Salesforce after Dreamforce 2026: what changes for midsize firms
Dreamforce 2026 brought Koa, long-horizon agents, Salesforce in Claude and consumption billing. For a midsize company, what matters: data, permissions, cost.
Leer en español →Dreamforce 2026, held September 15 to 17 in San Francisco, made Salesforce’s direction clear: a CRM used less and less through screens and more and more through agents, whether inside Salesforce, in Slack or in Claude. Five announcements carried the most weight: Koa, Salesforce’s own reasoning model for CRM tasks; long-horizon agents that can pursue a goal over days or weeks; orchestration across multiple agents; Salesforce inside Claude, the result of the Anthropic partnership called Claudeforce; and an official Salesforce plugin for Claude Code. The last two had been announced a few weeks earlier.
For a midsize Spanish company, the useful question isn’t which of these is most impressive but three more prosaic ones: what is actually available and in which regions, how it’s paid for, and whether your data and permissions are in good enough shape for an agent to work with them without causing problems. If the answer to the third is no, it isn’t worth it yet, however good the other two answers are.
A note on timing: this analysis reflects what Salesforce had published by mid-October 2026. Several pieces had general availability planned for November or the winter, so check their current status before you decide.
What was announced and where it stood
| Announcement | What it is | Announced status (October 2026) |
|---|---|---|
| Koa | CRM reasoning model built on NVIDIA Nemotron, trained on synthetic scenarios | Pilot; general availability expected in winter 2026 in U.S. regions |
| Long-horizon agents | Pursue a goal over days or weeks, with memory and approvals based on your rules; the first is Hunter, for outbound sales | Pilot; general availability expected in November 2026 |
| Multi-agent orchestration | A primary agent splits work among specialized agents, including third-party ones through the A2A protocol | No date in the official recap; confirm which parts are live |
| Salesforce in Claude | Plugin with 37 sales skills that works with each seller’s permissions | Beta since September 15 on paid Claude plans, with prior approval |
| Claude Code plugin | Skills, MCP servers, specialized agents, hooks and commands for building on Salesforce | Available in Claude’s official plugin marketplace |
| Agentforce Coworker | Autonomous assistant inside Salesforce that finds answers and executes plans | Available |
What matters for a midsize Spanish company
Language and voice. Agentforce Voice added six languages, including Spanish, German and Italian. For a Canary Islands tourism business handling calls in several languages, that detail matters more than any benchmark.
Koa, with caveats. Salesforce says that on its own benchmark Koa is 11% more precise at choosing the right action than today’s default general-purpose models; that’s the vendor’s own measurement. Two details matter more: Salesforce controls the model weights and runs it on its own infrastructure, without using customer data to train it, and the announced general availability is for U.S. regions. If you need everything processed in Europe, confirm it before counting on Koa. Agentforce offers other models, including Claude, which according to Salesforce runs inside its trust boundary through Amazon Bedrock.
Long-horizon agents. They’re the most interesting launch for sales teams: you give the agent a goal such as “re-engage my at-risk deals” and it analyzes the data, proposes a dated plan and executes it, asking for approval according to the rules you set. They also demand the most preparation: Hunter, for example, asks for access to your email so it can write on your behalf, and its plan will only be as good as your opportunity data.
Salesforce inside Claude. If your sales team already uses Claude, the plugin lets them prep meetings, review the pipeline or update opportunities without opening Salesforce, under their own permissions and, by default, asking for approval before each change. Watch the requirements: Claude’s help center calls for a recent Sales Cloud enterprise edition and Salesforce’s approval for the beta, and you’ll be paying for both Claude and Salesforce.
The Claude Code plugin. It’s a tool for whoever builds or customizes Salesforce: it bundles about 40 development skills, three MCP servers, specialized agents and automated checks into a single install. It makes customization cheaper, but it doesn’t replace human review or a disciplined deployment process.
How you pay: licenses plus consumption
Salesforce’s model is no longer just a price per user. There are three layers worth understanding separately.
Editions. Since September 2026, Salesforce’s Sales Cloud price list in Spain offers Core at €195, Advanced at €395 and Max at €550 per user per month, billed annually. They include Agentforce, Slack, Tableau Next and Premier support, plus an annual pool of Flex Credits per org: 500,000 in Core, one million in Advanced and 2.75 million in Max. Starter Suite (€25) and Pro Suite (€100) remain available, although the pricing page itself says AI can be added from Core upward. Salesforce says existing customers keep the pricing of their legacy editions.
Agentforce consumption. According to its pricing page as of October 2026, Flex Credits cost €500 per 100,000; in Salesforce’s own examples, each action consumes 20 credits, or €0.10. For customer-facing agents there’s an alternative of €2 per conversation, which can’t be combined with Flex Credits in the same org. There are add-ons from €125 per user per month with unmetered use for employees, and a €5 per user per month license to give every employee access to agents, which requires Flex Credits. Unused credits don’t roll over, overages are billed monthly in arrears and Digital Wallet lets you track consumption with alerts.
External agent access. Salesforce published on September 17 that, with a November target, every external agent will have to be registered with its own identity, on timelines that depend on the type of connection and your renewal, and that each call it makes to the platform over MCP or API will count as an interaction that consumes Flex Credits. Rates had yet to be published, with 30 days’ notice promised before billing starts. Sandboxes aren’t metered, and traditional integrations keep their current pricing.
An illustrative example of why agent design matters: 2,000 customer inquiries a month cost €4,000 with per-conversation pricing. With Flex Credits, if each inquiry is resolved in three actions, it would come to €600; if it takes ten, €2,000. These are calculations with published prices, excluding Data 360 credits and other consumption, and they show that a poorly designed agent can multiply the bill.
Readiness: data, permissions and a narrow pilot
- Clean the data the agent will read. Sales stages that mean the same thing to everyone, realistic close dates, deduplicated contacts and up-to-date knowledge articles. An agent reads the same fields as your reports: if they’re wrong, so is the agent.
- Review permissions. The Claude plugin acts as the signed-in user, so an over-permissioned user becomes an over-permissioned agent. Use agents’ own identities to give them only what they need.
- Pilot in a sandbox and measure how many actions each real case consumes before moving to production.
- Keep human approvals at first and relax them one workflow at a time, based on data.
- Define a metric before you start: cases resolved without human intervention, hours saved per rep or revenue recovered.
- Name an owner for connections and consumption, and set up alerts from day one.
Don’t anchor your decisions to product names, which change often: what counts is what you sign, meaning products, credits and terms.
When it isn’t worth it yet
- If you’re on Starter or Pro Suite with a small team and basic CRM use. Moving from Pro Suite (€100) to Core (€195) nearly doubles the per-user cost; you need a use case that pays for it.
- If your data is messy or your sales process isn’t written down. An agent speeds up whatever is there, including the mess.
- If you rely on features that, for now, only reach U.S. regions and you need processing to happen in Europe.
- If you can’t estimate consumption. While external agent access rates remain unpublished, any budget is provisional.
- If the task is deterministic. Many so-called agent tasks are fixed-rule workflows that are better and cheaper to handle with Flow, n8n or a lightweight custom build.
It isn’t a fringe position: according to an Oppenheimer survey of customers at Dreamforce 2026, reported by Salesforce Ben, 85% had tested or deployed Agentforce, but about half were still in testing or pilot.
At Conceptual Creative we work with Salesforce, HubSpot and Zoho, and we can help you review data and permissions, design a pilot with a clear metric or decide whether an automation without agents would serve you better. See how we approach custom CRM and applied artificial intelligence.
