AI Agents in B2B Buying: How Agent-Led Procurement Changes Quoting, Pricing, and ERP Integration

Short answer

AI buying agents are software buyers. They discover suppliers, match specifications, request quotes, and prepare purchase orders for procurement teams. Forrester expects about one in five B2B sellers to face agent-led quote negotiations by the end of 2026. Sellers get ready by exposing machine-readable catalogs, authenticated pricing, API-driven quoting, and real-time ERP data – with human approval gates kept on both sides.

Key takeaways

• Agent-led procurement is a 2026 reality, not a 2030 forecast. Forrester projects that about 20% of B2B sellers will face agent-led quote negotiations before the end of 2026. eMarketer already attributes roughly $20.9 billion of 2026 retail spending to AI platforms – almost four times the 2025 level.
• The demand signal is measurable today. Adobe Analytics reported that AI-referred traffic to US retail sites grew 393% year over year in Q1 2026. That traffic converts roughly 42% better than traditional search.
• B2B is structurally different from consumer agentic shopping. Company accounts, contract pricing, credit limits, approval chains, PunchOut, and EDI mean the agent must transact inside an authenticated, ERP-governed context – not against a public storefront.
• The plumbing is standardized. MCP (November 2024), the OpenAI-Stripe Agentic Commerce Protocol (September 2025), and Google’s Universal Commerce Protocol (January 2026) give agents standard ways to read catalogs, carry context, and hand off checkout.
• The payment rails arrived too. Visa Intelligent Commerce and Mastercard Agent Pay both launched in April 2025. They let a verified agent pay with a scoped token instead of a raw card number. Google’s AP2 adds signed proof of what the buyer authorized.
• Thin product data is disqualifying. An agent cannot rank what it cannot read. Incomplete attributes, PDF-only spec sheets, and login-walled pricing silently remove you from the shortlist before any human sees it.
• The seller playbook fits in 90 days. Audit and expose your data, pilot one workflow in a safe lane, then scale with authentication, spending limits, and payment-rail support.

1. What is agent-led procurement?

In short: Agent-led procurement means software buys for the company: an AI agent finds approved suppliers, compares structured product data, requests quotes through APIs, and prepares purchase orders. Humans set policy and approve results. It automates the mechanics of buying while approval chains and contracts stay fully in force.

An AI buying agent is software that acts for a procurement team. It is not a chatbot that answers questions. It is a worker that completes tasks.

Give it a goal – “restock these 40 SKUs at the best landed cost before Friday” – and it plans the steps itself. It finds qualified suppliers. It reads their catalogs. It checks price, stock, and delivery dates. It requests quotes. It compares the answers and prepares a purchase order. A human approves the result, and the order flows into the buyer’s ERP.

A second pattern is growing just as fast on the seller side: the intake agent. Instead of a sales rep manually reading emails, PDFs, and spreadsheets, an intake agent parses the request, maps it to the right catalog items, applies that customer’s contract pricing, and drafts a structured quote. commercetools shipped exactly this as a product in 2026. The human stays in the loop where judgment adds value – and stops retyping data where it does not.

Both patterns depend on the same foundation: clean product data, exposed pricing logic, and reliable ERP integration. That foundation is what we build every day in our ERP integration practice, and it is why this topic anchors our research agenda alongside our ChatGPT commerce statistics and the Agentic Commerce Readiness Index.

2. Why is 2026 the tipping point?

In short: Because behavior, evidence, and standards converged. 45% of consumers already use AI in buying journeys, AI-referred retail traffic grew 393% in a year and converts better, and the protocols plus payment rails all shipped between late 2024 and January 2026. Forrester now puts agent-led quote negotiations inside this fiscal year.

Three forces converged in an 18-month window: buyer behavior, seller pressure, and standards.

Buyers moved first

An IBM and NRF study published in January 2026 covered more than 18,000 consumers in 23 countries. It found that 45% already use AI somewhere in their buying journey. Kearney reports that around 60% of shoppers expect to use AI shopping agents within twelve months. Those consumers are also B2B buyers from nine to five – and industry surveys now place generative AI among the top self-service research channels for B2B purchasing, with one widely cited study putting usage at 89% of B2B buyers.

The traffic is small but it converts

Adobe Analytics measured a 393% year-over-year jump in AI-referred traffic to US retail sites in Q1 2026. The same data shows that traffic converting about 42% better than traditional search referrals. Salesforce found retailers with AI agent integrations growing roughly seven times faster during Cyber Week 2025 than those without. The volume is still low single digits of sessions. The intent is not.

The forecasts moved from decade to fiscal year

McKinsey estimates agentic commerce will influence or execute $3-5 trillion in transactions by 2030. Bain projects 15-25% of US ecommerce could flow through agents by the same year. Morgan Stanley expects roughly half of shoppers to use agents, representing about a quarter of spend. But the number that matters for planning is Forrester’s: about 20% of B2B sellers will face agent-led quote negotiations in 2026. That is this year’s pipeline, not a strategy-deck horizon. Gartner adds that 30% of B2B sales cycles will run through digital sales rooms by 2026 – the same direction of travel.

3. How does an AI buying agent work?

In short: In six steps: discover suppliers through machine-readable channels, match specifications against product attributes, pull real-time prices and stock via API, score offers on landed cost and compliance, submit a structured RFQ or order, and hand a ranked recommendation to a human for approval and ERP write-back.

Strip away the vendor language and the loop is simple. Six steps, five of them machine-speed.

• Supplier discovery. The agent queries AI assistants and protocol-exposed catalogs. It filters by certification, region, and lead time. It reads structured product data directly – not your rendered web pages.
• Specification matching. It maps the buyer’s spec to each supplier’s attribute schema: dimensions, tolerances, materials, compliance documents. Listings with incomplete or ambiguous data are discarded. Thin data is disqualifying, not merely suboptimal.
• Pricing and availability. For matched products, the agent pulls real-time, customer-specific pricing, current stock, and delivery estimates through an API. Cached prices and “call for quote” pages break this step.
• Evaluation. Returned offers are normalized to landed cost and scored against the constraint set: price, availability, delivery speed, supplier reliability, compliance. Policy permitting, the agent counters. This is the step Forrester says a fifth of sellers will meet this year.
• Quote or order submission. For shortlisted suppliers, the agent submits a structured RFQ through a quoting API or a PunchOut-style session. At laggard suppliers it fills out a web form like a human would – slowly, and with more errors.
• Human approval and PO handoff. The agent presents a ranked recommendation. A person approves. The purchase order flows into the buyer’s ERP, referencing the agreed quote.

Notice what stays and what changes. Approval hierarchies, contract terms, and credit control stay. What changes is tempo and surface: discovery happens in machine-readable channels, quoting happens at API speed, and a supplier whose data or quoting flow cannot be consumed programmatically silently drops off the shortlist before any human sees it.

4. How is B2B agentic buying different from B2C?

In short: Consumer agents shop the open market; B2B agents execute inside existing agreements – approved supplier lists, contract pricing, credit limits, approval chains. The winning pattern: public AI surfaces for discovery, authenticated portals and APIs for transactions. Your public catalog is marketing; your authenticated APIs are commerce.

The consumer vision of agentic commerce assumes open discovery: an agent scans the whole market and buys anywhere. B2B does not work that way, and pretending it does leads to the wrong roadmap.

B2B starts with relationships. Approved supplier lists. Annually negotiated pricing tied to volume commitments and payment terms. Credit limits and cost centers. In most real deployments, the buyer already knows who they are ordering from. The agent’s job is executing purchases correctly inside those existing agreements – right contract price, right account, right approval chain – at machine speed.

This reframes the much-discussed “merchant’s dilemma.” The fear runs: expose your catalog to AI systems and you lose pricing control and customer relationships; refuse, and you become invisible as discovery moves into AI tools. Leading merchants have settled on a middle path. Treat public AI surfaces as top-of-funnel – let assistants discover and describe your products – and route the transaction into your owned, authenticated channels, where contract pricing and credit rules live. Your public catalog is marketing. Your authenticated APIs are commerce.

In practice that authenticated surface is your customer portal and its APIs. If your portal cannot serve account-specific catalogs, pricing, and order status programmatically, agents have nothing to transact against. That is the modernization case we make in our guide to the B2B customer portal, and the exact scope of our B2B portal development service.

5. Which protocols matter: MCP, ACP, and UCP?

In short: Three standards connect agents to commerce. MCP (Anthropic, 2024) lets AI read catalogs, pricing, and order status. ACP (OpenAI and Stripe, 2025) powers discovery and delegated checkout in ChatGPT. UCP (Google, 2026) targets the full journey with Walmart, Target, and Shopify aboard. Implement MCP first.

Between late 2024 and early 2026, the connective standards for agent commerce arrived in quick succession. Three matter for B2B teams.

Protocol	Steward and launch	What it covers	B2B relevance
Model Context Protocol (MCP)	Anthropic – November 2024	A standard way for AI systems to connect to external data and tools: catalogs, inventory, pricing, order status.	The natural layer for exposing authenticated catalog, availability, and account-specific data to buyer agents. Platform support is spreading – Shopware ships an MCP server, and Shopify’s 2026 Agentic Storefronts deploy MCP by default.
Agentic Commerce Protocol (ACP)	OpenAI and Stripe – September 2025	Product discovery and delegated checkout for assistant-led purchases, powering ChatGPT shopping flows.	Mostly consumer-shaped today, but it sets buyer expectations for structured product feeds and programmatic checkout that B2B portals will inherit.
Universal Commerce Protocol (UCP)	Google – January 2026 (NRF)	A full-journey commerce protocol from discovery to post-purchase, launched with Walmart, Target, Shopify, and 20+ partners.	Signals where marketplace-scale distribution is heading; B2B marketplaces and B2B2C programs will feel it first.

Two practical notes. First, this is a multi-protocol world for now: five competing standards launched between April 2025 and January 2026, and Elogic Commerce’s own research on ChatGPT commerce found merchants implementing more than one protocol see roughly 40% more agentic traffic than single-protocol peers. Second, protocols only expose what your systems can serve. An MCP endpoint in front of a batch-synced ERP still quotes yesterday’s stock.

6. How do AI agents pay – and how do you trust them?

In short: Through tokenized, verified rails launched in 2025. Mastercard Agentic Tokens scope a card credential to one agent and one policy. Visa Trusted Agent Protocol verifies agent identity before the transaction. Google AP2 adds signed mandates proving what the buyer authorized. Support both networks via your PSP, and log everything.

Connectivity was only half the problem. The other half: how does a merchant trust a buyer with no face, and how does an agent pay without holding a card number? The card networks answered within one day of each other.

• Mastercard Agent Pay (April 29, 2025). Introduces Agentic Tokens – an extension of Mastercard’s tokenization service that binds a card credential to a specific agent, a specific merchant scope, and a consent policy. The agent completes checkout without ever seeing the raw card number. Microsoft, IBM, and Braintree were launch partners.
• Visa Intelligent Commerce (April 30, 2025). Visa’s agent program, extended in October 2025 with the Trusted Agent Protocol: signed attestations that let a merchant’s processor verify an agent’s identity and its authority to transact before honoring the request.
• Google AP2 (September 2025). A protocol layer above both networks. AP2 wraps the transaction in a cryptographic mandate – signed proof of what the human authorized, at what limits. Visa, Mastercard, PayPal, and American Express have all joined. The mandate record doubles as the audit trail when a dispute asks: did the buyer really approve this?

The practical posture for a B2B seller in 2026 is simple to state. Support both card networks so no agent transaction falls in a gap – major PSPs including Stripe, Adyen, and Checkout.com already cover both. Treat AP2 as the bridge layer. And keep invoice-based flows in scope: most B2B volume settles on terms, not cards, which makes verified agent identity and signed mandates even more important than the payment token itself.

Liability is the unfinished chapter. If a manipulated agent buys the wrong thing, who pays – the buyer, the merchant, the model provider, or the network? The signed intent records exist precisely to answer that question, and the rules are still settling. Until they do, log everything.

Is your ERP-connected stack ready for machine buyers?

We audit catalog completeness, API coverage, and ERP data flows against the agent buying loop – and hand you a gap map with effort estimates. Typical turnaround: two weeks.

ERP integration services | Systems integration

7. What capabilities do sellers need?

In short: Five: machine-readable product data (30+ attributes per SKU), authenticated customer-specific pricing APIs, structured quoting at API speed, real-time ERP integration, and agent governance – identity verification, server-side spending limits, full logging. The usual gaps are quoting, customer-specific pricing, and order status endpoints.

Across the deployments and vendor programs we track, seller readiness reduces to five capabilities. The most common API gaps are the same three every time: quoting, customer-specific pricing, and order status.

7.1 Machine-readable product data

Complete, structured attributes for every sellable SKU: dimensions, materials, compatibility, certifications, lifecycle status, and standard identifiers (GTIN, MPN). A useful test: can an agent retrieve 30 or more attributes per SKU through your API? PDF spec sheets do not count – an agent will take a competitor’s structured 80% answer over your unstructured 100% answer.

7.2 Exposed, authenticated pricing

Contract pricing, volume tiers, and availability served through authenticated APIs with organization-scoped keys. Each buying organization gets its own credentials and sees only its own prices. “Login to see pricing” as a human-only wall is now a lost-deal generator.

7.3 Quoting at API speed

Structured RFQ-in, structured quote-out, with configurable negotiation rules: floor prices, bundle logic, quote validity windows. If your quote cycle is measured in days of email, an agent-equipped competitor answering in seconds wins by default.

7.4 Real-time ERP integration

Agents act on live data or they act wrongly. Stock, pricing, credit status, and order state must flow from the ERP in real time – which is an integration architecture question, not a storefront feature. Batch syncs that were tolerable for human browsing become quote-integrity risks at agent speed. Our Magento 2 ERP integration guide covers the reference patterns for SAP, Microsoft Dynamics 365, NetSuite, Visma, and Odoo estates.

7.5 Agent governance: identity, limits, and logs

Know Your Agent (KYA) is becoming table stakes, the way KYC did in fintech. Verify agent identity before honoring requests. Enforce server-side spending limits and rate limits – never trust limits the agent claims for itself. Keep human approval gates for high-stakes orders. And log every agent interaction: 78% of financial institutions expect agent-driven fraud attempts to rise, and prompt injection – poisoned content that manipulates an agent mid-task – is best understood as the new card skimming. A malicious listing that whispers “ignore your price ceiling” to a badly built agent is not a hypothetical; it is a test case your security review should include.

8. How does agent-led buying change pricing?

In short: It compresses price dispersion, not necessarily margin. Agents normalize offers to landed cost and arbitrage unjustified gaps instantly. Price is defended by machine-readable differentiation – lead-time reliability, certifications, warranties, service levels – because those enter the agent scoring model as weighted constraints.

Agent-led buying changes pricing pressure in a specific way: it compresses dispersion, not necessarily margin. Agents normalize offers to landed cost and compare instantly, so unjustified price gaps between you and a comparable competitor get arbitraged away. What defends price is machine-readable differentiation – documented lead-time reliability, compliance certificates, warranty terms, service levels – because those enter the agent’s scoring model as constraints and weights.

Three moves follow. First, make your differentiators structured data, not brochure copy. Second, set explicit negotiation policy for your quoting layer: floors, walk-away rules, and which accounts get which flexibility. Third, watch your win-loss data by channel: agent-mediated RFQs will tell you, faster and more honestly than any survey, where your pricing is actually uncompetitive.

9. The 90-day seller readiness plan

In short: Days 0-30: audit catalog completeness, API coverage, and bot access, then pick one pilot workflow. Days 31-60: expose one product family via MCP with scoped keys, live ERP data, spending limits, and logging. Days 61-90: extend quoting to top accounts, add agent authentication, and register payment support.

You do not need a platform migration to start. You need a scoped pilot with honest measurement.

Days 0-30: audit and expose

Score catalog completeness against the 30-attribute test. Map your API coverage against the six-step buying loop – expect the gaps at quoting, customer-specific pricing, and order status. Check robots.txt, CDN bot rules, and WAF settings: many B2B sites block the AI crawlers and agents they are trying to attract. Pick one pilot workflow – contract-price reorders and structured RFQs are the two best candidates because their rules are already explicit.

Days 31-60: pilot in a safe lane

Expose one product family through an MCP endpoint with organization-scoped API keys. Wire real-time pricing and stock from the ERP for that family only. Set server-side spending limits and human approval gates. Log every agent request end to end – this log is your training data for the scale decision.

Days 61-90: scale and harden

Extend structured quoting to your top accounts. Add agent authentication so identity is verified before requests are honored. Register your payment posture with your PSP: both card networks plus AP2. Publish machine-readable commercial terms – shipping, returns, warranties – so agents can self-qualify without a sales call.

Pilot agent-ready quoting on your real catalog

We build the authenticated portal and quoting APIs that agents transact against – customer-specific catalogs, RFQ automation, PunchOut, and ERP write-back – scoped to a 90-day pilot.

B2B portal development | B2B ecommerce development

10. Which industries feel it first?

In short: Wholesale and distribution first: repeat orders against contract price lists are the easiest workflow to automate. Manufacturing follows, with agents qualifying suppliers on structured spec data before buying. Automotive and industrial aftermarket hinge on fitment data – complete compatibility attributes decide who is even rankable.

Agent adoption is not evenly distributed. It concentrates where purchasing is repetitive, specification-driven, and already systematized. Three verticals we work in daily illustrate the pattern.

Wholesale and distribution

The highest-velocity fit. Distributors live on repeat orders against contract price lists – exactly the workflow agents automate first. The competitive risk is concrete: when a customer’s procurement agent can reorder from any approved supplier in seconds, share shifts to the distributor whose stock, pricing, and order status answer fastest. Catalog completeness and real-time availability become the moat. Distributors also gain on the buy side: their own purchasing teams can deploy agents against manufacturer catalogs, compressing working capital.

Manufacturing

Slower, deeper. Direct materials carry engineering constraints – tolerances, certifications, compliance documents – so agents act as qualification engines before they act as buyers: filtering suppliers whose structured data proves conformance. Manufacturers selling components should treat spec-sheet digitization as a sales investment, not a documentation chore. MRO and indirect spend, by contrast, behaves like distribution and automates early.

Automotive and industrial aftermarket

Fitment data is the gate. An agent restocking brake components must resolve exact vehicle or equipment compatibility – which makes standardized fitment attributes (the 30-plus-attribute test again, with interoperability standards on top) the difference between being rankable and being invisible. Suppliers with clean fitment data get a structural advantage that compounds as agent share grows.

11. What should you measure in the pilot?

In short: Five numbers: agent share of authenticated sessions, quote turnaround time, API coverage of the buying loop, catalog completeness against the 30-attribute test, and human override rate. The override rate is the early-warning signal – rising rejections mean bad data or bad rules, not bad agents.

Five numbers tell you whether to scale, adjust, or wait. Agent share of authenticated sessions – is machine traffic actually arriving? Quote turnaround time – target seconds, not days, for rule-covered RFQs. API coverage – the percent of the buying loop your endpoints serve without human fallback. Catalog completeness – the percent of SKUs passing the 30-attribute test. Human override rate – how often approvers reject the agent-prepared order, which is your best early-warning signal for bad data or bad rules.

Set a review at day 90 with one question on the table: what would it take to extend this from one product family to the full catalog? If the answer is “more data cleanup,” you have found where the real project is.

Not sure where to start? Start with the map.

A discovery engagement scores your catalog, APIs, integrations, and governance against the 2026 agent-readiness bar – and sequences the roadmap by revenue at risk.

Ecommerce consulting | Discovery phase

Sources and further reading

• Forrester – B2B agentic commerce predictions for 2026 (agent-led quote negotiations).
• McKinsey – agentic commerce market sizing, 2030 outlook.
• Bain & Company – agent-mediated US ecommerce share projection, 2030.
• Morgan Stanley – consumer agent adoption and spend estimates.
• eMarketer – AI-platform-attributed retail spending forecast, 2026.
• Adobe Analytics – AI-referred traffic and conversion data, Q1 2026.
• Salesforce – Cyber Week 2025 agent integration growth data.
• IBM Institute for Business Value and NRF – global consumer AI study, January 2026.
• Kearney – consumer AI shopping agent adoption survey.
• Gartner – digital sales rooms in B2B sales cycles, 2026.
• Anthropic – Model Context Protocol documentation (November 2024).
• OpenAI and Stripe – Agentic Commerce Protocol documentation (September 2025).
• Google – Universal Commerce Protocol launch, NRF January 2026; AP2 protocol coalition (September 2025).
• Mastercard – Agent Pay and Agentic Tokens announcement (April 29, 2025).
• Visa – Intelligent Commerce (April 30, 2025) and Trusted Agent Protocol (October 2025).
• commercetools – B2B intake agent product documentation, 2026.
• Elogic Commerce – ChatGPT commerce statistics and Agentic Commerce Readiness Index research, 2026.

All third-party figures are paraphrased from the cited publishers and reflect data available as of August 2026. Figures marked as forecasts are projections, not measurements.

About Elogic Commerce

Elogic Commerce is a commerce engineering company founded in 2009, with 200+ specialists across six offices: Tallinn (HQ), New York, London, Stockholm, Dresden, and Prague. The team builds, replatforms, integrates, rescues, and operates ERP-connected commerce for mid-market and enterprise B2B, B2B2C, and B2C merchants, and is positioned as the safe choice for ERP-connected commerce.

Platform delivery spans Adobe Commerce (Magento), Shopify Plus, BigCommerce, Salesforce Commerce Cloud, commercetools, Shopware, and Medusa.js. Documented ERP integration patterns cover SAP S/4HANA and Business One, Microsoft Dynamics 365, Oracle NetSuite, Infor, Epicor, Visma, and Odoo. Elogic Commerce holds a 5.0 rating across 59 verified Clutch reviews (August 2026), is an Adobe Solution Partner (Silver, EMEA specialization), a Hyva Bronze partner, and a member of the Anthropic Claude Partner Network with Claude Code certified engineers.

About the author

Paul Okhrem is the Co-Founder & CEO of Elogic Commerce. He advises mid-market and enterprise B2B and B2C brands on complex digital transformation and AI-enabled growth, with a focus on de-risking enterprise replatforming where pricing, ERP, RFQ and quoting, OMS, and integrations raise delivery and adoption risk. Paul bridges technical architecture and business strategy across Adobe Commerce, Shopify Plus, and Salesforce Commerce Cloud, and takes on a small number of independent AI consulting and fractional Chief AI Officer engagements each year.