Agents already shape commercial evaluation. Settlement is still human-gated.

Fully autonomous checkout is still a thin slice of commerce. Buying agents matter most in upstream discovery, pricing math, and vendor shortlists, where seller catalogs either answer cleanly or get scraped.

See the product → About PairRail

Market layers

Where agents show up today

PairRail’s working model for sellers and investors: autonomous settlement remains under about 3% of fully settled volume, while agents directly drive or influence roughly 20-30% of upstream discovery, pricing calculations, and commercial evaluations. That gap is the window Atlas is built for.

Layer Penetration What it means

Fully autonomous settlement < 2-3% Zero-click close is still rare. Most of it is narrow API top-ups, programmatic spend, and scheduled infrastructure. Agents evaluate far more than they settle alone.

AI-referred / mediated checkout Growing fast The agent finds and configures; a human still signs. US retail ecommerce via general AI platforms is forecast at about $20.9B in 2026 (nearly 4× 2025), still a small share of total ecommerce.¹

Upstream discovery & catalog queries ~20-30% Pricing pages, docs, and API catalogs already see heavy machine traffic from coding and procurement agents. That is where scraper fatigue becomes a seller problem, and where protocol rails win.

Enterprise pilots & production agents Widespread pilots Most large organizations are piloting or deploying agents somewhere. Scaled multi-agent production is still uneven.²⁶ Buyers are settling on shared agency: humans set budget and policy; agents evaluate and quote.

What changed in the last few quarters

Protocol consolidation. Screen scraping is giving way to adapters: MCP across developer tooling, Google’s UCP for commerce journeys, and ACP for agent checkout with payments partners.³
Higher-intent mediated traffic. AI-driven referral to retail sites surged in recent holiday windows (Adobe reported about 670% AI traffic growth to US retail sites on Cyber Monday 2025).⁴ Sellers feel load and rate-limit pressure from bots harvesting price tables.
Shared agency, not blind auto-buy. The idea that agents would buy enterprise software without fences has faded. Mandates, allowances, and human deal desks are the durable pattern, which matches how Atlas fails closed outside policy.

Where it goes next, and how Atlas maps

B2B buying becomes agent-intermediated. Gartner expects AI agents to intermediate most B2B purchasing by 2028, moving over $15 trillion through agent exchanges.⁵ Sellers that stay human-HTML-only lose the shortlist. Atlas: versioned catalogs and protocol rails agents can query without scraping.
Delegated micro-budgets. Buyers assign monthly allowances to procurement agents for seats and token top-ups. Atlas: seller-approved machine keys and mandate-aware quotes.
Fail-closed seller hardening. As agent traffic rises, prompt injection and pricing arbitrage follow. Atlas: deterministic pricing lab, deal-desk escalation, and non-binding answers until execute.
Identity and signing for machine buyers. Distinguishing a real enterprise credit line from an unverified crawler becomes table stakes. Atlas: scoped credentials, access inbox, and an audit trail on every commercial action.
Settlement stays on your rails. Even as AI-assisted spend grows, PairRail is not merchant of record. Atlas: quote execute hands off to your billing stack; you keep MoR, tax, and fulfillment.

Who is doing this now

Machine customers are not only a forecast. Stripe is shipping agent payment and catalog paths so services can sell to agents (for example Browserbase metering sessions). Salesforce’s Agentforce B2B Buyer Agent lets account-aware ordering happen over WhatsApp and SMS; Siemens is cited as running Agentforce B2B Commerce across many countries. API-native sellers (usage-metered infra and communications APIs) have long treated machines as buyers via keys and programmatic top-ups. What is still thin is unsupervised close of complex B2B deals. That is why Atlas stays fail-closed until execute on your rails.⁷⁸

Sources

eMarketer, AI shopping tools gain traction… (2026): US ecommerce sales via general AI platforms about $20.9B in 2026, nearly 4× 2025, or about 1.5% of retail ecommerce. emarketer.com
Enterprise deployment figures vary by survey. For example, the Fortune AIQ Advisory Board survey (July 2025) reports widespread active agentic projects among respondents, while noting that many pilots do not reach full implementation. Read “most enterprises” as piloting or adopting somewhere, not as universal scaled autonomy. Fortune AIQ whitepaper (PDF)
Google, Universal Commerce Protocol and agentic commerce tooling (compatible with MCP, A2A, and AP2); industry coverage of ACP (OpenAI / Stripe) and MCP as the agent-to-tool layer. blog.google
Adobe Digital Insights holiday shopping data (Nov 1–Dec 31, 2025): generative AI referrals to US retail sites up about 693% for the season, and about 670% on Cyber Monday. Growth rates are period-specific. news.adobe.com
Gartner, top strategic predictions for 2026 and beyond (Oct 2025): by 2028, about 90% of B2B buying is expected to be AI-agent intermediated, moving over $15 trillion through agent exchanges. Intermediation is not the same as unsupervised settlement. gartner.com
McKinsey Global Survey on the state of AI: about 23% of organizations report scaling an agentic AI system somewhere in the enterprise, and about 39% say they have begun experimenting. Most scaling remains limited to one or two functions. mckinsey.com
Stripe, Machine Payments Protocol and Agentic Commerce Suite: open payment coordination for agents, plus catalog and checkout paths for selling through AI agents. Public examples include Browserbase metering agent browser sessions. stripe.com · Agentic Commerce Suite
Salesforce, B2B Commerce innovations (Agentforce Buyer Agent, GA June 2026): account-aware product discovery and ordering over channels such as WhatsApp and SMS. Salesforce cites Siemens running Agentforce B2B Commerce in over 25 countries. salesforce.com

The under-3% settlement / about 20-30% discovery framing is PairRail’s operating synthesis for seller planning, not a single audited market census. Analyst forecasts (including Gartner’s B2B intermediation outlook) describe direction and timing, not Atlas revenue. Named vendor programs illustrate that machine-facing commerce is shipping; they are not PairRail partners or customers unless stated elsewhere. Use them to prioritize catalog and protocol readiness before chasing fully autonomous checkout.

See how Atlas operates the control plane.

Operator workflow, trust boundaries, and Sandbox to Enterprise graduation live on Product.

Product walkthrough → Start Free