At Dreamforce 2026 in late September, Salesforce moved AI agents from experiments to everyday work: prebuilt agents for sales and service, tools to test and monitor them, and Salesforce data inside Slack and Claude. The keynote's message was about results, with Salesforce saying 30,000 customers are now live on Agentforce. The catch is simple: an agent can only be as good as the data and setup it runs on.
What Salesforce announced
| Announcement | What it does | Status, per Salesforce |
|---|---|---|
| Job-ready agents: Piper, Fin, Hunter | Prebuilt role agents. Piper handles inbound leads; Fin handles service cases | Most available now; Hunter in November |
| Agentforce Coworker | An AI teammate that answers questions across CRM and Slack and carries out plans | Available |
| Slackforce and Claudeforce | Salesforce data and workflows inside Slack and Claude | Slackforce available in October; Claudeforce in beta |
| Agent Optimizer | Finds where agents fail in production, suggests fixes and generates tests | Available in October |
| Agent Health Monitoring | Always-on reliability metrics with custom alerts | Available |
| Long-horizon agents | Agents that pursue goals over days, with memory and approval guardrails | Pilot; general availability planned for November 2026 |
| Koa | Salesforce's first CRM reasoning model | Pilot; expected Winter 2026 |
| Einstein Bot upgrade | Converts existing Einstein Bots into Agentforce service agents | Announced |
Throughout the keynote, Salesforce stressed human oversight: points where a person can step in, redirect or approve what an agent does.
Why most orgs aren't ready yet
An agent reads your records, follows your automations and acts with the access you give it. Duplicate accounts, dead fields, half-retired automations and out-of-date knowledge articles don't disappear when you add AI. The agent repeats them, faster and at scale.
A seven-point readiness check
- Clean the core records. Merge duplicate accounts, contacts and leads, and remove fields nobody fills in.
- List every automation. Document flows, legacy workflow rules and process builders, and retire the ones that overlap.
- Review access. Agents act with a user's permissions, so check profiles and permission sets before an agent inherits them.
- Fix your knowledge base. Service agents answer from your articles, so update or archive anything out of date.
- Decide the handover points. Write down when the agent must pass a lead or case to a named person.
- Test on real history. Run the agent against past cases or leads before customers see it.
- Agree on the numbers. Track resolution rate, handover rate and response time from day one.
Start with one agent, one queue
The lowest-risk start is a single job-ready agent on a single, well-defined queue, such as Fin on one type of service case, measured for 30 days. If the numbers hold, widen the scope. If they don't, the fix list usually points straight back to the data.
How Solinovo helps
We start with an org health check: duplicates, automations, permissions and knowledge, with a priced fix list. Then we set up and measure a first agent pilot. We've built AI support agents before, including one that now resolves 65% of a SaaS client's tickets without a human. Book an org health check.