Dreamforce 2026 – San Francisco
Many agents, many models, one priority: reliable data and context
What I took away from Dreamforce 2026 about enterprise AI, agents, and architecture
After three days at Dreamforce 2026 in San Francisco, one thing stood out to me more than any individual announcement: the conversation around enterprise AI has changed very quickly.
Beyond the keynotes and product announcements, Dreamforce also had a very personal side for me: reconnecting with former colleagues and partners from earlier stages of my career. Reconnecting, exchanging perspectives, and seeing how everyone has evolved professionally is also part of what makes events like this valuable.
Not long ago, most conversations were centered on copilots, chatbots, and content generation. Now the focus is shifting toward something much more complex:
how to give AI a real understanding of our business, how to let it act on real processes, and how to do so with control, security, and measurable results.
Over these three days, I attended the Main Keynote, Enterprise AI Harness, Life Sciences, the conversation between Marc Benioff and Sam Altman, and the Sales, Agentforce, Service, Commerce, and Marketing keynotes.
These are the main ideas I am taking away.
1. AI is moving from answering to executing
This was probably the most consistent message throughout Dreamforce.
Until now, we have mainly used AI to ask questions, summarize information, generate content, or support decisions.
The next phase is different: we want AI to do things.
Query systems, retrieve information, execute workflows, update data, or trigger actions.
For years, we have worked roughly like this:
User → application → menu → screen → action
Increasingly, we will see:
User → intent → agent → multiple systems → outcome
This changes the question technology leaders need to ask.
It is no longer just:
“Which AI assistant should we use?”
It becomes:
“Which business processes are we actually ready to let an agent execute?”
2. From applications to agents: Salesforce and Microsoft are moving in the same direction
Another idea that caught my attention was the shift toward increasingly headless architectures. And importantly, this is not something happening only inside Salesforce.
Salesforce is moving toward a model where users can work from environments such as Slack while agents access capabilities from Sales, Service, Commerce, and other systems behind the scenes.
At Dreamforce, I could already see this pattern in practice, with users interacting from Slack while agents accessed data and capabilities from systems behind the scenes.
Microsoft is moving in a very similar direction through Teams and Microsoft 365 Copilot.
Two different ecosystems are converging toward the same underlying pattern:
User → Slack / Teams / Copilot → agents → ERP, CRM, ecommerce, and data
The fact that vendors with different platforms and starting points are moving toward the same model feels like a strong signal of where enterprise software may be heading.
Applications will not disappear. We will still need ERP, CRM, ecommerce, and data platforms as transactional systems and sources of truth.
But they may stop being the primary interface for many tasks.
Users may no longer need to know which application contains a particular piece of information or functionality. They may simply express what they need and let agents coordinate the systems behind the scenes.
The important change is not that applications disappear, but that our interaction with them may no longer need to be direct.
That could fundamentally change how we think about enterprise architecture, regardless of which vendor we choose.
3. But no agent understands your company by itself
Within the Salesforce ecosystem, we will have more and more specialized agents for Sales, Service, Commerce, Marketing, and other business processes.
We will also have access to multiple models with increasingly powerful capabilities.
But having the agent or the model does not solve the problem by itself, because models do not know our company.
They do not know our customers, products, processes, rules, priorities, or exceptions. That knowledge lives inside our systems and inside the organization itself.
For an agent to genuinely help us, it needs to understand things such as who the customer is, what they have purchased, which commercial conditions apply, what inventory is available, what happened previously, which rules apply, and which actions it is authorized to perform.
The real challenge is giving agents reliable data and the right context so they can understand the business and act correctly.
That makes master data, integration, data quality, and governance fundamental parts of AI infrastructure.
And the more agents we deploy, the more important it becomes that they work with consistent data, share context, and follow the same rules.
4. CRM wants to stop recording work and start doing it
This was particularly clear during the Sales Keynote.
For years, one of the biggest challenges with CRM has been obvious: sales teams need to continuously enter information to keep the system updated.
The direction now is almost the opposite.
The system should be able to capture signals, maintain context, understand what is happening with the customer, and support the salesperson without requiring every interaction to be manually recorded.
If this evolution works, CRM can progressively move from being mainly a system of record to becoming a system of action as well.
And that is a much deeper change than simply adding a chatbot inside the CRM.
5. Agentforce now has to prove ROI
This was something I particularly liked about the Agentforce Keynote.
The conversation is shifting from impressive demos to measurable business value.
The key questions are becoming:
What problem does it solve? What process does it improve? What value does it create?
We should talk less about “AI use cases” and more about:
“business processes we want to improve using AI.”
The strongest examples I saw at Dreamforce followed that pattern: specialized agents, embedded in specific processes, with clear objectives and boundaries.
As agents become capable of taking more actions, we will also need to rethink processes, permissions, and responsibilities around them.
6. Sales, Service, Commerce, and Marketing form one customer journey
On the second day, I attended the Sales, Agentforce, Service, and Commerce keynotes back-to-back, followed by Marketing on the third day.
Taken together, the conclusion is simple.
Customers do not experience Sales Cloud, Service Cloud, Commerce Cloud, or Marketing Cloud.
They experience one relationship with the brand:
discover → ask → buy → receive → get support → buy again
For agents to support that whole journey, they need to share data and context.
Without connected data, there is no connected intelligence.
Models will continue to improve, and we will have more and more agents available from an increasing number of providers.
But none of them comes with our company built in.
Reliable data, the right context, and clear rules for taking action will remain our responsibility, not the model’s.
The question I am taking home is not which agent or model to choose next.
It is simpler, and more uncomfortable:
Of the processes that today still depend on a person looking at a screen, which ones are we actually ready to entrust to an agent?
