Salesforce's own list of Dreamforce '26 announcements runs to 21 items, and they are not all at the same stage. Agentforce Coworker is generally available now. Slackforce was slated for general availability in October. Salesforce in Claude, the centrepiece of the partnership with Anthropic, is in beta. Koa, Salesforce's first CRM reasoning model, is a pilot, with general availability expected in Winter 2026, and long-horizon agents are a pilot until November. Salesforce also notes that availability may vary by region. [S1, S2]

That is the 20%: the product. The rest of this piece is about the other 80%, through the three questions I hear most from Australian teams.

The headlines give people reason to ask. Dreamforce opened amid calls for curbs on AI development, and the people on Salesforce's stage did not dodge it. Sam Altman said people are right to be afraid of where AI is heading. Dario Amodei repeated his concerns about the risks. [S6]

There are two ways to look at that problem. The AI in those headlines is a frontier model in a lab test rig, with the internet wide open. The AI in your business is a different animal: it has a job, set permissions, and an audit trail. It serves customers, drafts, routes, and resolves. One is a research lab. The other is closer to a hospital, where every action is logged and somebody is accountable for it.

So respect the risk, then build with the second kind. Pick a goal you can name, and build the guardrails around that goal before you build the agent.

Good choice of model. I would pick it too. But look at who shared Salesforce's stage in San Francisco: Sam Altman, Dario Amodei, and Jensen Huang. [S6] Anthropic could build its own CRM. It partners with Salesforce instead. Worth asking why.
My read is that the model is becoming the part you rent. Bain's Technology Report 2026 points to price pressure on mature model tokens [S7], and a better model lands every few months. If everyone can rent the same model, the model stops being the advantage. What nobody can rent is your customer record: the data, the business rules, the permissions, and the audit trail that shows who did what.

That gives the build question three parts. Build the agent that makes you different. Rent the model, and switch it as models get cheaper and better. Buy the record, the rules, and the audit trail.

Then look below the waterline. The build case counts the app, the model, and the cancelled licence. Running it adds servers, access control, security, updates, and engineers who now spend their time reviewing what the AI wrote. Every token you spend building, you spend again maintaining, so spend it on what gives you a unique moat.

Great. Who hired them, and who manages them?
In Australia that stops being a rhetorical question on 10 December. From that date, an organisation's privacy policy must disclose the kinds of personal information it uses in automated decisions, and the kinds of decisions a computer program makes, or substantially helps make, where they could significantly affect someone's rights or interests. [S8] You cannot disclose a decision you cannot see.

The Dreamforce launches help here, by design. Salesforce says every request runs on your existing permissions and business rules, so an agent only sees what the person asking can see. [S2] That is the right design. It also means the access model somebody configured years ago is now part of your AI governance, and soon part of your privacy disclosure. Carlos Langle, our Director of Data & Integration, puts it plainly: if you want an agent tightly scoped, you do not narrow the scopes, you narrow the user.

So onboard every agent like a new hire. Run an agent census, and give each one four things: a job, an owner, an access level, and an off switch. The organisations getting a return already work this way. KPMG's Q3 2026 Global AI Pulse found that 86 percent of organisations with established ROI run a formal cross-functional or enterprise-wide AI management layer, against 31 percent of those still experimenting. [S9]

Every failed AI pilot I have seen died on the same thing, and it was never the model. The model is roughly 20% of the work. The other 80% is the delivery discipline around it: the data, the evals, the governance, and the person who owns it after go-live.

Bain puts a ratio on it: for every dollar spent on technology, four are spent on people and process. [S7] Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. [S4] None of those causes appear in a demo.

If you cannot name the human, you are not in production. You are in a long pilot with good lighting.

J4RVIS works with both Salesforce and Anthropic, so we see this partnership from both sides. We also run what we sell on ourselves first. Nothing goes to a client until it has run on us, including the launches we rate most highly.

Our team will be at World Tour Essentials in Sydney on 20 October and Melbourne on 27 October, and at the Salesforce Summit in Brisbane on 5 November.