, watch the two numbers that matter: the tier 4 plus tier 5 share, and the doubling time. While the first sits near zero, you have time. When it moves and the second shortens, you don't.

The AI revolution won't announce itself with a press release. It will show up as a ratio, quietly compounding: the share of the economy where the machines do the finding, then the buying, then the selling. Businesses that dismissed websites in 1998 didn't fail in 1998; they failed years later, when the customers had moved and their names never came up. The website era gave laggards a decade to catch up. This one won't.

If you have better data for any cell in that table, or a series that could track a tier over time, I'd like to see it. The estimates are there to be improved, and the index only becomes useful once someone is publishing it regularly.

References

Adobe, 2026a. Adobe: U.S. online holiday shopping season hits record $257.8 billion. [online] Adobe News, January. Available at: https://news.adobe.com/news/2026/01/adobe-holiday-shopping-season [Accessed 15 September 2026].

Adobe, 2026b. Adobe Digital Insights: Prime Day 2026. [online] Adobe Business Blog, 29 June. Available at: https://business.adobe.com/blog/2026-prime-day-insights [Accessed 15 September 2026].

Amazon, 2025. Amazon is testing "Buy for Me", an AI shopping agent that buys from other brands' sites. [online] About Amazon, 3 April. Available at: https://www.aboutamazon.com/news/retail/amazon-shopping-app-buy-for-me-brands [Accessed 15 September 2026].

Ant Group, 2026a. Alipay's AI Pay exceeds 120 million transactions in one week. [online] Business Wire, 13 February. Available at: https://www.businesswire.com/news/home/20260213770962/en/ [Accessed 15 September 2026].

Ant Group, 2026b. Alipay launches AI Wallet and Token Pay; AI Pay passes 300 million transactions. [online] Business Wire, 26 May. Available at: https://www.businesswire.com/news/home/20260526337824/en/ [Accessed 15 September 2026].

Anthropic, 2026. Project Deal. [online] Anthropic, 24 April. Available at: https://www.anthropic.com/features/project-deal [Accessed 15 September 2026].

CNBC, 2026. OpenAI revamps shopping experience in ChatGPT after Instant Checkout. [online] CNBC, 24 March. Available at: https://www.cnbc.com/2026/03/24/openai-revamps-shopping-experience-in-chatgpt-after-instant-checkout.html [Accessed 15 September 2026].

Cooley, 2026. Ninth Circuit rules on AI agent access to third-party websites under the CFAA. [online] Cooley LLP, 6 August. Available at: https://www.cooley.com/news/insight/2026/2026-08-06-ninth-circuit-rules-on-ai-agent-access-to-third-party-websites-under-cfaa [Accessed 15 September 2026].

Dixon, R.J., 2025b. Generative Engine Optimisation (GEO): Preparing Organisations for the AI-Driven Web. [online] LinkedIn. Available at: https://www.linkedin.com/pulse/generative-engine-optimisation-geo-preparing-ai-driven-rob-dixon-zsgme [Accessed 15 September 2026].

Dixon, R.J., 2026b. Purpose, Execution, Judgement: A Framework for Human-AI Collaboration and Organisational AI Capability in the Post-ChatGPT Era. SSRN Electronic Journal. Available at: https://dx.doi.org/10.2139/ssrn.6282200 [Accessed 15 September 2026].

Dixon, R.J., 2026e. AI Innovation Architecture: Nested Purpose, Closed-Loop Dynamics, and the Chief Agency Officer. SSRN Working Paper. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6546178 [Accessed 15 September 2026].

Dixon, R.J., 2026. Our Viking moment: why the winning strategy changed on 30 November 2022. [online] robdixon.ai. Available at: https://www.robdixon.ai/post/our-viking-moment-strategy-change-ai [Accessed 15 September 2026].

Dixon, R.J., 2026. The CEO's Guide to AI: what I told a room full of CEOs in Chicago. [online] robdixon.ai, 3 September. Available at: https://www.robdixon.ai/post/ceos-guide-to-ai-chicago-keynote [Accessed 15 September 2026].

Eurostat, 2025. E-commerce statistics. [online] Eurostat Statistics Explained. Available at: https://ec.europa.eu/eurostat/statistics-explained/index.php?title=E-commerce_statistics [Accessed 15 September 2026].

FRED, 2026. E-Commerce Retail Sales as a Percent of Total Sales (ECOMPCTSA). [online] Federal Reserve Bank of St. Louis. Available at: https://fred.stlouisfed.org/series/ECOMPCTSA [Accessed 15 September 2026].

Google, 2026. Agent Payments Protocol: "Human Not Present" mode and contribution to the FIDO Alliance. [online] Google Keyword blog, 28 April. Available at: https://blog.google/products-and-platforms/platforms/google-pay/agent-payments-protocol-fido-alliance/ [Accessed 15 September 2026].

IMF, 2019. The Global Informal Economy: Large but On the Decline. [online] IMF Blog, 30 October. Available at: https://www.imf.org/en/Blogs/Articles/2019/10/30/the-global-informal-economy-large-but-on-the-decline [Accessed 15 September 2026].

Linux Foundation, 2026. Linux Foundation announces operational launch of x402 Foundation to standardize internet-native payments for AI agents and applications. [online] Linux Foundation, 14 July. Available at: https://www.linuxfoundation.org/press/linux-foundation-announces-operational-launch-of-x402-foundation-to-standardize-internet-native-payments-for-ai-agents-and-applications [Accessed 15 September 2026].

Mastercard, 2026. Mastercard: powering ASEAN's AI ambitions in the future of payments. [online] Mastercard Newsroom (Asia Pacific). Available at: https://www.mastercard.com/news/ap/en/newsroom/press-releases/en/2026/mastercard-powering-asean-s-ai-ambitions-in-the-future-of-payments/ [Accessed 15 September 2026].

McKinsey, 2025. The automation curve in agentic commerce. [online] McKinsey QuantumBlack, October. Available at: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-automation-curve-in-agentic-commerce [Accessed 15 September 2026].

McKinsey, 2026. Europe's agentic commerce moment: decision influence is here, execution is coming. [online] McKinsey QuantumBlack, 2 March. Available at: https://www.mckinsey.com/capabilities/quantumblack/our-insights/europes-agentic-commerce-moment-decision-influence-is-here-execution-is-coming [Accessed 15 September 2026].

Morgan Stanley, 2025. Agentic commerce: market impact and outlook. [online] Morgan Stanley Insights, 8 December. Available at: https://www.morganstanley.com/insights/articles/agentic-commerce-market-impact-outlook [Accessed 15 September 2026].

National Bureau of Statistics of China, 2026. Total Retail Sales of Consumer Goods in December 2025. [online] NBS, 20 January. Available at: https://www.stats.gov.cn/english/PressRelease/202601/t20260120_1962354.html [Accessed 15 September 2026].

Netlify, 2026. Compete in the OpenAI WebMCP Challenge. [online] Netlify Blog, 25 August. Available at: https://www.netlify.com/blog/compete-openai-webmcp-challenge/ [Accessed 15 September 2026].

OpenAI, 2025. Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol. [online] OpenAI, 29 September. Available at: https://openai.com/index/buy-it-in-chatgpt/ [Accessed 15 September 2026].

OpenAI, 2026. Scaling AI for everyone. [online] OpenAI, February. Available at: https://openai.com/index/scaling-ai-for-everyone/ [Accessed 15 September 2026].

PayPal, 2025. PayPal and Perplexity launch Instant Buy. [online] PayPal Newsroom, 25 November. Available at: https://newsroom.paypal-corp.com/2025-11-PayPal-and-Perplexity-Launch-Instant-Buy [Accessed 15 September 2026].

Rogers, E.M., 2003. Diffusion of Innovations. 5th ed. New York: Free Press. Available at: https://www.simonandschuster.com/books/Diffusion-of-Innovations-5th-Edition/Everett-M-Rogers/9780743222099 [Accessed 15 September 2026].

Salesforce, 2026. 2025 holiday shopping data. [online] Salesforce News, January. Available at: https://www.salesforce.com/news/stories/2025-holiday-shopping-data/ [Accessed 15 September 2026].

Santander, 2026. Santander and Mastercard complete Europe's first live end-to-end payment executed by an AI agent. [online] Santander Press Room, 3 March. Available at: https://www.santander.com/en/press-room/press-releases/2026/03/santander-and-mastercard-complete-europes-first-live-end-to-end-payment-executed-by-an-ai-agent [Accessed 15 September 2026].

Securities Times (证券时报), 2026. JD.com releases A2P2 agent autonomous payment protocol. [online] stcn.com, 11 June. Available at: https://www.stcn.com/article/detail/3956646.html [Accessed 15 September 2026]. In Chinese.

SecurityBrief Asia, 2026. Alipay launches AI commerce platform for merchants. [online] SecurityBrief Asia, August. Available at: https://securitybrief.asia/story/alipay-launches-ai-commerce-platform-for-merchants [Accessed 15 September 2026].

Stripe, 2025. Stripe and OpenAI launch Instant Checkout. [online] Stripe Newsroom, 29 September. Available at: https://stripe.com/newsroom/news/stripe-openai-instant-checkout [Accessed 15 September 2026].

Stripe, 2026. Stripe helps Meta's Muse shop with Link. [online] Stripe Newsroom, 8 September. Available at: https://stripe.com/newsroom/news/stripe-helps-meta-muse-shop-with-link [Accessed 15 September 2026].

TechCrunch, 2026a. India has 100M weekly active ChatGPT users, Sam Altman says. [online] TechCrunch, 15 February. Available at: https://www.techcrunch.com/2026/02/15/india-has-100m-weekly-active-chatgpt-users-sam-altman-says/ [Accessed 15 September 2026].

TechInformed, 2026. Visa opens one integration for AI agent payments. [online] TechInformed, 9 April. Available at: https://techinformed.com/visa-opens-one-integration-for-ai-agent-payments/ [Accessed 15 September 2026].

TechNode, 2026. QuestMobile: China's AI-native apps reach 499 million monthly active users. [online] TechNode, 14 July. Available at: https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/ [Accessed 15 September 2026].

The Next Web, 2026. China's tech giants are replacing the search bar with AI agents that shop for you. [online] TNW, 15 May. Available at: https://thenextweb.com/news/chinas-tech-giants-are-replacing-the-search-bar-with-ai-agents-that-shop-for-you [Accessed 15 September 2026].

US Census Bureau, 2026. Quarterly Retail E-Commerce Sales, 2nd Quarter 2026. [online] US Census Bureau, 18 August. Available at: https://www.census.gov/retail/mrts/www/data/pdf/ec_current.pdf [Accessed 15 September 2026].

VentureBeat, 2026. Google Chrome ships WebMCP in early preview, turning every website into a structured tool for AI agents. [online] VentureBeat, 12 February. Available at: https://venturebeat.com/infrastructure/google-chrome-ships-webmcp-in-early-preview-turning-every-website-into-a [Accessed 15 September 2026].

Web Machine Learning Community Group, 2026. WebMCP. Draft Community Group Report, W3C, 14 September. Editors: B. Walderman (Microsoft), K. Sagar and D. Farolino (Google). Available at: https://webmachinelearning.github.io/webmcp/ [Accessed 15 September 2026].

White & Case, 2026. EU AI Omnibus enters into force, amending the AI Act. [online] White & Case LLP, July. Available at: https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act [Accessed 15 September 2026].

World Bank, 2021. Widespread informality likely to slow recovery from COVID-19 in developing economies. [online] World Bank, 11 May. Available at: https://www.worldbank.org/en/news/press-release/2021/05/11/widespread-informality-likely-to-slow-recovery-from-covid-19-in-developing-economies [Accessed 15 September 2026].

Worldpay, 2025. 10 years of cash, cards and crypto: Worldpay's Global Payments Report. [online] Business Wire, 11 March. Available at: https://www.businesswire.com/news/home/20250311959755/en/ [Accessed 15 September 2026].

The regional AEI estimates in the table are the author's first-pass calculations from the sources above, scoped to consumer commerce; they are offered as a baseline to be improved, not as measured fact. The five-tier model, the weights, and the phase thresholds are the author's.

Rob Dixon is Founder and CEO of Dixon AI Ltd and lectures at Manchester Metropolitan University. He writes and speaks on AI capability, strategy, and human-AI collaboration. Bibliography · Google Scholar · ORCID · LinkedIn · dixonai.com