A Brief History of Autonomous Commerce

By the Agentic Commerce Report editorial team — covering the AI-to-checkout layer since September 2024.

Autonomous commerce — the execution of buying, selling, and procurement decisions by software systems with little or no human intervention — is not a product of the 2020s AI boom. Its roots run back more than six decades, to the era of mainframes and paper-to-wire document conversion.

What has changed, dramatically and irreversibly, is the degree of autonomy: from rule-following scripts that required human configuration at every step, to AI agents capable of browsing suppliers, negotiating terms, placing orders, and reconciling invoices without a human ever touching the workflow.

This article traces the full history of autonomous commerce across five distinct eras, identifies the milestones that defined each period, and explains precisely what makes the 2024–2026 inflection point different from every prior wave of commercial automation.

If you are new to the underlying concept, start with our primer on what agentic commerce is before continuing.

What Is Autonomous Commerce? A Working Definition

For the purposes of this history, autonomous commerce describes any system in which software initiates, negotiates, approves, or completes a commercial transaction — including purchasing, selling, bidding, invoicing, or procurement — based on programmed rules, learned models, or AI reasoning, rather than direct human action at the moment of execution.

The spectrum runs from fully deterministic (a pre-coded reorder script that fires when inventory drops below a threshold) to fully generative (an LLM agent that reads a supplier's website, compares it with three alternatives, drafts a purchase order, and submits it). Most real-world deployments have always occupied the middle: partially automated, partially supervised, with humans setting the rules and reviewing exceptions.

Understanding that spectrum is essential to interpreting the history correctly. Each era did not replace the previous one — it added a new layer of capability on top of infrastructure that already existed.

Era 1: EDI and Electronic Procurement (1960s–1990s)

The history of autonomous commerce begins not in Silicon Valley but in the logistics offices of the US military and the rail freight industry.

Electronic Data Interchange (EDI) was developed in the late 1960s to replace paper-based purchase orders, invoices, and advance shipping notices with standardized electronic messages transmitted between computer systems. The Transportation Data Coordinating Committee formalized early EDI standards in 1968. By the 1970s, major retailers including Sears and K-Mart were piloting EDI with their largest suppliers to automate replenishment orders.

The critical commercial breakthrough came in 1975, when Walmart adopted EDI as a condition of doing business with its suppliers — effectively mandating that vendors connect their inventory systems directly to Walmart's purchasing platform. This was, in a meaningful sense, the first large-scale deployment of machine-to-machine commerce: software on one side of a trading relationship sending a legally binding purchase order to software on the other side, with no human drafting the document.

The ANSI X12 standard (published 1979) and the international EDIFACT standard (1987) created the lingua franca that allowed disparate systems to exchange transaction data reliably. By the late 1980s, EDI was processing hundreds of billions of dollars in B2B transactions annually across retail, automotive, healthcare, and government procurement.

What EDI automated: document creation, transmission, and receipt acknowledgment — but nothing else. A human buyer still set the reorder points, negotiated the contracts, and decided which suppliers to use. EDI was a pipe, not a brain.

Key constraint: EDI required both trading partners to implement the same standard, making onboarding expensive and slow. Only large enterprises and their tier-1 suppliers could participate meaningfully. Smaller suppliers continued to receive faxed purchase orders and re-key them manually.

Era 2: eCommerce and Algorithmic Purchasing (1994–2010)

The public internet transformed commerce faster than almost any prior technology. Amazon launched in July 1994; eBay followed in September 1995. But the impact on autonomous commerce was subtler than headlines suggested.

The most significant development for automated buying in this era was not the consumer storefronts — it was the programmatic price comparison and algorithmic repricing systems that emerged alongside them.

By the early 2000s, Amazon and its marketplace sellers were deploying automated repricing bots that adjusted prices in real time based on competitor pricing data. A seller's software would poll competitor listings every few minutes and adjust its own price to undercut or match — a closed-loop feedback system operating entirely without human intervention. The phenomenon became visible enough that Amazon's seller policy had to be updated repeatedly to manage bot behavior on its own platform.

In B2B procurement, the period saw the rise of e-procurement platforms — Ariba (founded 1996, acquired by SAP in 2012), Coupa (2006), and Oracle Procurement Cloud — that extended EDI's logic into a web-native interface. These systems allowed buyers to configure sourcing rules, approval workflows, and preferred supplier catalogues so that routine purchases under a spending threshold could complete without a purchase manager's signature.

The Dutch auction model pioneered by FreeMarkets (acquired by Ariba in 2004) took this further: software managed reverse auctions in which pre-qualified suppliers competed in real time to win a contract, with the platform automating bid validation, ranking, and award notification. Hundreds of millions of dollars in supply contracts were allocated by automated auction logic, with human buyers reviewing only the final result.

What this era automated: pricing decisions, catalog purchasing below approval thresholds, supplier auctions, and spend analytics. The human buyer's role shifted from transactional to strategic — setting the rules under which software operated.

Key constraint: every automation was explicitly rule-based. If a situation arose that the rules did not cover — a new supplier, an unusual product specification, a currency fluctuation outside the programmed range — the system stopped and escalated to a human.

Era 3: Programmatic Advertising and Bot-Driven Buying (2010–2018)

The decade of programmatic advertising demonstrated, at scale, that software could execute billions of commercial transactions per day with no human involvement in any individual decision. This era is sometimes overlooked in histories of autonomous commerce because "advertising" feels different from "procurement" — but the underlying dynamic was identical: software agents buying and selling access to audiences through automated auctions, in milliseconds.

The Google Display Network launched its real-time bidding (RTB) infrastructure in 2009. DoubleClick Bid Manager (now DV360) extended this to third-party exchanges in 2011. By 2017, the Interactive Advertising Bureau estimated that more than 80% of US digital display advertising was transacted programmatically, representing over $26 billion in annual spend allocated by automated bidding algorithms.

RTB is worth dwelling on as an engineering feat: when a user loads a webpage, an auction opens, SSPs broadcast the impression opportunity to hundreds of DSPs, each DSP evaluates the bid request against its campaign criteria and responds with a bid price — all within 100 milliseconds, before the page finishes loading. The "buyer" (a brand's DSP algorithm) evaluates audience signals, frequency caps, placement quality, and campaign pacing to decide whether to bid and how much. No human approves any individual impression purchase.

The dark side of bot-driven commerce also became visible in this period. Invalid traffic — fraudulent bot impressions generating fake ad spend — cost advertisers an estimated $6.5 billion in 2017 (ANA/White Ops study). The automation that made programmatic efficient also made fraud at industrial scale possible, because there was no human auditing individual transactions.

In parallel, algorithmic trading in financial markets (a form of autonomous commerce in securities) matured rapidly. High-frequency trading firms accounted for more than 50% of US equity volume by 2012. The 2010 Flash Crash — in which the Dow Jones Industrial Average dropped nearly 1,000 points in minutes before recovering — demonstrated both the power and the systemic risk of fully autonomous commercial decision-making at speed.

For a detailed look at how machine-to-machine transactions operate at the infrastructure level, see our article on machine-to-machine commerce.

What this era automated: bid evaluation, auction participation, budget pacing, audience targeting, and creative selection — all in real time, at internet scale. The feedback loop (bid → impression served → conversion tracked → bid model updated) operated continuously without human input.

Key constraint: programmatic systems were still rule-and-model driven, not reasoning-capable. They optimized within defined objective functions (maximize clicks, minimize CPA) but could not interpret context, handle novel situations, or pursue goals that required multi-step planning.

Key Milestones in the History of Autonomous Commerce

Era 4: RPA and Process Automation (2015–2022)

Robotic Process Automation (RPA) brought a different kind of autonomous commerce capability: software robots that could navigate existing user interfaces — websites, desktop applications, legacy ERPs — to execute tasks that previously required human hands on keyboards.

UiPath (founded 2005 in Bucharest, IPO 2021 at $29 billion valuation), Automation Anywhere (founded 2003), and Blue Prism (founded 2001) built the platforms that enabled enterprises to deploy "bots" that logged into supplier portals, extracted invoice data, matched it to purchase orders in SAP, flagged discrepancies, and posted approved invoices to accounts payable — all without human action on any routine transaction.

By 2020, Deloitte's Global RPA Survey found that 78% of enterprises that had adopted RPA were scaling or planning to scale their deployments. In procurement specifically, RPA was applied to:

• Purchase order processing: Bots extracted requisitions from email or ERP, generated POs, emailed them to suppliers, and logged confirmations.
• Invoice matching: Three-way matching (PO vs. goods receipt vs. invoice) automated for invoices meeting tolerance thresholds.
• Supplier onboarding: Form completion, data validation, and system entry across multiple platforms.
• Contract management: Extracting key terms from PDF contracts and populating master data fields.

The COVID-19 pandemic accelerated RPA adoption sharply. When office workers could not access physical mail or on-premise systems, bots that could navigate digital workflows became operationally critical. The RPA market grew from approximately $1.4 billion in 2019 to $2.9 billion in 2021 (Grand View Research).

What this era automated: the execution layer — form-filling, data entry, document matching, portal navigation. RPA bots are essentially very fast, very consistent digital workers following explicit step-by-step scripts.

Key constraint: RPA bots are brittle. A minor change to a supplier portal's UI — a button relocated, a field renamed — breaks the bot. Maintenance overhead scales with the number of bots deployed. And RPA bots cannot handle unstructured inputs: a supplier email written in an unusual format, a PDF with a non-standard layout, an exception that the script author did not anticipate.

Era 5: LLM-Powered Agentic Commerce (2023–Present)

Every prior era of autonomous commerce shared a fundamental constraint: systems could only act within explicitly defined rules or within the optimization boundary of a trained model. They could not reason about novel situations, interpret unstructured information, or plan multi-step sequences of actions toward an open-ended goal.

Large language models changed this — not incrementally, but categorically.

The release of GPT-4 in March 2023, followed by Claude 2, Gemini, and a proliferation of open-weight models, gave developers access to AI systems capable of understanding natural language instructions, reading and summarizing documents, generating structured outputs, and — crucially — calling external tools. When an LLM is connected to tools (web browsers, APIs, database queries, form-submission capabilities), it becomes an agent: a system that can perceive its environment, plan a course of action, execute steps, observe results, and adjust — in a loop, without human intervention at each step.

Applied to commerce, this changes the fundamental unit of automation. Where RPA automated a fixed sequence of steps that a human had pre-specified, an LLM agent can be given a goal — "find the three best-rated suppliers of industrial fasteners under $0.08/unit who can deliver within two weeks, compare their current pricing, and prepare a draft purchase order for the lowest-qualified bidder" — and pursue it autonomously, adapting to whatever it encounters along the way.

Specific capabilities that distinguish this era:

• Unstructured document comprehension: An agent can read a supplier's PDF price sheet, extract line items, and compare them to incumbent pricing — without any template or parser being pre-built for that document's format.
• Web navigation: Via browser-use tools and computer-use APIs (Anthropic's computer-use capability, released October 2024), agents can interact with any website as a human would, without requiring a supplier to expose an API.
• Multi-step planning: Agents can decompose complex procurement workflows into sub-tasks, execute them in sequence or parallel, and handle exceptions by reasoning through them rather than escalating.
• Natural language negotiation: Early deployments in 2024–2025 demonstrated agents negotiating terms in email threads, adjusting positions based on supplier responses, and escalating only when pre-defined parameters were exceeded.
• Agent-to-agent coordination: In multi-agent architectures (LangGraph, AutoGen, CrewAI), a buying agent and a selling agent can interact directly, negotiating and transacting without either having a human counterpart in the loop.

For concrete examples of how these capabilities are deployed today, see our roundup of agentic commerce examples.

What Makes 2024–2026 Categorically Different

It is tempting to describe LLM-powered agentic commerce as "just another wave of automation." This framing misses the structural break.

Prior automation automated tasks. Agentic commerce automates judgment.

Every prior era — EDI, e-procurement, programmatic advertising, RPA — required a human to have already made the strategic decisions and encoded them as rules. The software executed those rules faithfully and quickly. When a situation arose outside the rules, a human decided what to do.

LLM agents apply judgment to novel situations. They can evaluate a supplier they have never encountered, read a contract in a format they have never seen, adapt to a procurement policy described in natural language rather than code, and respond to an unexpected price increase with a reasoned counter-proposal. This is not a difference of degree — it is a difference of kind.

The second structural difference is composability. RPA bots are built for specific workflows; connecting them requires custom integration work. LLM agents communicate in natural language, which means an agent designed for supplier research can hand off to an agent designed for contract review, which can hand off to an agent designed for PO generation — with the handoff requiring no more than a structured message. This composability is what makes multi-agent procurement workflows viable at scale for the first time.

The third difference is the speed of capability expansion. EDI took roughly 20 years to reach mainstream adoption. Programmatic advertising took about 8 years to dominate digital ad spend. RPA took roughly 6 years to reach broad enterprise deployment. Agentic commerce frameworks went from research prototype to production deployment in under 24 months, driven by the compounding pace of foundation model improvement and the explosion of tooling built on top of those models.

What Comes Next

The trajectory from EDI to AI agents reveals a consistent direction: each era reduced the number of decisions that required a human, and expanded the complexity of decisions that software could handle. There is no reason to expect this trend to reverse.

Near-term (2026–2028): The primary development will be standardization. Just as EDI required ANSI X12 and EDIFACT to reach scale, agentic commerce will require agreed protocols for agent identity, authorization, negotiation, and audit logging. Early frameworks — including Anthropic's Model Context Protocol (MCP), emerging agent identity standards, and platform-specific agent APIs from major enterprise vendors — are competing to become the infrastructure layer for agent-to-agent commercial interaction.

Medium-term (2028–2032): Expect the emergence of agent networks as market participants. Rather than a human company with an AI agent, commerce will increasingly involve AI-native organizations where the decision-making infrastructure is itself the company. Procurement agents for one firm will interact directly with sales agents for another, with human oversight reserved for policy-setting and exception review above defined materiality thresholds.

Long-term: The boundary between automation and agency will continue to blur. The more interesting question may not be "how autonomous can commerce become?" but "what governance structures, liability frameworks, and verification mechanisms allow humans to remain appropriately in control of increasingly autonomous commercial systems?" That is a legal, ethical, and institutional question as much as a technical one — and the answer will shape the trajectory of autonomous commerce more than any individual model release.

Frequently Asked Questions

When did autonomous commerce begin?

The history of autonomous commerce traces to the late 1960s, when Electronic Data Interchange (EDI) standards enabled computers at separate companies to exchange purchase orders and invoices directly, without human transcription. Walmart's 1975 mandate requiring suppliers to use EDI is often cited as the first large-scale deployment of machine-to-machine commerce.

How is agentic commerce different from RPA?

Robotic Process Automation (RPA) executes fixed, pre-scripted sequences of steps that a human has explicitly defined in advance. Agentic commerce, powered by large language models, can interpret goals described in natural language, reason about novel situations the script author never anticipated, and adapt its approach based on what it encounters — without requiring a human to re-script the process. RPA automates tasks; agentic commerce automates judgment within those tasks.

Did programmatic advertising count as autonomous commerce?

Yes. Real-time bidding (RTB) systems executing billions of impression purchases per day, with no human involved in any individual transaction, are a clear form of autonomous commerce. Programmatic advertising demonstrated at unprecedented scale that software could handle commercial transactions faster and at higher volume than any human team — and also demonstrated the fraud and systemic risk that accompany fully autonomous buying systems.

What are the biggest risks of autonomous commerce?

The risks vary by era and mechanism. Rule-based automation creates brittleness (a small change breaks the system). Algorithmic systems create optimization pathologies (a bot optimizing for click-through rate may generate brand-unsafe placements). Agentic systems introduce new risks: agent error (the AI makes a wrong judgment), agent manipulation (a bad actor crafts inputs to mislead the agent), scope creep (an agent takes actions beyond its intended authority), and accountability gaps (when an autonomous agent makes a costly mistake, assigning responsibility is legally and operationally complex).

Which industries are furthest ahead in autonomous commerce adoption?

Financial services (algorithmic trading, automated lending decisions) and digital advertising (programmatic buying) have operated at near-full autonomy for transactional decisions since the mid-2010s. In B2B procurement, large retail and manufacturing companies with mature EDI and ERP infrastructure are the fastest adopters of agentic workflows. Healthcare and government procurement lag due to regulatory requirements for human approval on spending decisions.

Will autonomous commerce eliminate procurement jobs?

Historical patterns suggest substitution of specific tasks rather than wholesale elimination of roles, at least in the near term. EDI eliminated data-entry roles but created demand for EDI administrators and supply chain analysts. Programmatic advertising eliminated media buyers focused on manual insertion orders but created demand for programmatic traders and data analysts. The current LLM-powered wave is automating routine sourcing, vendor comparison, and PO processing — shifting procurement professionals toward supplier relationship management, risk oversight, and strategic category management. The pace of this wave is faster than prior transitions, which will compress the adjustment period.

The history of autonomous commerce is, at its core, a history of humans delegating progressively more complex commercial decisions to software — and the technology finally catching up to the full scope of what humans do when they buy and sell. We are at the earliest stage of that catching-up, not the end of it.

Further reading: What Is Agentic Commerce? — the foundational guide to how AI agents initiate, negotiate, and complete commercial transactions. | Machine-to-Machine Commerce — how AI agents transact with each other at the infrastructure level. | Agentic Commerce Examples — concrete deployments of agentic purchasing in production today.