What to act on now, what to pilot carefully, and what can wait until the technology matures.
If you watched the Dreamforce keynotes, you probably came away with two feelings. Excitement, because some of the demos were genuinely impressive. And a nagging worry that your competitors are already rolling all of it out.
Some of it is ready today. Agentforce Coworker, several of the new prebuilt agents and Multi-Agent Orchestration are all generally available. But many of the most interesting new capabilities Salesforce showed in San Francisco from September 15 to 17 are still in beta, pilot, or early rollout. That gives mid-market companies something rare in this industry: time to get the groundwork right before the tools mature.
We’ve been implementing Salesforce and Oracle systems since 1999, and we’ve sat through a lot of keynotes. Our advice after this one comes down to a single idea. Pick one business workflow, validate the data, permissions and integrations behind it, run one controlled AI pilot, and measure a business outcome before you scale.
Below is how that plays out if you run sales, service or IT at a company with a few hundred to a few thousand employees. What’s worth money this quarter, what deserves a small test, and what can comfortably wait until next year.
The big idea: Salesforce is moving beyond the Salesforce UI
Parker Harris, Salesforce’s co-founder and the person who built Lightning, asked the audience why anyone should need to log into Salesforce at all anymore. That question sums up the whole event.
The product behind it is AIforce. Salesforce describes AIforce as a live interface layer that brings Salesforce data, workflows, business logic, permissions and governance into the interfaces where people and agents already work, such as Slack, Teams, Claude and its own Agentforce agents. Agents work inside your existing permissions, so they can only see and do what the person using them can.
The rest of the announcements hang off that idea:
- Agentforce Coworker, an assistant built into the Lightning search bar. Generally available now.
- Salesforce in Claude and Slackforce. Salesforce in Claude is in beta, while Slackforce is expected to reach general availability in October 2026.
- Seven prebuilt “job-ready” agents with names like Casey (customer service), Piper (lead qualification), Paige (IT and HR requests) and Hunter (outbound sales). Casey, Piper and Paige are generally available. Hunter is in pilot, with general availability expected in November 2026.
- Koa, Salesforce’s first in-house AI model, built with NVIDIA specifically for CRM work. It’s in pilot, with general availability expected Winter 2026 in U.S. regions.
- Builder Central, for building apps and agents by describing them in plain English. Beta.
- Security Mesh, which gives you a unified view of security threats, is generally available. MCP server monitoring and risk scores, which rate the risk of outside tools your agents connect to, remain in beta.
One more thing is worth watching: pricing and consumption economics are still an important open question.
What we’d spend money on this quarter
Two things, and only two.
The first is Coworker. It needs no migration and runs on the data you already have, which makes it the cheapest way to find out whether your people will actually use an AI assistant day to day. Salesforce says 100,000 users switched it on in its first 35 days. Pick one group, say account managers heading into renewal season, and track how much prep time it saves them over a month. If the answer is “not much,” you’ve learned that cheaply too.
The second is a single job-ready agent, tied to a cost you already measure. If your support queue spikes after hours, look at Casey. If inbound leads sit untouched for a day before anyone calls, look at Piper. If a small internal team is buried in IT and HR tickets, look at Paige.
What we’d avoid is switching on several agents at once. When the numbers improve, you won’t know which one did it. Multi-Agent Orchestration, which coordinates a team of agents, is generally available too, but it only earns its keep once you have more than one agent doing real work.
What deserves a small test
Anything labelled beta belongs in a sandbox with a handful of users, not in front of the whole company. Features and terms can change before general release, and you don’t want your sales process depending on something that shifts under your feet.
If your teams practically live in Slack or Claude, these integrations are worth a pilot with 10 to 20 people. Salesforce in Claude is still in beta, and Slackforce is only just reaching general availability, so treat both as tests for now. Keep the test concrete. Can a sales manager get a pipeline-risk summary in under a minute, without asking ops to build a report? That’s a question with a yes or no answer, which is exactly what you want from a pilot.
Builder Central is worth a look if your admins are stretched thin. Start with one internal app and see how its built-in testing and deployment controls fit the way you already release changes.
And if you plan to connect agents to outside systems, bring your security team in now rather than later. Security Mesh is available today. The MCP risk scores for external connections are still in beta, but they’re precisely the kind of control an auditor will ask about, so it’s worth understanding them early.
What can wait
Koa is the announcement we find most interesting and are least in a hurry to buy.
The logic behind it is sound. Routine, well-defined tasks go to a cheaper model tuned for CRM work, and the harder judgment calls go to larger models like Claude or Gemini. If that holds up in practice, running agents at scale gets noticeably cheaper. But the accuracy claims so far come from Salesforce’s own benchmark, and there’s no price yet. We’d wait for customers outside the pilot program to report back.
Hunter sits in the same bucket. An agent that can pursue a goal for weeks, like reviving a stalled deal, is impressive on stage. It’s also the kind of autonomy you want firm rules around before it starts emailing your customers.
If your renewal is coming up, ask your Salesforce account team three things. How will AIforce, Koa and agent usage be priced? Which beta features are expected to reach general availability, and when? And what happens to your data when an agent acts from Claude, Slack or Teams? The answers will tell you a lot about how ready this really is.
The part nobody put on a keynote slide
Every demo on stage assumed clean, connected data. In most mid-market Salesforce orgs we look at, that isn’t the reality.
The same few problems come up again and again. The CRM can’t see what’s in the ERP, so an agent will happily quote a customer who hasn’t paid their last three invoices in Oracle or NetSuite. Duplicate accounts mean the agent works from whichever record it happens to find first. And permissions written years ago for people now also apply to software that moves far faster than people do. If a rep can see records they shouldn’t, so can the agent working on their behalf.
None of this is glamorous. All of it decides whether the new tools pay off. We went deeper on the data side in Why Data Cloud Is the Real Bottleneck in Your Agentforce Strategy.
A realistic plan for the rest of 2026
If we were running a mid-market Salesforce team, this is roughly how we’d spend the fourth quarter.
October: choose one workflow and audit the data, permissions and integrations behind it.
November: run one controlled pilot, with a person approving what the agent does.
December: measure the impact and decide whether to scale.
Whatever you measure, measure outcomes rather than logins. Hours of prep saved per rep, cases closed without escalation, and how quickly a new lead hears back from you all tell you far more than adoption dashboards do.
Where Selectiva fits
Most Salesforce partners stop at the CRM. We also implement Oracle E-Business Suite, Oracle Fusion and NetSuite, and we build the integrations between them on MuleSoft, Boomi, Oracle Integration Cloud and Informatica. That matters right now, because the gap between CRM and ERP data is usually the first thing that trips up an AI agent.
Selectiva can help you identify one practical AI use case, assess the data and integration gaps behind it, and decide whether it deserves a pilot now or later. Talk to our team about the workflow you have in mind.
