Salesforce acquired Cimulate, and most coverage treated it as a Commerce Cloud feature announcement. That framing misses the architectural significance. The agentforce commerce conversational AI direction this signals is a fundamental shift in how discovery, intent, and transaction get wired together across the Salesforce platform.
The keyword search model has been the default for 25 years. Cimulate’s acquisition was a direct bet that it won’t be the default for the next five. That bet is now paying out in shipping product.
Why Keyword Search Breaks at Enterprise Scale
The problem isn’t that keyword search is slow. The problem is that it externalizes the cognitive work onto the customer.
A shopper who types “blue running shoes” is doing the platform’s job for it. They’re translating an intent (“I want to run a 10K without knee pain”) into a query format the system can process. Every step of that translation is a place where intent gets lost, results get irrelevant, and conversion drops.
At scale, this compounds. Enterprise commerce orgs with 50,000+ SKUs, multiple product lines, and complex attribute taxonomies see this acutely. Faceted search helps. Merchandising rules help. But both are still fundamentally reactive to the query the customer typed, not the goal they arrived with.
Conversational discovery inverts this. The system does the translation work. The customer states a goal, a constraint, a context. The architecture resolves that into a product recommendation, a guided flow, or a clarifying question. The cognitive load shifts from customer to platform.
That shift has real conversion implications. In commerce orgs with high SKU complexity, the gap between “customer arrived with intent” and “customer found the right product” is where most abandonment lives.
What the Cimulate Capability Has Become
Cimulate built a conversational product discovery layer trained specifically on commerce data. Not a general-purpose LLM wrapper. A model that understands product catalogs, attribute relationships, and purchase intent signals at a level that generic AI doesn’t.
That capability has now matured well beyond its initial Spring ‘26 beta. Guided Shopping is generally available. Contextual Search has evolved into semantic, intent-aware search powered by a commerce-optimized small language model that operates on shopper intent and session context rather than keyword matching. The architecture is no longer a chatbot sitting in front of a search bar. It’s a reasoning layer operating on the product graph itself, now extended to external discovery surfaces you don’t control.
The channel surface has expanded significantly. Beyond the original ChatGPT catalog feed integration, Salesforce has formalized the Agentic Commerce Protocol (ACP) as the mechanism for agent-to-agent and agent-to-shopper commerce flows across external AI surfaces. Google AI commerce surfaces now support native checkout paths. One-click checkout with deeper Stripe and Adyen payment orchestration enables direct purchase completion inside these AI channels without redirecting to your storefront.
The merchandising automation layer has also broadened. Agentic Boost and Bury, now in Winter ‘26 beta as Agentforce Actions for Merchandising, automatically promotes trending or high-margin items and suppresses out-of-stock ones based on real-time catalog signals. This is no longer a configuration option. It’s an agent action that runs continuously.
The integration surface architects need to account for:
- Product catalog structure: attribute hierarchies, variant relationships, bundle configurations
- Inventory and availability: real-time stock signals that constrain recommendations, surfaced through Agentforce Actions for Merchandising
- Customer profile data: purchase history, preference signals, segment membership from Data Cloud
- Pricing and promotions: entitlement logic triggerable via Marketing and Commerce Connected Journeys, now including payment orchestration for AI channel checkout
- External AI channel feeds: ACP-compliant catalog and checkout endpoints for ChatGPT, Google AI surfaces, and any future agent-to-agent commerce flows
Getting conversational discovery right requires all five to be queryable in near-real-time. That’s a Data Cloud architecture problem as much as a Commerce Cloud problem.
The Data Cloud Dependency Is Non-Negotiable
Most architects will try to wire conversational commerce directly to Commerce Cloud APIs and call it done. That works for simple catalogs. It breaks for anything complex.
The reason is context. A conversational agent that can only see the product catalog will give technically correct answers that are commercially wrong. It might recommend a product the customer already bought last month. It might suggest a bundle that’s out of stock in the customer’s region. It might miss that this customer is on a B2B contract with different pricing.
The architecture that survives here puts Data Cloud at the center. The Unified Individual profile, Calculated Insights for purchase behavior, and real-time Data Streams from Commerce Cloud all feed the context layer that the Agentforce agent reasons against.
Specifically, the pattern that works:
- Data Cloud ingests Commerce Cloud order history, browse behavior, and cart events via Data Streams
- Identity Resolution unifies these signals against the customer profile (critical for B2C orgs where guest checkout creates fragmented records)
- Calculated Insights compute preference signals: category affinity, price sensitivity, brand loyalty scores
- Data Graphs pre-compute the joins between customer profile, product preferences, and segment membership
- The Agentforce agent queries this materialized context at inference time, not raw transactional data
The Data Graph step is where most implementations skip corners and pay for it later. Pre-computing joins is the difference between a 200ms response and a 2-second response. In a conversational interface, 2 seconds kills the experience.
Data 360 Zero Copy Access for Commerce, slated for Winter ‘26 GA, extends this pattern by unifying storefront, service, marketing, and inventory data without replication overhead. When it ships, it removes one of the main reasons orgs have been reluctant to put Data Cloud at the center of commerce architecture.
The Data Cloud identity resolution architecture breakdown maps how the identity layer handles the fragmented-record problem that’s endemic in commerce.
Agentforce Agent Design for Commerce Contexts
The current release ships several agent patterns with distinct design requirements: Guided Shopping Agents for on-site assistance (now GA), Two-Way Email agents that turn promotional sends into interactive purchase threads, and emerging Point-of-Sale agents in Spring ‘26 pilot. The design logic differs between them, but all expose the same underlying constraint: commerce agents operate in high-volume, low-context environments.
A service agent might handle 500 conversations a day with rich case history. A Guided Shopping Agent might handle 50,000 sessions a day where the customer has provided two sentences of context and expects a useful answer in three seconds. Two-Way Email adds a different constraint: the agent must maintain coherent context across an asynchronous thread that might span hours. Point-of-Sale adds yet another: the agent must operate within the latency envelope of a physical transaction.
That changes how you configure Topics, Actions, and Instructions.
Topics need to be scoped tightly. A single “product discovery” topic that tries to handle everything from initial browsing to post-purchase upsell will produce inconsistent behavior. Separate discovery, comparison, and purchase assistance into distinct topics with clear handoff logic. For Two-Way Email specifically, add a thread-context topic that reconstructs session state from prior email turns before the Atlas Reasoning Engine begins its reasoning loop. For ACP-based external channel flows, add a channel-context topic that normalizes the incoming intent signal before routing to discovery or checkout actions.
Actions need to be fast. Every Action call adds latency. In a commerce context, the Atlas Reasoning Engine needs to complete its reasoning loop quickly enough that the experience feels conversational. This means pre-fetching context via Data Graphs rather than making live API calls to Commerce Cloud at inference time. Guided Discovery’s smart alternatives for zero-result queries should be implemented as a pre-computed Action against the Data Graph, not a live catalog scan. Blocking on analytics writes during a live session is an avoidable latency source.
Instructions need to encode merchandising logic that would otherwise live in rules engines. Promotional priorities, brand restrictions, margin-sensitive recommendations, and the Agentforce Actions for Merchandising rules all need to be expressed as behavioral constraints in the agent’s Instructions, not as post-processing filters. Filtering after the fact produces recommendations that feel arbitrary to the customer.
The Agentforce agent design patterns for enterprise article covers the broader Topic/Action/Instruction architecture. Commerce contexts add the constraint that latency tolerance is lower and session context is shallower than most enterprise agent deployments.
Integration Implications for Existing Commerce Architectures
Most enterprise commerce orgs aren’t starting from a clean slate. They have existing search infrastructure, often Solr or Elasticsearch-based, with years of merchandising rules, synonym libraries, and relevance tuning baked in.
The question architects will face: replace or augment?
Replacing existing search with a conversational layer is the right long-term direction. Guided Shopping’s GA status and the maturity of intent-aware search make that path more concrete than it was a year ago. But it’s still a 12-18 month program for any org with significant search customization. The merchandising logic embedded in existing search configurations doesn’t automatically translate to agent Instructions.
Augmenting is the practical near-term path. Run intent-aware search and Guided Shopping Agents as parallel channels alongside existing search. The ACP-based ChatGPT and Google AI channel integrations are relatively low-code activations that extend your discovery surface without touching existing search infrastructure. Use that parallel operation as a data collection phase to understand which intent patterns the conversational layer handles better, and which still benefit from tuned keyword relevance.
The integration architecture for augmentation: conversational agent handles initial intent capture and disambiguation, hands off to existing search infrastructure for catalog retrieval when the intent is sufficiently specific, then re-engages for comparison and decision support. MuleSoft is the right integration layer where existing search infrastructure is external to Salesforce, because you need bidirectional data flow with transformation logic that Commerce Cloud’s native connectors won’t handle cleanly. This is also where Order Routing for Order Management (Winter ‘26 GA) connects: routing logic needs to be aware of both the conversational session context and inventory state, which requires a transformation layer between systems.
This buys 18-24 months of parallel operation while the conversational layer matures and merchandising logic migration happens incrementally. Holiday 2025 data from Salesforce shows 35% more site time and doubled AI-search traffic for orgs running conversational features, which gives you the business case to fund the migration while the legacy infrastructure stays live.
What Architects Should Do Now
Agentforce Commerce has moved from roadmap to shipping product across multiple GA features. The architectural decisions you make now determine whether you activate cleanly or spend 18 months untangling technical debt.
Three things to address immediately:
First, audit your Data Cloud implementation against the commerce context requirements above. If you don’t have Commerce Cloud order history flowing into Data Cloud via Data Streams, Identity Resolution configured to handle guest checkout fragmentation, and Calculated Insights computing purchase behavior signals, you’re not ready for conversational commerce regardless of what’s already in your org. Data 360 Zero Copy Access for Commerce, when it reaches GA in Winter ‘26, will reduce the replication overhead, but the identity and insight layers still need to be in place before it can help you.
Second, document your existing search merchandising logic. Not the Solr configuration files. The business logic: what gets promoted, what gets suppressed, what rules govern recommendations by customer segment. This is the hardest migration artifact to reconstruct later, and most orgs have it living in the heads of two or three people who’ve been tuning the search engine for years. Agentforce Actions for Merchandising will need this logic expressed as Instructions and agent actions, not inferred from configuration files.
Third, evaluate your Agentforce maturity against commerce-specific requirements. If you haven’t deployed an Agentforce agent in production yet, start with a lower-stakes use case first. The Agentforce implementation guide covers the foundational deployment patterns. Commerce is not the right first agent deployment for an org that hasn’t worked through the Topics/Actions/Instructions design cycle on a simpler problem. The Point-of-Sale pilot and ACP channel integrations add further complexity that compounds quickly if the foundational agent design patterns aren’t already solid.
The current decision on Data Cloud architecture determines whether conversational commerce is a 3-month activation or a 12-month rebuild.
Key Takeaways
- Guided Shopping is generally available; intent-aware semantic search, Agentforce Actions for Merchandising (beta), and Order Routing for Order Management (Winter ‘26 GA) have materially expanded what’s shipping versus the original Spring ‘26 release.
- Conversational commerce requires Data Cloud as the context layer. Commerce Cloud APIs alone are insufficient for personalized, real-time recommendations at enterprise scale.
- Data Graphs are the performance-critical component. Pre-computing joins between customer profile, purchase history, and product preferences is the difference between a viable conversational experience and an unusable one.
- The Agentic Commerce Protocol (ACP) is the mechanism for agent-to-agent commerce flows across ChatGPT, Google AI surfaces, and future external channels. Orgs without ACP-compliant catalog and checkout endpoints will be invisible on these surfaces.
- Orgs that don’t have Data Cloud identity resolution configured for guest checkout fragmentation will face a foundational rebuild before conversational commerce can work correctly.
- The merchandising logic embedded in existing search configurations is the hardest migration artifact. Document it now, before the migration pressure arrives.
