Agentforce Commerce reached general availability in early July 2026 with three specialized agents: Shopper Agent, Buyer Agent and Merchant Agent. It's the biggest agentic commerce launch Salesforce has shipped — and the announcement material is generous on conversion numbers, thin on guardrails.
That's not an accident of communication. It's a choice of emphasis any decision-maker needs to recognize before turning the agent loose on their own storefront: the vendor shows what the agent gains in speed and conversion, but leaves it to the customer to decide where a sale needs a human nearby before it closes.
What Agentforce Commerce actually shipped
The three agents cover different roles in the sales chain, and understanding the split keeps you from treating "Agentforce Commerce" as a single product.
Shopper Agent carries the end buyer's conversation from start to finish — product discovery, checkout and post-purchase service — while keeping the brand's voice on the customer's own storefront. It checks inventory, carrier cutoff times and store-pickup options within the same conversation, without handing the buyer off between systems.
Buyer Agent tackles the other side of commerce: B2B procurement over WhatsApp and SMS. The example Salesforce uses is direct — a buyer texts "need 40 cases of the 16oz fasteners, same as the March order," the agent sends a product image to confirm the SKU, shows the current contract price and completes the order, with no portal login and no phone call.
Merchant Agent sits on the operations side: the catalog team organizes products, creates "boost and bury" rules and adjusts display order by describing what it wants, instead of navigating an admin menu.
The launch numbers reinforce the sense of urgency: during last year's holiday season, AI influenced 20% of global online sales — roughly $262 billion — and retailers already running their own shopper agent grew sales 59% faster than those who hadn't adopted yet. Traffic referred by AI converts at 8x the rate of social traffic. Early Merchant Agent customers report an 88% reduction in catalog task completion time.
What the launch coverage doesn't mention
None of the sources covering the announcement — press release, trade press, market analysis — included a line about approval limits, transaction value caps, or what happens when Buyer Agent confirms the wrong SKU before the buyer notices. The emphasis is 100% capability and adoption, 0% risk.
That doesn't mean the risk isn't there — it means Salesforce is selling the platform, and it's on the customer to design the control. It's the same pattern we already covered when mapping where MCP breaks: the protocol itself isn't the problem, it's the distance it creates between an automated decision and human review — and that distance is a configuration decision, not a product one.
It's also the same question we already asked about Agentforce in human-staffed service: what to automate and what not. There the criterion was emotional capacity and case ambiguity; here it's financial reversibility and contract exception — the framework shifts axis, but the underlying logic is the same.
The launch sells autonomy. The guardrail against a costly mistake is still the customer's job, not the product's.
Buyer Agent is the clearest case. Confirming a SKU by image reduces friction — but it also reduces the number of points where a human intercepts a wrong order before it turns into an invoice. Automatically displaying contract pricing is real convenience, until the contract has an exception the agent wasn't trained to recognize.
Three variables that decide if the agent can sell on its own
The right question isn't "can the agent sell on its own" — it's "under what conditions." Three variables, in the order worth checking:
- Reversibility of the transaction. A low-value item exchange is cheap to undo. A B2B order for 40 cases at the wrong contract price generates an invoice, freight and reconciliation rework — an expensive, non-trivial reversal.
- Exception complexity. How much of the sales process depends on negotiation, off-catalog discount or a non-standard contract condition? A fixed-price catalog tolerates full automation. A contract with a customer-specific clause doesn't.
- Ticket size and frequency. A recurring low-value purchase — the same risk logic that already applies in Salesforce's ROI matrix, where average ticket decides whether the platform pays for itself — tolerates automatic decisions. A one-off high-value purchase calls for approval before completing.
None of the three variables decides alone. A Shopper Agent selling a standard catalog item, low ticket, reversible transaction, meets all three — full autonomy is a low-risk call. A Buyer Agent closing a B2B order with contract pricing, high ticket, expensive reversal, fails at least two — there the right design is automatic approval within a predefined range, escalating to a human outside it.
Where full autonomy already makes sense — and where it's still early
Applying the criterion to the three agents:
Shopper Agent in standard-catalog B2C is already mature enough for full autonomy. Fixed-price item, visible inventory, clear return policy — all three variables in the criterion point to low risk. It's where the conversion gain cited at launch (8x on AI-referred traffic) makes the most sense to capture quickly.
Buyer Agent in B2B with contract pricing still calls for an approval design — not because the agent is bad, but because contract exceptions are the norm in B2B procurement, not the rare case. A value range with automatic approval, and anything outside it (a new SKU, a discount beyond what was agreed, a new customer's first order) going to human review before completing, is the design that avoids the costly mistake without bringing back the friction the agent was meant to remove.
Merchant Agent sits in the middle: the action itself (changing a sort rule, adjusting boost/bury) is internal and reversible — it doesn't close a sale, doesn't generate an invoice. The real risk here isn't financial, it's operational: a bad rule can hide the right product or surface the wrong one for days before anyone notices. Change logging and easy rollback solve that risk without requiring prior approval for every adjustment.
- Shopper Agent, standard catalog, low ticket: full autonomy, no human approval in the flow.
- Buyer Agent, B2B contract, price exception: automatic approval within a defined range; outside it, a human before completing.
- Merchant Agent, catalog rule: no prior approval, but with change logging and fast rollback available.
Low ticket and standard SKU tolerate full autonomy; contracts and price exceptions call for a human in the loop.
Fast adoption doesn't replace control design
The launch numbers are real and so is the incentive to adopt fast — whoever waits misses the window of AI-referred traffic that Salesforce is natively connecting to ChatGPT, Google AI Mode and the Gemini app through the second half of 2026. But adopting fast and adopting without control design are different decisions, and the launch material doesn't separate the two.
The work left to the customer — because no launch coverage addressed it — is deciding, agent by agent, where the transaction is reversible enough to run on its own and where a contract exception still calls for a human before completing. Whoever skips that design doesn't find the mistake on launch day. They find it on the first exception the agent treated as the rule.
Questions that keep coming back
Before wrapping up, the three most common questions about the launch.
What is Agentforce Commerce?
Agentforce Commerce is Salesforce's agentic commerce architecture, launched into general availability in July 2026, with three specialized agents: Shopper Agent (serves the end buyer from discovery through post-sale), Buyer Agent (B2B procurement via WhatsApp and SMS) and Merchant Agent (natural-language catalog management). All three have native connections to inventory, order management and customer data — that integration is what sets them apart from a generic chatbot, letting the agent promise a delivery date, honor contract pricing and keep context without handing the buyer off between systems.
Can Agentforce Commerce sell without human review?
It depends on the agent and the transaction — there's no single answer. Three variables decide: transaction reversibility, exception complexity (negotiation, discount, off-catalog contract) and ticket size/frequency. A Shopper Agent on a standard, low-ticket catalog meets all three low-risk conditions — full autonomy already makes sense. A Buyer Agent closing a B2B order with contract pricing fails at least two — expensive reversal, frequent exceptions — so it calls for automatic approval only within a predefined value range, escalating to a human outside it.
Why doesn't Salesforce talk about guardrails at launch?
Because the launch material — press release, trade coverage, market analysis — was built to sell capability and adoption speed, not to design risk controls. None of the announcement sources mention an approval ceiling, a transaction limit, or what happens when the agent confirms the wrong SKU before the buyer notices. That's not a product flaw: it's a configuration decision that falls to the customer, the same way it does for any system that automates a decision with real-world effect — the closer the action gets to generating an invoice or moving inventory, the closer it needs to sit to human review before completing.