Salesforce and AWS Take Enterprise AI Beyond the System of Record

Key Highlights

• Salesforce and AWS have deepened their collaboration to bring Salesforce context into Amazon Quick and AWS agents into Slack.
• The integration delivers zero-copy data access to additional AWS data services without requiring large-scale data migration.
• Agentforce users can now tap into Amazon Bedrock to access models from Anthropic and NVIDIA, with OpenAI models expected to follow.
• Planned real-time bidirectional voice capabilities connect Agentforce Voice with Amazon Connect in fall 2026.
• The announcement signals a broader shift toward decoupled intelligence, where enterprise context becomes available across more of the applications employees use.

The News

Salesforce and Amazon Web Services announced a significant expansion of their strategic collaboration on September 15, 2026. The integrations bring Salesforce data and context into Amazon Quick and make the AWS DevOps Agent available in Slack, with additional AWS agents planned for fall 2026. The partnership also expands Data 360 zero-copy access to AWS data services and gives Agentforce customers access to additional models through Amazon Bedrock. Read the press release here.

Analyst Take

We have been analyzing the evolution of enterprise software for the better part of a decade, and for most of that time, the operating assumption was simple: the system of record was the interface. Your CRM was where you lived. This announcement from Salesforce and AWS reflects a change in that model. Work happens across too many tools for any single application to own the experience.

Our discussions with the Amazon Quick team have helped us understand where AWS intends to take the product, and this integration aligns with the roadmap its product leaders shared with us. The integration is intended to make Salesforce context available wherever the user is already working, rather than requiring the user to return to Salesforce. That raises the more interesting question of who owns the screen. Salesforce may continue to own the customer data and business logic, but users may increasingly reach that context through Quick, Slack, or another AI interface. The system of record still matters, but it may no longer determine how employees interact with its data.

That shift also makes governance more complicated. Moving an agent into Slack does not resolve the harder questions about what it can see and do. Enterprises still need consistent identity, permissions, approval requirements, and audit trails across Slack, AWS, and the underlying systems the agent can change. The value is not simply keeping engineers in one interface. It is preserving those controls as the agent crosses application boundaries.

The same distinction applies to model choice. Model access is not the same as model portability. Enterprises need to know whether they can change models without rebuilding prompts, evaluations, security policies, and workflow logic around each one. Bedrock gives Agentforce customers more options, but the practical value will depend on how consistently Salesforce exposes and governs those models inside an agentic workflow.

What Was Announced

The announcement covers four areas: embedded workflows, data access, model choice, and voice.

Enterprise users can now access Salesforce data, context, and skills from Amazon Quick through an integration built on MCP and Salesforce’s headless architecture. Separately, Salesforce expanded Data 360 zero-copy access to AWS Glue-managed Apache Iceberg tables, S3-backed Apache Iceberg tables, Amazon Aurora, Amazon RDS, and SageMaker Lakehouse. The companies say these connections allow customers to ground agents in enterprise data without migrating or duplicating it while maintaining the governance controls established in each environment.

Zero-copy access reduces duplication, but it does not eliminate the underlying integration problem. Enterprises will still need to test query latency, semantic consistency, permission enforcement, and data freshness across both environments. The value will depend on whether those controls continue to work against the complexity of an actual enterprise data estate.

Agents embedded in the conversation stream. The AWS DevOps Agent is now available in Slack. AWS plans to add its Security, FinOps, and Partner Central agents in fall 2026. The plan is to expand to the AWS Security Agent, AWS FinOps Agent, and AWS Partner Central agents. These agents carry conversational context within channels, so a technical team investigating an incident can stay in the thread rather than switching to a separate console.

Bringing operational agents into Slack could help engineering teams investigate incidents without moving repeatedly between conversations and AWS consoles. Adoption will depend on whether the agents surface useful information without overwhelming channels with alerts, diagnostics, and suggested actions.

Agentforce customers can now access Anthropic and NVIDIA models through Amazon Bedrock, with OpenAI models expected to follow. Salesforce and AWS say Bedrock applies zero data retention and does not allow model providers to use customer data for training. The Salesforce Trust Layer adds capabilities including dynamic grounding, toxicity detection, and audit trails. The companies position the combined offering for workloads subject to requirements associated with HIPAA, PCI, SOC 2, and ISO 42001.

HyperFRAME Research found that 79% of organizations plan to use multiple models. The Salesforce and AWS integration reflects that direction by giving customers more options within Agentforce. It also adds operational work: enterprises must evaluate models for quality, cost, latency, security, and availability, then continue monitoring them as providers release updates. Regulated industries may benefit from the additional choice, but they will also need stronger model approval and lifecycle-management processes.

Salesforce and AWS plan to connect Agentforce Voice and Amazon Connect Customer through real-time bidirectional audio over WebSockets using A2A support in fall 2026. Separately, voice dictation for Slack is available now, using Salesforce speech models optimized on AWS Trainium chips.

Voice remains the least mature piece of the puzzle. Real-time bidirectional audio between an AI agent and a human caller, routed through two distinct platforms, is technically ambitious. Latency tolerances for voice are far tighter than for text-based interactions. We will reserve judgement until we see production benchmarks.

Looking Ahead

The broader pattern here is clear enough: enterprise intelligence is becoming less tied to a single application. MCP is emerging as one way to make data, context, and tools available across platform boundaries. Vendors that rely on monolithic ecosystems to capture value, keeping data and context locked inside their own walls, will face increasing pressure from composable, cross-platform alternatives.

That said, the gap between an announcement and sustained adoption is wide. We will be tracking three things in the coming quarters. First, real-world latency and governance performance of the zero-copy data strategy at scale, particularly for organizations operating across multiple regulatory jurisdictions. Second, actual usage metrics for the embedded Slack agents: are engineering teams adopting them in daily workflows, or do they become another notification source that gets muted? Third, the maturation of the voice capabilities, where AWS Trainium-optimized speech processing will need to demonstrate that it can meet enterprise latency and reliability requirements at volume.

The direction is right. The execution will determine whether this is a genuine inflection point or another well-engineered partnership announcement that underdelivers in practice. We have seen both outcomes before.