Dreamforce 2026: what's new in AI agents and what it means for your company

Over the past few years, plenty of companies rolled out assistants that could summarize information, draft text or answer questions. Dreamforce 2026 pushed the conversation onto more operational ground. An enterprise agent has to know the customer, understand internal policy, work with approved tools and log the outcome of what it does.

That raises the bar. You can generate a useful answer from limited information. Taking a correct action is harder: it needs identity, permissions, current data, business logic and a way to supervise the result. That is why the announcements paired new AI capabilities with a control architecture and cross-platform integrations.

AIforce and the CRM beyond the traditional interface

AIforce was the headline. Salesforce frames it as an interface layer that carries the business knowledge sitting inside Salesforce out to the places where people and agents actually work. That knowledge covers data, flows, rules, permissions, security and the semantics of the business.

In practice, someone could look up an account, update a record or kick off a flow from Slack or Claude without touching the usual CRM screens. The action still runs on the rules and permissions defined in Salesforce. The interface changes; the system keeps the context it needs to get the work done.

For companies, the payoff is less friction. The CRM usually has the information already. Adoption breaks down when people have to leave what they are doing, hunt for a screen and fill in fields that feel like paperwork. Bringing the information and the actions to where the work happens can lift adoption, provided the process is well designed.

Specialized agents for specific functions

Salesforce expanded Agentforce with agents built for specific jobs. The lineup covers customer service, IT and HR support, commerce and sales development. Each agent ships with the actions and data structures tied to its role, and you can then tune it to the way your company works.

That cuts the need to build every use case from scratch. A preconfigured agent still won't do the implementation for you. You have to define the data sources, the escalation rules, the exceptions, the owners and the metrics. Early speed only pays off if you keep control of the process.

When you pick a first use case, favor tasks with real volume, clear rules and an outcome you can see. Triaging requests, answering common questions or qualifying inbound leads is a far more manageable starting point than handing an agent an ambiguous commercial decision.

Koa and reasoning applied to the CRM

Koa is Salesforce's first reasoning model built specifically for CRM, developed on NVIDIA Nemotron. Salesforce says it was post-trained on synthetic scenarios that mirror business processes across industries, and that no customer data went into its training.

The point of Koa is multi-step work. An agent might check the state of an opportunity, look back at earlier interactions, apply a rule, update the CRM and schedule a next step. Writing a convincing sentence is easy. Choosing and running the right sequence is the hard part.

Koa is in limited pilots inside Agentforce, with general availability slated for winter 2026 in US regions. Companies in Latin America should confirm availability, data residency and pricing before they put it on a roadmap.

Slackforce, Claudeforce and Agentforce Coworker

AIforce shows up first through three experiences. Slackforce puts Salesforce data and actions inside Slack conversations and channels. Claudeforce wires Salesforce into Claude, with ready-made skills for commercial tasks. Agentforce Coworker lives inside the Salesforce interface and can orchestrate specialized agents the organization has already deployed.

All three run on the same idea: work in context. A sales manager could spot inactive accounts from Slack, review the related service cases and assign a follow-up. A rep could prep for a meeting from a chat interface backed by CRM data. The upside comes from fewer tool switches and having the right information on hand at the moment of the decision.

It also raises the stakes on permissions. As the CRM shows up in more interfaces, the company has to make sure every user and agent sees only what they are cleared to see, and that every action is logged. Easier access can't come at the cost of traceability.

Governance and control to scale agents

Salesforce introduced its Trusted Enterprise AI Harness, an architecture that pulls together context, action, governance, security and model selection. It also announced an AI Control Plane to discover and register agents, manage identities and policies, watch how agents behave, grade their results and keep costs in check. Some of the underlying technology is available now; the newer capabilities and a single unified experience start rolling out during Salesforce's fiscal 2028.

This direction matters because risk climbs with every extra agent and system in the mix. You can babysit a single pilot by hand. An operation with agents spread across sales, service and admin needs an inventory, consistent permissions, activity logs, evaluations and spending limits.

Governance can't be an afterthought. Before an agent goes live, the organization has to decide what it is allowed to decide, which actions need a human sign-off, what data it can touch and how a mistake gets handled.

A more open ecosystem

Dreamforce 2026 also laid out an interoperability play. Salesforce deepened its partnerships with AWS and Google Cloud to connect data, agents and infrastructure. With Google Cloud, the idea is for Salesforce agents and Gemini Enterprise to reason and act on shared ground. With AWS, the integrations aim to carry CRM context into work tools, connect agents to Slack and enable voice.

AgentExchange rounds this out as a marketplace to find and deploy agents, actions, integrations and skills built by partners. For customers, that openness can cut down on one-off development and widen their technology choices. It also raises the bar on vetting each component for ownership, compatibility, maintenance and security.

Cases that show the path to production

The customer stories from Dreamforce help show where this is landing in practice. Salesforce said Live Nation uses an agent to field common questions across 120 venue sites, with the potential to handle more than 300,000 queries a year. Siemens described Agentforce helping triage and qualify over 2,500 inbound leads a month for a sales organization of 18,000 people.

These numbers come from Salesforce and the companies themselves, so read them as specific cases rather than guaranteed results. What they show is worth noting: high interaction volume and a process where you can measure coverage, handoff and outcome.

Implications for companies in Latin America

The announcements open doors, and they also expose a few gaps that are common across the region.

Data turns into operating infrastructure. Duplicate records, half-empty fields and disconnected systems no longer just skew your reports; they blunt an agent's ability to read context and act correctly.

Automation needs explicit processes. If every team handles an exception its own way, the agent has no stable rule to follow. Writing down decisions, owners and escalation criteria becomes part of the AI project itself.

Adoption needs new habits. People have to learn what they can delegate, how to check a result and when to correct or escalate. Training has to blend how the tool works with real business judgment.

Return needs a baseline. Before the pilot, measure your times, volumes, costs, conversions or service levels. Without that reference point, you won't know whether the agent improved the process or just added another layer of technology.

Regulation and availability vary. Companies have to check the terms by country, sector and cloud region. A global announcement doesn't mean the feature is live in every market.

The next step

Dreamforce 2026 showed a platform built to fold agents into everyday work. AIforce widens where you can use Salesforce; Agentforce adds more specialized agents; Koa brings reasoning to CRM processes; and the Enterprise AI Harness is meant to keep control as autonomy grows.

For business leaders, the smart next move is to pick one process that matters, get its data ready, agree on the limits and measure the result, instead of chasing every new feature. That discipline is what separates a slick demo from an operational gain that lasts.

Is your organization weighing AI agents on Salesforce? Before you build, it pays to pick the right process, check your data quality and set a governance model, which is exactly what our AI agents service at GoCode focuses on. A maturity assessment can turn interest in AI into a roadmap with clear priorities, owners and metrics.