Salesforce wants ASEAN customers to stop thinking in tokens

Salesforce is keeping tokens out of its pricing talks with customers in Southeast Asia, where businesses will no longer fund AI projects that can’t show a return. “If I’m going to my CFO [chief financial officer] and I say, ‘I need five million tokens’, what does it do? What am I going to get for that?” said Gavin Barfield, vice-president and chief technology officer for solutions at Salesforce ASEAN. “We’re taking that layer away from customers,” he said. “Customers don’t talk to us about tokens. Customers talk to us about Flex Credits, which relate to actions.” Flex Credits, the consumption-based units of payment for Salesforce Agentforce , are drawn down only when an agent does something, such as looking up a record or taking an action in a system. “I did an action, so I pay something for it – as opposed to, I’ve burnt 50,000 tokens and I don’t quite know what I’ve done with them,” Barfield said. Barfield and Paul Carvouni, senior vice-president and general manager for ASEAN at Salesforce, were speaking to Computer Weekly at Dreamforce 2026 in San Francisco. At the event, the company unveiled AIforce , which lets people work with data and workflows in its customer relationship management (CRM) platform from inside Claude and Slack, or through its Agentforce Coworker assistant. For customers seeking certainty on cost, Carvouni pointed to the Agentic Enterprise licence agreement. It lets them “size up their consumption and then have essentially an unlimited commitment”, he said, with Salesforce taking on “a lot of that token burden”. The pitch comes as AI bills across Asia-Pacific keep climbing . Tokens cost less than they did in 2023, but Kearney estimates the number used per request has risen 184% over the same period because newer models do more reasoning. Salesforce’s own pricing mechanisms have evolved as well. Agentforce launched at $2 per conversation, switched to Flex Credits in May 2025 and later gained per-user licences. Carvouni said the mix gives choice and flexibility to customers who consume in different ways, and includes outcome-based pricing for ready-made agents where it makes sense. Finding value The generative AI wave that came before AI agents produced piles of proofs of concept (POCs) and pilots, Barfield said, and most never left that phase. “There was no real link to value. It was technology looking for a business problem.” Customers are now far more deliberate about tying projects to value, he added. “The days of them being able to have a blank chequebook to play around with things in POC are gone. Their leadership is now saying, ‘Show me value. Show me how these AI tools are actually going to change this metric. Show me how it’s going to reduce my call handling time.’” Barfield’s team now sends forward-deployed engineers into selected customers for four to six weeks at a time. “Instead of doing presentations and diagrams and videos, we’re building these things in a few weeks,” he said. Once a build shows value, it goes into production and the team moves on to the next use case. Ready-made agents are part of the same push. Salesforce released a batch ahead of Dreamforce, including Casey, which handles customer queries over voice, SMS, WhatsApp and web chat. Marshall runs supply chain and back-office processes, while Hunter, an outbound sales agent, is in pilot until general availability in November. Carvouni expects agents to boost CRM adoption in markets where it lags. “If I’m using a small percent of my CRM, agents can get a lot more out of that,” he said. Users might ask for their top 10 lost deals, how to de-risk the next renewals or “why am I winning deals in Thailand?” “Arguably, AI is able to use software in some circumstances better than humans, because humans don’t always want to write detailed notes,” said Barfield. “Humans don’t always want to fill in every field.” Within a few years, Barfield expects companies to run “tens and hundreds and sometimes thousands of different agents” and draw on a wider choice of large language models (LLMs) than they do today. Keeping tabs on them falls to MuleSoft Agent Fabric, Salesforce’s control plane for agents from multiple suppliers. It can discover agents and show what data they touch and who has access to them, said Barfield, with cost management capabilities in the pipeline. Many customers are “using a sledgehammer to cut a nut”, he said, running a massive LLM on something as simple as a case summary when a smaller or open-source model could do the job very well. Customers can already pick a model down to the prompt. With Agent Fabric they can instead set boundaries and let the platform choose a model depending on latency and task, he said. Digital natives such as Grab are “very mature” in matching models to use cases, added Carvouni. Some organisations may prefer home-grown models such as AI Singapore’s Sea-Lion . Barfield said a range of models can be used for prompts through Amazon Bedrock and other platforms, but reasoning engines have to be certified and run on particular models. The newest of these is Koa, Salesforce’s first CRM reasoning model, built by post-training Nvidia’s Nemotron 3 Super on synthetic data modelled on nearly three decades of CRM deployments. “It’s probably not very good at looking up a cheesecake recipe, but that’s not what customers want to do,” said Barfield, who expects many more small models doing narrow tasks. Asked whether AIforce is largely a repackaging of what Salesforce already had, Barfield said it was “bringing together a number of capabilities that have come out in parts” – including headless tooling shown at its TDX developer conference earlier this year – and setting out the roadmap. “The Claude stuff is new,” he said. Salesforce in Claude ships with 37 prebuilt sales skills, which he said were built with Anthropic’s engineers rather than “just knocked up”. Previous instalments of Dreamforce were about adding features to different Salesforce cloud services, Barfield said, but “this one has been quite different because we’re talking about fundamental engineering and plumbing”. Read more about AI in ASEAN Anthropic has hired a former Google Cloud executive to run its ASEAN business , days after CEO Dario Amodei called on the AI industry to slow down. NUS Libraries’ AI Sense Maker draws on 150,000 research papers and digitised Southeast Asian materials to help students and researchers generate insights from trusted sources. DBS has strung together as many as 80 agents to prepare credit approvals for big corporate clients but says the industry’s ability to police such systems is advancing at a fraction of the pace of the technology itself Ryt Bank is using its own LLM and agentic AI framework to allow customers to perform banking transactions in natural language, replacing traditional menus and buttons. [Read Full Story on Techtarget →](https://www.computerweekly.com/news/366650532/Salesforce-wants-ASEAN-customers-to-stop-thinking-in-tokens)