Salesforce has published customer accounts that put some numbers and operating rules behind agentic AI deployments. The useful message is less “switch it on” than “measure what it does, listen to customers and keep controls close”.
Salesforce Watch analysis
What happened
In a Salesforce article published on 1 October, customers including Canada Goose, AT&T, Southwest Airlines and Sammons Financial Group describe how they are using AI agents and what they learned along the way. Canada Goose says Agentforce autonomously resolves 89% of routine messaging enquiries and 15% of calls, freeing its Style Experts to focus on personal shopping and more tailored customer interactions.
The examples also show different approaches to oversight. AT&T says it built a trust layer and uses its own large language model for work involving proprietary intellectual property, while allowing agents to be turned on or off. Sammons Financial Group says it ran more than 200 guardrails and tests before launch, prevents its model from learning from live conversations and uses a supervisor agent to monitor calls. Read Salesforce’s customer accounts.
Why it matters
These accounts give businesses a more useful starting point than the usual choice between “AI will replace everyone” and “AI will make everything delightful”. In these examples, the work being handed to agents is bounded, customer feedback helps shape what comes next, and oversight is part of the deployment rather than an afterthought.
The reported numbers and practices come from Salesforce’s customer feature, not independently audited comparisons. Still, they make the operational questions concrete: which enquiries are routine enough to automate, what happens when an agent gets it wrong, and who can stop it?
Our read
The most valuable lesson here is not that every company needs an agent. It is that a deployment needs a job description, a way to learn from actual users and controls that match the stakes. A kill switch is not glamorous, but it beats discovering that the only off switch is a meeting invite.
Treat the customer figures as examples, not forecasts for your own service desk. Start with a narrow task, decide what success means before launch and work out how a person can step in when the conversation leaves the script.
What to watch
- Whether the customers publish comparable results over time, including service quality as well as automation rates.
- How companies decide which interactions stay with staff and which agents may handle autonomously.
- Whether customer feedback leads to visible changes in agent scope and controls.
Discussion spark: Should customer-service agents be allowed to resolve routine enquiries autonomously once they pass a company’s tests, or should a person review every AI-handled case?
Sources and evidence
- Getting to ROI: How Brands Turn Cost Centers into Revenue Engines with AI (1 October 2026, 15:00 UTC)
not affiliated with or endorsed by Salesforce