# McKinsey's Agentic Commerce Framework Explained: The Six-Level Automation Curve

*McKinsey projects $3-5 trillion in global agentic commerce by 2030. Their Automation Curve framework lays out exactly how we get there -- not through a single leap, but through six distinct levels of delegation between humans and AI agents. Here is what strategy leaders need to know.*

**Last updated: March 2026**

---

## The Framework That Is Shaping Boardroom Conversations

When McKinsey published "The Agentic Commerce Opportunity" in October 2025, followed by "The Automation Curve in Agentic Commerce" in January 2026, it did more than forecast a market. It gave executives a shared vocabulary for one of the most consequential shifts in retail history.

The numbers alone command attention. McKinsey projects that AI agents will orchestrate $3 to $5 trillion in global retail spending by 2030, with up to $1 trillion from the United States alone (McKinsey, October 2025). Morgan Stanley independently estimates $190-385 billion in US e-commerce agentic spending by 2030, while Bain projects $300-500 billion at 15-25% of online retail. The convergence of these forecasts from top-tier firms signals that agentic commerce is not speculative -- it is an emerging structural reality.

But the more important contribution is not the dollar figure. It is the framework McKinsey built to explain how delegation between humans and AI agents actually evolves. They call it the Automation Curve, and it is already becoming the standard lens through which enterprises evaluate their agentic commerce readiness.

---

## The Six-Level Automation Curve

McKinsey's Automation Curve defines six levels of agent autonomy, from Level 0 (no AI involvement) through Level 5 (fully autonomous multi-agent networks). Critically, McKinsey describes this as "a curve, not a ladder" -- higher levels are not inherently better. The goal, as their researchers put it, is "optimal delegation, not maximum autonomy."

### Level 0: Programmatic Convenience ("Set and Forget")

This is the pre-agent baseline. Rule-based automation handles recurring purchases through subscription models. Amazon's Subscribe and Save is the canonical example, with approximately 23% of US Amazon users maintaining active subscriptions as of 2024.

Level 0 is useful but brittle. The automation is blind to context. When a consumer's needs change -- a dietary shift, a new household member, a budget constraint -- the rules break and the human steps back in to reconfigure manually. There is no intelligence, only repetition.

**Where most companies are today**: Any business with subscription or auto-replenishment capabilities operates at Level 0.

### Level 1: Assist ("Cognitive Sidekick")

At Level 1, AI agents support the shopping process through analysis and recommendation, but without executing any transactions. The agent might find gifts under a specific budget, compare product specifications across brands, or summarize reviews. The human retains full control over every decision and purchase.

This is where the majority of current AI shopping tools operate. ChatGPT handling 53 million daily shopping queries (Digital Commerce 360), Perplexity's "Buy with Pro" tool, and Google's AI Overviews reaching 1.5 billion users monthly all function primarily at Level 1. The agent informs. The human decides and acts.

**Key limitation**: No cart assembly, no transaction preparation. The agent is a research assistant, nothing more.

### Level 2: Assemble ("Personal Shopper")

McKinsey identifies Level 2 as a "qualitative turning point." Here, the agent transitions from analysis to orchestration. It resolves trade-offs, accounts for taxes, shipping costs, and loyalty programs, and returns a checkout-ready basket for the consumer to review.

The consumer's role fundamentally shifts. Instead of comparing options and building a cart, the consumer is now approving a proposed solution. This is a meaningful psychological and operational change -- the locus of effort moves from the human to the agent.

**Why it matters**: Level 2 is where the consumer begins to experience genuine time savings, and where merchants must start designing for agent-readable product data rather than human-browsable interfaces.

### Level 3: Authorize ("Supervised Executor")

At Level 3, consumers delegate rules-based purchasing authority to agents. The agent executes end-to-end workflows when predefined conditions are met. McKinsey offers concrete examples: "If groceries are under $120 and arrive Friday 6-8 p.m., place the order," or "If my preferred sneakers drop below $80 from merchants I trust, buy them."

The agent acts autonomously within guardrails. It escalates to the human only when conditions fall outside the defined parameters. This is supervised autonomy -- the consumer sets the boundaries, and the agent operates freely within them.

**Strategic implication**: Level 3 requires robust payment infrastructure, agent authentication protocols, and clear accountability frameworks. This is where trust architecture becomes a competitive differentiator.

### Level 4: Autonomize ("Intent Steward")

Level 4 represents a shift from transactional to strategic delegation. Agents no longer respond to individual purchase triggers. Instead, they operate toward long-term consumer objectives -- optimizing total spending across categories, managing loyalty program value, anticipating needs before they arise.

The agent becomes proactive rather than reactive. It might notice that a consumer's coffee consumption pattern has shifted and adjust orders accordingly, or identify that switching brands on a household staple would save $240 annually without sacrificing quality.

**Key distinction from Level 3**: The agent is no longer executing discrete rules. It is interpreting intent across time and categories, making judgment calls within a broader mandate.

### Level 5: Network Autonomy ("Multi-Agent Commerce")

McKinsey describes Level 5 as "still nascent." In this model, commerce defaults to agent-to-agent negotiations. A consumer's personal agent communicates with specialized networks of merchant agents, logistics agents, and payment agents. Pricing is optimized, delivery is coordinated, loyalty is maximized -- all through machine-to-machine negotiation.

Multi-agent marketplaces emerge where intent is brokered, trust is carried through reputation signals, and transactions are settled through shared protocols. McKinsey envisions this as "procurement as a service" running continuously in the background.

**Reality check**: No commercial system operates at Level 5 today. But the infrastructure is being built. Google's Agent Payments Protocol (AP2) with 60+ partner organizations, OpenAI and Stripe's Agentic Commerce Protocol (ACP), and the Model Context Protocol (MCP) are all laying groundwork for this future.

---

## "Optimal Delegation, Not Maximum Autonomy" -- The Key Insight

The most important sentence in McKinsey's entire framework may be this: the goal is "optimal delegation, not maximum autonomy." This is a direct counter to the Silicon Valley instinct to automate everything to the highest possible degree.

McKinsey's research shows that delegation acceptance is most sustainable when it is bounded (clear limits), episodic (not continuous), and easy to reverse. It becomes contested when it shifts toward continuous execution with implicit authority.

This has profound implications for product strategy. Companies that race to build Level 5 systems for categories where consumers are comfortable only at Level 2 will face rejection. The framework is not a maturity model where everyone should aspire to the top. It is a matching exercise between category characteristics, consumer trust, and agent capability.

---

## Three Interaction Models

McKinsey identifies three architectural patterns for how agents interact with commerce infrastructure:

**Agent-to-Site**: The agent interacts directly with merchant platforms, scanning websites, comparing options, and surfacing results. A travel agent checking multiple hotel sites against user preferences is the simplest example. This model works within existing e-commerce infrastructure.

**Agent-to-Agent**: Autonomous agents transact directly with other agents. A personal shopping agent negotiates bundle discounts with a retailer's AI commerce agent. This model requires new protocols and trust frameworks but enables far richer negotiation than agent-to-site interactions.

**Brokered Agent-to-Site**: Intermediary systems facilitate multi-agent and multi-platform interactions. A restaurant-booking agent contacts OpenTable to find availability and apply loyalty discounts. The broker adds a coordination layer that neither the consumer's agent nor the merchant could efficiently maintain alone.

Each model has different infrastructure requirements, different trust profiles, and different competitive dynamics. Most current implementations are Agent-to-Site. The protocols being developed by Google (AP2), OpenAI/Stripe (ACP), and the broader MCP ecosystem are building toward Agent-to-Agent and Brokered models.

---

## Category Delegation Patterns

McKinsey's framework becomes especially actionable when applied to specific product categories. Willingness to delegate varies sharply along four dimensions: ticket size, emotional salience, identity signaling, and regret risk.

**Task-Oriented Categories (High Delegation)**: Household goods, consumables, and groceries are natural candidates for the highest autonomy levels. Brand narrative matters less than accuracy and reliability. Delegation climbs rapidly once operational reliability is verified. These categories may reach Level 3-4 delegation within the forecast period.

**Complex Purchases (Selective Delegation)**: Travel, consumer electronics, and home goods see "selective and situational" delegation. Agents autonomously handle research, comparison, and monitoring (Levels 1-2), but humans escalate when meaningful trade-offs are involved. McKinsey notes that "specification-driven purchases" -- products compared primarily on measurable attributes -- show the earliest vulnerability to agent-mediated displacement.

**Identity-Oriented Categories (Low Delegation)**: Luxury goods, fashion, and milestone purchases see delegation stall at Levels 1-2. The discovery process and personal expression are part of the product's value. Human involvement is not a friction to be eliminated -- it is integral to what makes the purchase meaningful.

This category framework explains why a single company may need to operate at multiple levels simultaneously across its product lines.

---

## The 4,700% Traffic Growth and What It Means

Among the most cited statistics in McKinsey's research: shopping-related GenAI queries grew 4,700% year-over-year through July 2025 (corroborated by Adobe and BCG data). AI traffic to US retail sites on Black Friday 2025 surged 805% year-over-year (Adobe).

These numbers need context. ChatGPT's share of total e-commerce sessions remains below 0.2% (Kaiser and Schulze). And ChatGPT referrals convert 86% worse than traditional affiliate links by some measures, even as Amazon's Rufus AI converts 60% higher than non-AI sessions within Amazon's own ecosystem.

The paradox is instructive. External AI agents (ChatGPT, Perplexity) are driving massive traffic growth from a tiny base with inconsistent conversion. Embedded AI agents (Amazon Rufus, Salesforce-powered retailer agents) show strong conversion within controlled environments. The growth trajectory is undeniable -- ChatGPT now drives over 20% of referral traffic to Walmart -- but the conversion economics are still being worked out.

For strategists, the signal is clear: AI-referred traffic is growing at rates that will make it material within 12-18 months, regardless of current conversion gaps. Companies that wait for the conversion economics to prove out before optimizing for agent discoverability will find themselves invisible to a fast-growing channel.

---

## Four Technical Protocols McKinsey Identifies

McKinsey's framework identifies four emerging protocols that will form the infrastructure layer of agentic commerce:

**MCP (Model Context Protocol)**: Enables context sharing between AI models, allowing agents to maintain state and preferences across interactions and platforms.

**A2A (Agent-to-Agent Protocol)**: A communication standard for inter-agent interaction, enabling the Agent-to-Agent model described above.

**AP2 (Agent Payments Protocol)**: Led by Google with 60+ partner organizations, AP2 establishes payment authorization frameworks for agent-mediated transactions.

**ACP (Agentic Commerce Protocol)**: Co-developed by OpenAI and Stripe, ACP enables in-ChatGPT purchases and is already live with over 1 million Shopify merchants opted into the OpenAI commerce ecosystem.

These protocols are not competing standards in the traditional sense. They address different layers of the stack -- context, communication, payment, and commerce -- and will likely coexist and interoperate. Companies should monitor all four rather than betting on a single standard.

---

## Risk Categories and Mitigation

McKinsey identifies seven distinct risk categories that enterprises must address as they move up the Automation Curve:

**Trust and Transparency**: As agents become primary transaction participants, trust frameworks must evolve beyond human-facing UX to include agent-verifiable signals. Eighty-three percent of consumers express data privacy concerns even as 52% willingly share data with AI agents (IBM-NRF, January 2026).

**Identity Management**: Secure authorization protocols for agent-mediated purchases require new approaches to authentication -- an agent acting on behalf of a consumer is neither the consumer nor a traditional bot.

**Fraud Prevention**: Seventy-eight percent of financial institutions expect AI agent fraud spikes (Salesforce). Banks must develop protocols for algorithmic purchasing authorization that can distinguish legitimate agent activity from malicious automation.

**Data Privacy**: The privacy paradox -- consumers want personalized agent experiences but fear data exposure -- requires transparent data governance that goes beyond current cookie-consent frameworks.

**Consumer Behavior**: Reduced friction may lead to higher spending. The consumer protection implications of frictionless AI purchasing are not yet addressed by existing regulatory frameworks.

**Accountability**: When an autonomous agent makes a poor purchasing decision, the chain of accountability (consumer, agent provider, merchant, protocol layer) is legally untested.

**Regulatory**: Competition, privacy, data protection, and consumer safeguard frameworks all require updates. The EU, with 84% of surveyed consumers already using AI tools daily, is likely to lead regulatory development.

---

## Strategic Recommendations for Different Company Types

McKinsey's framework implies different strategic priorities depending on where a company sits in the value chain.

**For large retailers and marketplaces**: Prioritize data quality and machine readability immediately. Clean product catalogs, real-time inventory APIs, and structured pricing data are table stakes. Invest in agent authentication infrastructure and auditable transaction logs. The 97% of retailers planning to increase AI spending next fiscal year (NVIDIA) suggests competitive intensity will be high.

**For brands and manufacturers**: Focus on the four critical infrastructure capabilities McKinsey identifies: identity resolution across surfaces, real-time data synchronization, intent capture at the first mile, and cross-surface normalization. Brands that cannot be accurately represented in agent-mediated contexts will lose share to those that can.

**For mid-market merchants**: The asymmetry between Level 0-1 (accessible with current tools) and Level 3+ (requiring significant infrastructure investment) creates a strategic window. Merchants that invest early in structured product data and protocol compatibility will disproportionately benefit as agent traffic scales.

**For technology providers**: The protocol layer (MCP, A2A, AP2, ACP) is where platform leverage will be built. Companies that establish themselves as infrastructure for agent-mediated commerce -- analogous to payment processors in traditional e-commerce -- will capture outsized value.

McKinsey notes that AI systems "ride the rails" of existing commerce infrastructure rather than requiring entirely new systems, enabling a faster transition than the web and mobile revolutions. This means the window for strategic positioning is shorter than many executives assume.

---

## Where Your Company Sits on the Curve: A Self-Assessment

Use these questions to locate your organization on McKinsey's Automation Curve:

**Level 0 indicators**: You offer subscription or auto-replenishment but no AI-driven personalization. Product data is optimized for human browsing, not machine readability.

**Level 1 indicators**: You have deployed chatbots or AI search that help customers find products, but no agent can assemble a cart or execute a transaction on your platform.

**Level 2 indicators**: Your systems can respond to agent queries with structured product data, pricing, and availability. An external agent could theoretically assemble a checkout-ready basket from your catalog.

**Level 3 indicators**: You support authenticated agent transactions with rules-based purchasing. Your payment infrastructure can process agent-initiated orders with appropriate authorization.

**Level 4 indicators**: Your systems support ongoing agent relationships -- preference learning, proactive recommendations, and optimization across multiple interactions over time.

**Level 5 indicators**: Your infrastructure participates in multi-agent negotiation protocols. Your commerce APIs support machine-to-machine price discovery, bundle negotiation, and automated settlement.

Most enterprises today cluster between Levels 0 and 1. The strategic question is not "How do we reach Level 5?" but rather "What is the optimal level for each of our product categories and customer segments, and what infrastructure do we need to support it?"

---

## Frequently Asked Questions

**What exactly is McKinsey's Automation Curve?**
The Automation Curve is a six-level framework (Level 0 through Level 5) published by McKinsey in January 2026 as a companion to their October 2025 agentic commerce report. It maps the progression of AI agent autonomy in commerce from rule-based subscriptions (Level 0) to fully autonomous multi-agent networks (Level 5). McKinsey emphasizes it is "a curve, not a ladder" -- the optimal level varies by product category and consumer context.

**How big is the agentic commerce opportunity according to McKinsey?**
McKinsey projects $3 to $5 trillion in global retail spending orchestrated by AI agents by 2030, with up to $1 trillion from the United States alone. These figures represent spending that agents influence or execute, not the value of the agents themselves. Independent forecasts from Morgan Stanley ($190-385 billion US) and Bain ($300-500 billion US) corroborate the directional scale.

**Where are most companies on the Automation Curve today?**
The vast majority of enterprises operate between Level 0 (rule-based subscriptions) and Level 1 (AI-assisted research and recommendations). Some leading retailers are beginning to enable Level 2 capabilities (agent-assembled carts), but Level 3 and above remain largely experimental. McKinsey describes Level 5 as "still nascent" with no commercial implementations.

**Why does McKinsey say "optimal delegation, not maximum autonomy"?**
Because consumer willingness to delegate varies dramatically by category. Groceries and household consumables are natural candidates for high autonomy (Levels 3-4), while luxury goods and fashion purchases see delegation stall at Levels 1-2 because personal discovery and expression are part of the product's value. Pushing automation beyond what consumers accept in a given category creates rejection, not efficiency.

**What are the four protocols McKinsey identifies for agentic commerce infrastructure?**
MCP (Model Context Protocol) for context sharing between AI models, A2A (Agent-to-Agent Protocol) for inter-agent communication, AP2 (Agent Payments Protocol) led by Google with 60+ partners for payment authorization, and ACP (Agentic Commerce Protocol) co-developed by OpenAI and Stripe for in-platform purchases. These protocols address different layers of the stack and are expected to coexist.

**How should companies prioritize their agentic commerce investments?**
McKinsey recommends three immediate actions: prioritize data quality (clean catalogs, real-time inventory), invest in machine readability (APIs for pricing, inventory, and promotions), and support agent authentication (budget constraints and auditable logs). The first step for any company is ensuring its product data can be accurately consumed by AI agents -- without this foundation, no higher-level capability is possible.

**What is the 4,700% traffic growth figure and should I take it seriously?**
Shopping-related GenAI queries grew 4,700% year-over-year through July 2025, according to data corroborated by Adobe and BCG. The growth rate is real but the base is small -- ChatGPT represents less than 0.2% of total e-commerce sessions. However, it already drives over 20% of referral traffic to Walmart. The trajectory suggests AI-referred traffic will become material within 12-18 months, making early optimization a strategic advantage.

---

*This analysis is based on McKinsey's "The Agentic Commerce Opportunity" (October 2025) and "The Automation Curve in Agentic Commerce" (January 2026), supplemented by data from Morgan Stanley, Bain, BCG, IBM-NRF, Adobe, Salesforce, Gartner, Forrester, and industry sources. All statistics are attributed to their original sources where available.*