Last Year's AI Is Good Enough. Dreamforce 2026 Proved It.
Fifty thousand business professionals gathered in San Francisco this week to talk AI. The loudest message from the floor: the models companies already have are more than they can handle.
Key points
- Salesforce's Dreamforce 2026 conference, held 15-17 September in San Francisco, drew 50,000 attendees and put the gap between AI frontier research and everyday business use on full display.
- Agentforce, Salesforce's customer service automation tool, runs on older AI models rather than the newest releases such as Anthropic's Claude Fable 5.1 or OpenAI's GPT-6 Astra.
- Switching from flat software subscriptions to usage-based AI pricing could cut gross margins from above 85% to around 45%, according to G2 chief innovation officer Tim Sanders.
- The safety debate dominated the keynote stage and barely registered on the expo floor.
The headline debate at Dreamforce this year was supposed to be about AI moving too fast. Nvidia CEO Jensen Huang told the crowd to "run as fast as you can." Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman had spent the days before the conference calling for a slower pace, after an Anthropic researcher resigned publicly and accused the top labs of "gambling with our lives." Our coverage of Jensen Huang's argument that markets can keep AI safe sets out why that confidence is contested.
Walk the expo floor, though, and the conversation was different.
"It's already hard enough to keep up," said Alec Bronston, senior Salesforce director at Chicago retail data company Spins, speaking to CNBC. A slowdown would give companies like his a chance to "even catch up and get our feet wet."
What models are businesses actually using?
Most are not using the newest ones. Older, cheaper models handle the bulk of real business workloads, and many companies deliberately wait before upgrading.
Engineers at Nagarro, a large technology services firm, wait roughly three months after a new model releases before plugging it in. Agentforce, which handles customer service and sales tasks automatically, does not rely on Anthropic's Claude Fable 5.1 or OpenAI's GPT-6 Astra, according to Salesforce's own support documentation. AI2Day has tracked Agentforce since 27 August; our report on the Koa model Salesforce built with Nvidia shows how far the company has moved toward running its own infrastructure rather than paying frontier labs.
"The majority of agentic outcomes aren't driven by frontier capabilities," said Tim Sanders, chief innovation officer at software review company G2. "They're driven by last year's AI."
Docusign, the electronic signature company, uses the most powerful models only for the heaviest tasks: complex contract analysis, reasoning across many documents at once. For everything else it routes requests to cheaper options automatically. That technique, called model routing, is when software decides on the fly which AI model to use based on task difficulty. Nice, which makes cloud contact centre software, uses the same approach.
What does this mean for software companies' profits?
The financial shift is the part most investors are not watching closely enough.
Traditional software companies charge flat monthly or annual subscriptions. AI flips that to a usage model: you pay per "token," a unit roughly equal to three-quarters of a word. Sanders told attendees that this change could drag gross margins, the profit remaining after direct costs, from above 85% down to around 45%. That is not a rounding error. It is a different business.
Databricks, a data-analytics company, gave all 3,500 of its software developers access to GPT-6 Astra this week. Engineering vice president Patrick Wendell said publicly that Astra "unambiguously outperforms" previous top models on complex tasks. Databricks is the exception at Dreamforce, not the rule.
The most trafficked booth on the expo floor was Anthropic's, where young staff in oversized white sweaters ran demos. Claudeforce, a tool Salesforce unveiled in August that lets salespeople pull data from inside Anthropic's Claude chatbot, was the hottest product topic.
Safety dominated the stage. It barely came up on the floor.
That gap is worth watching. If the companies building on top of AI are still figuring out models from two years ago, the frontier race may matter far less to most businesses than labs and investors assume.
Honest takeaway: If you're a business owner or manager feeling behind on AI, you probably aren't. Pick one older, cheaper model, put it to work on a single real task, and measure the result before worrying about what comes next.
