The Agentic Work Unit: Converting Raw Intelligence into Real Work

For the last two years the industry has focused on raw model intelligence. But as we work with customers to move beyond the hype, a critical question remains: Are we actually using these models for anything useful?

To date, AI success has been measured through the consumption of tokens. But tokens only measure how much an AI talks, not the work it actually completes. That's why we're excited to introduce the Agentic Work Unit (AWU).

An AWU is one discrete task accomplished by an AI agent. It is the moment where raw intelligence is converted into real work. It's a prompt processed, a reasoning chain completed, or-most importantly-a tool invoked.

The AWU provides a platform-level metric that captures the breadth of activity happening across the Agentic Enterprise, from Agentforce to Slack AI. To provide a complete picture of this value, we are tracking two key figures:

• AWUs: The total volume of work our platform performs on behalf of customers.
• Tokens Processed: Our footprint in the global AI compute economy, grounding our scale in the infrastructure layer.

From Inference to Work

The relationship between tokens and AWUs is not a fixed ratio; it's elastic. As our platform innovation matures, and our customer's implementations improve, we expect to see a divergence in tokens versus AWUs, implying that more work is getting done for less cost.

High-frequency, deterministic tasks (like triggering a Flow or calling an API) become increasingly "token-lean". Conversely, complex reasoning and autonomous problem-solving may actually see an increase in input tokens. Especially as agents perform more sophisticated actions like running evaluations to determine the quality of work, designing and optimizing agents with vibe coding, and leveraging even more context for the best possible responses.

The objective here is a high "inference-to-work" ratio: using input tokens to produce concise, high-value output tokens. This is critical because in the world of LLMs, output tokens can be up to 10x more expensive than the input used to get there. Our goal isn't simply to use fewer tokens; it's to make sure that every expensive output token spent is maximizing the work being done.

Leading by Example

We aren't just selling this vision; we're living it. Salesforce is becoming an Agentic Enterprise where humans and agents work together to drive unprecedented efficiency. In our own internal operations, our support agents now handle 96% of cases autonomously, and we've saved over 50,000 seller hours by letting agents handle the "admin" of sales.

We're seeing rapid acceleration across our customers as well, with Service Agents growing 106% quarter-over-quarter to reach 129M Q4 AWUs and Employee Agents experiencing a 62% QoQ increase. In Slack, AI Search grew 116% QoQ, complemented by a 44% increase in File Summaries and the successful debut of our new Slackbot.

These aren't just usage stats; they're proof points of the Agentic Enterprise in action. By measuring AWUs, we are finally moving past the AI chatter to the reality of humans and agents working together. This is how we measure the true value of the Agentic Enterprise.

Go Deeper: