Mustafa Suleyman, Microsoft's CEO of AI, recently declared in MIT Technology Review that computing is at the threshold of "nearly human-level agents." Which is great, assuming we can figure out how to keep them from accidentally emailing your private medical records to a stranger.

According to Databricks' State of AI Agents report, only 19% of organizations have deployed AI agents, and even then mostly on a limited basis. Craig Wiley, Databricks' head of AI, told ZDNET that CFOs have three concerns: control, quality, and cost. So basically the same concerns they have about their teenage children, but with more data leakage.

Wiley's first best practice is governance - specifically, controlling what data an agent can access. He pointed to women's health app Flow, which has 75 million users and a legitimate fear of mixing up one person's ovulation chart with another's. "The last thing they would want is an app user to get a response that includes some other app user's information in it," Wiley said, with the understatement of a man who has never seen a data breach headline.

Asset manager Franklin Templeton faces a similar challenge with portfolio reports. Nobody wants to receive an email from their financial advisor that starts "Dear Client, here's someone else's net worth." Wiley emphasizes that data segmentation must be "deterministically forced," not merely suggested in a prompt - because AI, like a distracted intern, will take the path of least resistance.

The second practice is evaluation. When Flow's developers needed to ensure accuracy, they didn't ask programmers to judge the output; they asked actual physicians. "The software programmers write what's called the orchestration system," Wiley explained, "but it is physicians who were saying, 'this response over here needs additional context or color.'" Evaluation should be ongoing, checking not just final answers but every intermediate step of the agent's thinking. Companies that do this are six times more likely to get agents into production - which is either a testament to evaluation or a damning indictment of everyone else.

The third concern, cost, is essentially the reward for doing the first two things right. "Once you can do those two things, to be honest, the rest of it becomes implementation details," Wiley said. But cost must be considered upfront: "Is this something that we can solve today inside a reasonable cost envelope? And assuming we can solve it, is it actually going to move the needle in your company?"

Wiley advises starting small. Convenience store chain 7-Eleven built a "super assistant" for service techs that accesses tons of documentation about equipment, reducing the need to call a buddy and ask, "Have you seen this problem before?" The result: a 25% increase in first-time fix rates and a 40% drop in time to repair. Baylor University uses agents to review recordings of calls with prospective students, analyzing decision factors that human note-takers inevitably miss.

Franklin Templeton's automation of investment portfolio analysis identified over $15 million in new product opportunities - presumably by noticing gaps in client portfolios that humans were too busy to see.

Wiley compares the current state of agentic AI to "the equivalent of 2001 on the web, where companies are investing in their web pages but don't really understand the purpose of all this yet." The key takeaway: get your data in order first. "If your data's in good shape, we could do it [build and deploy an agentic system] this afternoon," Wiley said. "If your data is in rough shape, then the real problem is going to be how long it takes us to get your data in order."

So the path to nearly human-level agents is paved with clean data, rigorous evaluation, and the eternal hope that your AI won't accidentally reveal your deepest secrets to a stranger. Welcome to the future.