Salesforce’s biggest annual event returns to San Francisco this week, and the Dreamforce 2026 story is already taking shape. In the days before the event, Salesforce has announced long-running Agentforce agents capable of pursuing goals over days and weeks, a new Enterprise AI Harness, Slackforce Surfaces and a growing Headless 360 architecture. Claudeforce, announced in August, has also given us a much clearer picture of how Salesforce intends to work with AI assistants outside its own applications.

Alongside these announcements, Data 360 is becoming increasingly important as the context behind AI, while documents, images and other unstructured information are becoming useful inputs for multimodal AI models. Together, these developments point towards something much bigger than another annual collection of product releases. Dreamforce 2026 may be where the architecture of the Agentic Enterprise really starts to become clear.

The Salesforce Innovations We'll be watching

1. Agentforce moves from AI assistant to AI worker

2. The CRM starts to run more of the process itself

Long-horizon agents also change what we can expect from CRM. Traditionally, CRM has depended on people moving the process forward. Someone:

• updates the opportunity;
• creates the follow-up task;
• checks whether information is missing;
• sends the next communication;
• decides what needs to happen next.

AI agents can increasingly perform some of that operational work.

Salesforce’s new portfolio of job-ready agents gives us some practical examples. Hunter can work an outbound sales process over time. Piper engages and qualifies inbound prospects. Casey handles customer-service work across channels. Marshall orchestrates back-office processes.

That suggests an evolution from: System of record → System of engagement → System of action, Employees establish objectives, handle important relationships and exceptions, apply judgement and approve consequential decisions. AI can increasingly perform operational work between those points.

The interesting question at Dreamforce is now how far Salesforce intends to extend that model across Customer 360 — and how much of an end-to-end process organisations will ultimately be comfortable assigning to agents.

3. Headless 360 and AIforce take Salesforce beyond the Salesforce UI

Headless 360 could be one of the most strategically important Salesforce developments of 2026. The idea is relatively simple: using Salesforce no longer has to mean opening Salesforce. Headless 360 makes Salesforce data, workflows, business logic and actions available as capabilities that authorised AI agents and applications can discover and use. Salesforce is expanding this architecture through MCP servers, APIs, reusable Agent Skills and a Headless Experience Layer.

Dreamforce has now given this architecture another name: AIforce. It brings Salesforce data, applications, business logic and Agentforce capabilities into the AI interface an employee chooses to work in. Its first experiences include Claudeforce, Slackforce and Agentforce Coworker, with the underlying Headless Toolkit providing access through MCP, APIs, plugins and Agent Skills.

This makes the Headless 360 direction much easier to picture. An employee could work in Claude or Slack while Salesforce continues to provide the customer data, permissions, business logic and actions underneath. For IT leaders, the choice of employee interface can increasingly be separated from the systems that hold trusted data and execute business processes.

4. Salesforce embraces a more open AI ecosystem

Headless 360 tells us something important about Salesforce’s wider AI strategy: Salesforce increasingly appears comfortable providing the trusted business platform underneath AI experiences it doesn’t completely own.

Claudeforce, the recently announced partnership between Salesforce and Anthropic, is a particularly strong example. Salesforce in Claude launches with 37 prebuilt sales skills, enabling Claude to reason over live Salesforce revenue context, update pipeline and perform governed Salesforce actions. Claude brings the reasoning interface. Salesforce provides customer data, workflows, business logic, permissions and actions underneath it. Salesforce is also making Claude available within its own ecosystem, including Agentforce, Agentforce Coworker and Slack.

Salesforce is also developing its own specialised AI models. At Dreamforce it introduced Koa, its first CRM reasoning model, built on NVIDIA Nemotron and designed for multi-step CRM work such as qualifying opportunities, routing cases and scheduling follow-ups. Koa sits alongside Salesforce's partnerships with external model providers, giving organisations another option for applying AI to Salesforce processes within the Salesforce trust boundary.

Businesses therefore may not ultimately choose one AI assistant for everything. Employees could use different interfaces and models for different kinds of work while Salesforce continues to provide customer data, business context, permissions, workflows and approved actions. This gives Salesforce a potentially powerful position in an increasingly open AI ecosystem: the trusted enterprise platform connecting data and business action across whichever authorised AI experiences an organisation chooses to use.

5. Slack becomes the work interface for humans and agents

Slack is also becoming far more important to Salesforce’s AI strategy.

Salesforce has now started using another name for this vision: Slackforce.

The newly announced Slackforce Surfaces takes the concept further. Employees can ask Slackbot to turn Salesforce and other connected information into interactive dashboards, reports, presentations and other working interfaces inside Slack.

This begins to change the role of the user interface itself.

Salesforce provides CRM data, workflows and business capabilities. Agentforce and other AI agents can reason and act. Slack provides an environment in which employees and agents can work together — and increasingly create the interface required for the task at hand.

Record channels can already bring Salesforce records into Slack. Slackbot can work with natural-language requests. Human approvals can remain part of the process. Headless capabilities allow actions to flow back into Salesforce.

A process that once looked like: Open Salesforce → Find record → Investigate → Perform action → Update colleagues; And can increasingly become: Discuss → Ask → Analyse → Approve → Act.

For organisations considering Microsoft Copilot, Slack, Claude and other AI employee interfaces, this raises an interesting architectural question: which interface should employees actually work in? Dreamforce should give us a much clearer picture of Salesforce’s answer. The process will be inside the workspace where the team is already collaborating. For organisations considering Microsoft Copilot, Slack, ChatGPT, Claude and other AI employee interfaces, Dreamforce's vision for Slack will be particularly thought-provoking.

6. Data 360 becomes the context layer for AI

7. Multi-modal AI makes documents, images and other data actionable

A huge amount of useful business information doesn’t sit neatly in CRM fields. It lives in PDFs, emails, photographs, reports, presentations, transcripts, forms and other unstructured content.

Multi-modal AI models can work across several types of information, allowing documents and images to become part of an AI-supported Salesforce process. Imagine a business receiving a PDF application together with photographs and supporting documentation. AI could potentially: Extract → Understand → Compare → Reason → Recommend → Act.

An agent could examine the material, compare it with Salesforce records, identify missing evidence, recommend a decision and prepare Salesforce updates for human approval. The same principle can apply to contracts, invoices, claims, customer correspondence, property information, inspection reports, healthcare documentation and many other document-heavy processes.

Salesforce document automation has traditionally concentrated heavily on the outbound journey: Salesforce data → Generate → Document. Multimodal AI adds a powerful inbound journey: Document / image → Understand → Structured data → Salesforce action.

This turns information that previously had to be manually read and interpreted into context that can potentially become part of an automated business process. For organisations sitting on large quantities of unstructured information, this could become one of the most practical applications of enterprise AI.

8. AI governance becomes part of the architecture

Giving AI access to business information is one level of responsibility. Giving an agent permission to change records, communicate with customers, progress opportunities or execute business processes raises the stakes considerably.

Salesforce’s newly announced Enterprise AI Harness gives this issue a much clearer architectural framework.

It brings together six areas:

Trusted Context → Trusted Agency → Trusted Action → Trusted Governance → Trusted Security → Trusted Models

Alongside these, Salesforce is introducing an AI Control Plane intended to give organisations a common place to discover and register agents, establish identity and policy, manage their lifecycle, evaluate performance, observe behaviour and outcomes, and control costs across Salesforce and third-party AI.

This becomes increasingly relevant as agents operate for longer, use different models and act across multiple systems.

Organisations will need clear guardrails. Some actions may be fully automated. Others may operate within thresholds. Higher-risk decisions can require human approval or escalation.

Monitoring will also extend beyond what an agent did to what AI is operating across the organisation, which models it is using, what it costs and how effectively it is performing. Salesforce’s Digital Wallet already provides visibility into Agentforce and Data 360 consumption; the AI Control Plane points towards much broader operational governance.

For many Salesforce customers, AI governance is likely to become an architectural discipline in its own right.

Further reading: agentforce and Data 360 credit consumption.

9. Specialised agents move into real business processes

Salesforce’s latest Agentforce announcement also gives us a clearer indication of how it intends to move from general-purpose agents towards agents designed for particular jobs. Its new portfolio includes agents for customer service, employee support, shopping, outbound sales, supply chain, inbound pipeline generation and customer operations. Each comes with skills, actions and data models for its particular role and can then be adapted to the organisation.

This approach becomes particularly interesting in industries such as financial services, healthcare, insurance and the public sector, where AI needs to operate within specialised processes, permissions, terminology and regulatory requirements. The ability to reason is only part of the implementation challenge. Businesses also need to determine what information an agent can use, what actions it can perform, which decisions need approval, how exceptions are handled and how its work is audited. As Agentforce moves into more substantial operational roles, those implementation decisions become increasingly significant.

10. Vibe coding changes how Salesforce itself is built

AI isn’t only changing how employees use Salesforce; it is also changing how Salesforce solutions are built.

Vibe coding uses natural language to describe what you want to create or change, with AI producing some of the implementation. Salesforce’s Agentforce Vibes brings this approach into Salesforce development, while Slack Code extends AI-assisted building into the collaborative workspace. This could gradually change the division of work between administrator, developer, architect and AI. Salesforce professionals can increasingly describe a required outcome and use AI to help create the solution.

It also creates a new risk. A Salesforce org is an interconnected architecture of data, automation, permissions, integrations, consent rules and business processes. A solution can work perfectly in isolation while duplicating an existing process, conflicting with another automation, bypassing established controls or creating data problems elsewhere.

Vibe coding potentially gives far more people the ability to build, including people who don’t understand the architecture they are building into. The result could be an uncomfortable version of the Peter Principle for AI development: users gain implementation capabilities faster than they develop the expertise required to manage their consequences.

Architecture, governance and oversight therefore become increasingly valuable. Organisations will need clear rules around who can create and deploy AI-generated solutions, how those solutions are tested and how they fit into the existing Salesforce architecture. Making something work is becoming easier. Understanding what should be built, what already exists and how a change affects the wider system remains a much harder discipline.

What does all this mean for Salesforce customers?

The announcements immediately before Dreamforce make the emerging Salesforce architecture considerably clearer:

• Employees and customers interact through Salesforce, Slack, websites and external AI assistants.
• Agentforce and other AI agents reason, plan, coordinate and execute work — increasingly over longer periods.
• Data 360 supplies trusted enterprise context.
• Multimodal AI brings documents, images and other unstructured information into that context.
• Headless 360 makes Salesforce data and business capabilities available beyond the traditional Salesforce interface.
• The Enterprise AI Harness provides common context, agency, actions, governance, security and model choice.
• The Salesforce Platform remains underneath as the system for CRM data, workflows, permissions and business action.

For Salesforce customers, the opportunity extends well beyond adopting Agentforce.

Organisations need a clear view of where AI fits into their operating model: the data it can access, the processes it can execute, the interfaces employees will use, the models and agents they will allow, and the decisions that require human judgement. Salesforce’s announcements suggest that its ambition is to provide the trusted architecture connecting all of those pieces.ll use, and the decisions that require human judgement. Getting those foundations right will determine how successfully AI works across the business.

Dreamforce 2026 dates, location and Dreamfest

What we'll be watching at Dreamforce 2026

Several things we expected Salesforce to emphasise at Dreamforce have now been confirmed before the event even begins.

Long-running Agentforce agents are real. Multi-agent orchestration is here. Headless 360 is expanding. Salesforce has put a formal architecture around enterprise AI through its Enterprise AI Harness. Claudeforce demonstrates how Salesforce can sit underneath external AI interfaces. Slackforce is beginning to show what a dynamically generated employee interface could look like.

Dreamforce itself now needs to show us how these pieces work together in practice.

That may ultimately be the defining Dreamforce 2026 story: moving on from proving that AI agents can work to showing how businesses can put them to work at scale — across data, applications, processes and employee interfaces.

We’ll update this article as Dreamforce 2026 unfolds, adding the major announcements, what Salesforce actually launches and what those developments mean in practice for Salesforce customers. --

Last Update: Wednesday 16 September 2026.