AI is no longer just talking — it's working: agents now complete tasks at industrial scale
Salesforce coined a new unit of measure: the Agentic Work Unit (AWU) — a discrete task an AI agent completes autonomously: a decision made, a record updated, a workflow triggered. In their last quarter, their agents completed a huge volume of these units, and the cumulative total keeps growing quarter after quarter. Not conversations — finished work.
For years, the industry measured artificial intelligence in conversations, sessions, and messages. This month, Salesforce contributed a data point that changes that conversation: its agents now complete Agentic Work Units (AWUs) at industrial scale in production. An AWU is not a conversation — it's a discrete task an AI agent completed autonomously: a decision made, a record updated, a workflow triggered. In the last quarter of their fiscal year alone, Agentforce agents delivered a huge volume of these units, with strong growth over the prior quarter. AI is no longer just talking: it's working.
What exactly an AWU measures — and why it matters more than sessions
The distinction matters for any business evaluating whether an agent is worth the investment. A session counts that someone talked to the AI. An AWU counts that the AI did something. Inside Salesforce's own operations, their agents handled a massive volume of interactions and freed their teams from an enormous number of work hours through Agentforce in Slack. Those hours didn't disappear — they were redirected to decisions an agent can't make: contextual judgment, client relationships, creativity. The AWU isn't just an investor relations metric: it's the first honest way to measure whether AI did real work.
The model built for this kind of work
Anthropic launched Claude Sonnet 5 in late June — describing it as the model with “the largest gains over the previous version in coding and agentic tasks.” Unlike the higher-cost Opus-class models, Sonnet 5 is built for volume: completing agentic tasks repeatably, using tools, running workflows without interruption. It ships with adaptive thinking on by default, a very large context window, and reduced introductory pricing for a limited time. The industry's message this month is consistent: capable models are no longer exotic or expensive — they're the infrastructure for agentic work at scale.
- Review hundreds of incoming support tickets, classify them by urgency, and assign each to the right person — many AWUs, instantly.
- Cross-reference the monthly sales report with inventory and flag every discrepancy for review — one AWU per line processed.
- Draft a contract renewal for each client with an upcoming expiration — one AWU per client.
- Reply to prospects who didn't open the follow-up email with a version personalized to their industry — one AWU per reply.
For Fiona Tan, CTO of Wayfair, AI agents can be applied to so many use cases that practically every business should adopt them, and the savings can be pointed to in hard dollars. (Google Cloud Blog)