Salesforce’s write-up of customer sessions at Dreamforce 2026 puts one idea first: AI agents can turn cost centers into value centers. The brands it quotes, from Canada Goose to Canon, describe service teams that field fewer routine questions and sell more, with controls written in before launch.
What the brands told Salesforce
Salesforce published the account on October 1 under the headline “How Brands Turn Cost Centers into Revenue Engines with AI.” It names AT&T, Southwest, Canada Goose and others, and it reports what each said about running Agentforce in production.
Canada Goose gives the most concrete numbers. According to Salesforce, Agentforce autonomously resolves 89% of the company’s routine messaging inquiries and 15% of its calls. The questions it handles are the ones about when a jacket will arrive or how to swap a sweater for another size. The people who used to answer them now work as “Style Experts” on personalization, personal shopping and what Salesforce calls moments of connection.
The revenue claim comes from Dennis Liut, Head of Global Customer Experience and Revenue at Canada Goose. “And now we have an Experience Center that used to be a call center that is bringing in millions in revenue,” he said at Dreamforce 2026. Salesforce’s write-up does not say how that revenue is measured or what share comes from the agent-handled work, so the figure is best read as the company’s own framing.
Canon is also moving its service organization from a cost center to a value center, Salesforce reports. Bill Duval, VP of Service Information Management and Technology at Canon, described capturing what customers say in service interactions and feeding it into product development.
AT&T’s example sits closer to the sales floor. Its “agent briefing” uses check-in details captured through the AT&T retail app and packages a customer’s key information, so a retail expert knows who they are helping and why before the conversation starts. John Miller, VP of Consumer and Business Solutions at AT&T, said the goal is proactive customer service. Salesforce adds that making the briefing work required investing in data quality and governance up front.
Customers say the data comes from the chat itself
Southwest Airlines uses its chat logs to decide where to deploy agents next. Megan Rauber, Manager of Customer 360 at Southwest, said the airline is “using a lot of the data that we’re finding in the chat experience to help drive and know where our customers want to be serviced in this experience.” She added that Southwest is “continuing to see our CSAT score improve week over week” as it introduces new capabilities.
Rudi Khoury, Chief Digital Officer at Fisher & Paykel, made the same point from the other direction. Teams start with hypotheses about what customers want, he said, “and you quickly learn through doing what they actually need.” In both cases the service channel works as a research instrument, and the first deployment is treated as the start of a loop.
Guardrails came first in the regulated cases
The customers Salesforce describes as most durable built governance into the rollout from day one. AT&T can turn its agents on or off at any time and runs a trust layer plus its own LLM for anything that touches proprietary IP, according to the write-up.
Sammons Financial Group ran more than 200 guardrails and tests before going live and chose to stop the model from learning from live conversations, which Salesforce says removes drift in a compliance-sensitive setting. Andrew Walling, AVP of Capability Planning and Delivery at Sammons, described the fail-safe: “We have a kill switch, a supervisor agent that’s constantly listening to all of those phone calls, checking the values in the CRM against what was said on the line.”
Takeda is automating administrative work that consumes its sales reps’ time while staying inside pharmaceutical regulation, Salesforce reports, and Samer Ansari, Head of Commercial Innovation and External Experience, said the company has to keep up with the technology as it moves.
The model underneath
Salesforce is also training its own model for this work. In its Dreamforce takeaways post, the company describes Koa as its first purpose-built CRM reasoning model, built on the NVIDIA Nemotron 3 Super architecture. In internal pilots, Salesforce says Koa was 14% better than internal benchmarks at retaining context across long multi-turn conversations, gave 15% more relevant answers to end users and was 11% more precise in executing the correct tool calls. Those are Salesforce’s pilot results, and the post does not describe the benchmarks.
What it means for the marketing leader
Our read, not Salesforce’s: when service agents are justified on revenue, the metrics that judge them start to overlap with the ones marketing already owns. Resolution rate and handoff quality sit next to conversion and retention on the same dashboard, and the data the agents collect, such as what customers ask and where they stall, becomes first-party input for segmentation, offers and product messaging.
That raises a question of ownership that marketing teams have already run into with other agent deployments. GTM Teams Deploy AI Agents Faster Than Governance covered the gap between deployment speed and controls, and Own the Approval Step Before Agents Own Your Stack argued that someone in marketing should hold the approval step. The Sammons kill switch and the AT&T on/off control are concrete versions of that step.
What to do with this
Before approving budget for a customer-facing agent, ask for four things in writing. First, the resolution rate by inquiry type, so routine volume is separated from the cases that need people. Second, how any revenue attributed to the service team is measured and who signs off on it. Third, a named owner for the switch that turns the agent off, plus a supervisor process that checks agent statements against the system of record. Fourth, a plan for routing what customers say in chat to the teams that write the campaigns.
Teams that already run service and marketing on separate stacks can start smaller. Pick one high-volume inquiry type, set the success measure before launch, and and read the chat data as it comes in, which is close to what Southwest and Fisher & Paykel describe.
Source: Salesforce News
