In 2026 Adobe and Salesforce both repositioned their customer data platforms as the context layer for AI agents. At Dreamforce 2026, Salesforce described Data 360 (formerly Data Cloud) as a governed context service that any authorised agent can call, and shipped a hosted Model Context Protocol (MCP) server over its APIs. At Summit in April, Adobe launched CX Enterprise Coworker on Adobe Experience Platform (AEP), built on MCP and the Agent2Agent (A2A) protocol. Info-Tech Research Group’s Dreamforce 2026 note summarises the Salesforce position as “the runtime context layer between enterprise data and AI agents”; Adobe’s own announcement describes a near-identical role for AEP.

The practical consequence for buyers is that an Adobe vs Salesforce AI agents comparison conducted at the feature level now produces two broadly equivalent lists. Both vendors offer agents for audience building, journey creation, analytics and content; both expose them through open protocols; both federate queries into Snowflake, Databricks and BigQuery without copying data. The differences that remain are in where each vendor’s agents find their authoritative customer record, and in how each vendor charges for agent activity.

This analysis argues that the choice between the two is therefore decided by which system already holds the organisation’s operational customer record, not by agent capabilities. Where that record lives in a cloud data warehouse, either vendor’s agents become largely interchangeable consumers of it, and the evaluation shifts to cost model, governance and operating fit.

Architectural convergence: MCP access, zero-copy federation and the CDP as agent context

The two vendors’ agent stacks now share the same four layers. Mapping them side by side shows how little structural difference remains.

The mechanism behind the convergence is the same on both sides. An agent needs three things to act on a customer: an identity it can resolve, attributes and history it can read, and permitted actions it can call. In 2024 each vendor supplied all three only inside its own suite. By late 2026 each vendor exposes the profile and the actions through MCP (the pattern examined in MCP customer data access), so an agent built elsewhere can call them, and each vendor reads attributes from the customer’s warehouse at query time rather than requiring them to be ingested first.

Two consequences follow. First, agent features are no longer a durable source of differentiation, because an organisation can point a third-party agent at either platform’s MCP endpoint. Second, the zero-copy layer means neither vendor needs to hold the full customer history to serve an agent; it needs to hold, or reach, whichever record the organisation treats as authoritative.

The convergence is not total. Salesforce’s MCP server is described by Salesforce as hosted and managed, and Info-Tech notes that the broader AI control plane spanning Data 360, Tableau and Informatica begins rollout only in Salesforce’s fiscal 2028. Adobe announced CX Enterprise Coworker in April with general availability “in the coming months”, and buyers should confirm current availability status in writing during evaluation.

Salesforce: agents anchored to the CRM transaction record

Salesforce’s agent strategy is anchored to the operational record that Sales Cloud and Service Cloud already hold: accounts, contacts, opportunities, cases, entitlements and orders. Data 360 sits on the same core platform metadata, so an Agentforce agent reading a unified profile, assembled by Data 360 identity resolution, is reading the same objects that sales and service users update during the working day. For an organisation whose customer truth is the CRM, this removes a synchronisation step that would otherwise sit between the agent and the record.

Marketing Cloud Next extends the same logic to marketing. It is built natively on Data 360 and the core platform, uses the Salesforce Flow engine for campaign orchestration instead of Journey Builder, and serves B2B and B2C use cases in one application, replacing the separate Marketing Cloud Engagement and Account Engagement (formerly Pardot) products for new deployments. Agents on Marketing Cloud Next draft campaign briefs, generate segments and optimise send times against the same profile the service agent uses.

Pricing follows the architecture. Marketing Cloud Next requires Data 360 provisioning and is sold on consumption credits for messaging, data processing and AI, rather than the subscriber-based licensing of the legacy product. Agentforce actions are metered through Flex Credits, which Salesforce prices at $500 per 100,000 credits, with a standard agent action consuming 20 credits, or roughly $0.10. The commercial signal is strong: Salesforce reported Agentforce annual recurring revenue above $1.5 billion in its most recent quarter, up more than 240% year on year.

Two qualifications apply. Existing Marketing Cloud Engagement customers face a convergence rather than a migration, and Salesforce had not announced a sunset date for the legacy product as of early 2026; organisations with large Journey Builder and data-extension estates will typically run both stacks in parallel for an extended period. Consumption pricing also makes cost harder to forecast in advance, since agent volume, segment refresh frequency and federated query load all draw on the same credit pool.

Adobe: agents anchored to content, experience and discovery data

Adobe’s agent strategy is anchored to a different record: digital behaviour, content performance and experience delivery. AEP has historically ingested web and app event streams at scale, Adobe reports AEP powering more than one trillion experiences a year, and Customer Journey Analytics, Journey Optimizer, Experience Manager and Target all operate on that event and content layer, joined by AEP identity resolution. CX Enterprise Coworker is positioned to monitor signals across these applications, recommend next-best actions and execute them across channels with a human approval step.

The AEP Agent Orchestrator already routes natural-language requests to specialised agents: an Audience Agent that detects audience size changes and duplicates, a Journey Agent that creates, analyses and simulates journeys, a Data Insights Agent inside Customer Journey Analytics and an Experimentation Agent for Journey Optimizer. These agents are strongest where the work is content-heavy and experiment-driven, which reflects the data AEP holds most completely.

Adobe has also extended the record it controls into AI search. It completed the $1.9 billion acquisition of Semrush in April 2026, and its LLM Optimizer product added detection of content hidden from large language models and controlled citation experiments. Adobe’s 2025 holiday data reported LLM-referred traffic growing almost sevenfold year on year. For organisations whose acquisition depends on being discovered and cited by AI assistants, this is a capability Salesforce does not currently offer at comparable depth.

The qualifications are material. CX Enterprise Coworker was announced with general availability to follow, so parts of the orchestration story were roadmap at announcement. AEP is licensed as a platform with application add-ons, which tends to produce a higher fixed entry cost than starting with a single Salesforce cloud. Adobe also enters a leadership transition: Anil Chakravarthy, who leads the Customer Experience Orchestration business, becomes chief executive on 1 December 2026. Where an organisation’s customer truth is transactional and CRM-resident, AEP agents typically depend on a CRM connector to reach it, which reintroduces the synchronisation step that Salesforce avoids.

Decision rule: locating the operational customer record

The convergence described above reduces the vendor decision to a single prior question: which system does the organisation already treat as the operational customer record? The operational record is the store whose values are used to make commitments to the customer, such as price, entitlement, eligibility, consent and case status. It is distinct from the analytical record used for reporting.

Three patterns are common.

• The CRM holds the operational record. Accounts, opportunities, cases and orders are mastered in Salesforce, and other systems receive copies. Salesforce’s agents read that record natively, with no synchronisation step. Adobe remains a viable choice for content and experimentation but its agents depend on a connector for the authoritative record. Salesforce is typically the lower-friction choice.
• The experience and content layer holds the operational record. The customer relationship is predominantly digital, the CRM is thin or absent, and the most decision-relevant data is behavioural: sessions, content engagement, offer response and AI-search referral. Adobe’s agents operate on the record they already hold. Adobe is typically the lower-friction choice.
• A cloud data warehouse holds the operational record. Customer, transaction and consent data are modelled in Snowflake, Databricks or BigQuery, and the CRM and marketing tools are downstream consumers. Both vendors’ agents reach this record through zero-copy federation and MCP, and neither has a structural advantage in access. The evaluation should then be decided on cost model, governance controls and the organisation’s existing operating skills, and the agent layer can be treated as replaceable.

The third pattern is becoming more common in organisations that adopted warehouse-native or composable architecture, and it is the one in which vendor lock-in risk is lowest. It is not universally appropriate: organisations without a maintained warehouse customer model, or with real-time decisioning needs below the latency of federated queries, are usually better served by allowing one vendor’s platform to hold the operational record.

Commercial and operating-model considerations

The two vendors also differ in how agent activity is priced, and this difference becomes more significant as agent volume grows. Salesforce meters agent actions and data processing through credits, so cost scales with usage and is visible per action. Adobe licenses AEP and its applications as platform capacity, with agent capability dependent on which applications are licensed, so cost is more predictable but carries a higher fixed base. Buyers should model both against the same assumed volume of agent actions, federated queries and profile refreshes, rather than comparing list prices.

Market traction differs in scale. Salesforce reported Agentforce annual recurring revenue above $1.5 billion; Adobe reported AI-first annual recurring revenue above $650 million against total annual recurring revenue of $27.5 billion, with AEP annual recurring revenue growing above 30% earlier in its fiscal year. Neither figure measures marketing-agent adoption specifically, and both include products outside the scope of this comparison.

Analysts are consistent that the operating model, rather than the technology, constrains outcomes. Forrester’s Joe Cicman observed after Adobe Summit 2026 that “simply deploying an agent capability into an operating model does not move the needle”, and recommended that buyers ask each vendor how it has run its own platform internally before selling it. The same applies to Salesforce: agents that act continuously require approval rules, escalation paths and measurement that most marketing operating models, built around periodic campaigns, do not yet have.

Conclusion

Adobe and Salesforce now offer agent layers that are structurally equivalent: both expose the customer profile and its actions through MCP, both read warehouse data without copying it, and both position the CDP as the context source for agents. The remaining differences are in which customer record each vendor holds natively and in how each charges for agent activity. Salesforce is the lower-friction choice where the CRM is the operational record; Adobe is the lower-friction choice where digital experience and content are; where the warehouse holds the record, the agent layer becomes a replaceable consumer and the decision rests on cost and governance.

The first practical step is an inventory rather than a demonstration. Organisations evaluating either vendor should list the ten to fifteen attributes their agents would need to make a commitment to a customer, such as consent status, eligibility, open cases and lifetime value, and record which system masters each one today. The system that masters most of them indicates the lower-friction vendor. The evaluation should then test each vendor’s MCP and zero-copy access against the organisation’s own warehouse tables, and pilot one bounded agent use case with a measured outcome before committing to a suite.

Next step

A structured set of vendor questions shortens this evaluation considerably. The CDP data ownership and portability toolkit provides vendor questions and a migration-cost calculator for testing where each platform holds the customer record, and applies directly to an Adobe and Salesforce shortlist.

Frequently Asked Questions