Salesforce has released the second edition of what it calls its Agentic Enterprise Index, a study of how businesses use its AI agent product, Agentforce. The headline numbers are striking. Companies that stuck with the platform now run about 13 AI agents on average, up from 5 agents a year earlier. The time it takes to build and launch a new agent dropped by more than half, down to roughly 2 days.
To prove these agents are actually worth something, Salesforce built a new measurement called the agentic work unit, or AWU. Every task an agent finishes, a question answered, a form processed, a decision made, counts as one AWU. By this measure, work done by agents grew at a rate of 15% every month through April. Retail businesses alone accounted for 22% of all the work these agents did, and their AWU output grew 18 times over during the study period.
This is worth pausing on, because the AWU is Salesforce's own invention, graded on Salesforce's own scale, using Salesforce's own customer data. Outside analysts have already pointed out the flaw. A workflow can be triggered and still fail. A customer question can be answered and still be wrong. The AWU counts the attempt, not the outcome. One industry analyst put it plainly: the metric tracks activity, not quality.
That gap between "the agent did something" and "the agent helped the business" is not a small detail. It is the entire question companies are trying to answer before they spend real money on this technology. A separate survey by software company Aptean, covering more than 1,500 decision makers, found that less than half of businesses currently consider AI essential to their core operations. Gartner has gone further, predicting that more than 40 percent of agentic AI projects across the industry will be shut down by the end of 2027 because of rising costs and unclear payoff.
None of this means the technology is fake. Salesforce points to real examples: PenFed, a credit union serving military families, built an agent called Ace that checks account balances, tracks loan applications, and moves funds for members, plus a voice version called Echo. Salesforce also reports that even as agent conversations scaled up sharply, the share of conversations that had to be handed off to a human stayed close to 32 percent, meaning roughly 7 in 10 customer service chats got resolved by the AI alone.
That last number matters more than the AWU count. It tells you agents are handling real conversations without falling apart at scale, which is the actual worry most operators have. If you are evaluating agent tools for your own business, treat growth metrics from any vendor, including Salesforce, the way you would treat a sales pitch. Ask instead for the numbers that are harder to fake: how many issues got solved without a human, how much time or money it saved, and how often it got escalated because it got something wrong.