AI pilots are stalling at rollout, executives warn

Corporate AI pilots can look promising in controlled tests and still falter once companies try to use them across the business. Executives at Fortune Brainstorm Tech said the failures often stem from weak planning, vague goals and data problems rather than the AI systems alone.

The discussion, reported by Fortune, centered on a common problem for companies investing in AI: a test project wins approval after a pilot, then underperforms or stops delivering the expected business value. The executives said companies need tighter decisions about which pilots advance, clearer definitions of success and earlier checks on data access, privacy and security.

Governance before scale

Sean Bruich, chief technology officer at Amgen, said companies should encourage many AI experiments but apply strict standards before approving broad deployment. He said pilots can multiply quickly inside large organizations, and that experimentation is useful only if leaders decide carefully which projects deserve more resources.

Bruich told the roundtable that “the key to making pilots scale successfully” is pairing a broad set of ideas with “very tight governance” over which pilots move forward. Fortune reported that he framed this as a way to separate useful projects from tests that may work in a limited setting but lack broader value.

Lashonda Anderson-Williams, Salesforce’s chief customer and commercial officer, said companies also need to define the business result they want before they judge an AI project as successful. According to Fortune, she said some organizations focus too heavily on whether AI features have been installed and too little on whether those features improve business outcomes.

That creates disappointment, Anderson-Williams said, when a tool works technically but fails to change results in a meaningful way. Her point was that the target cannot be the presence of AI; it has to be the business change the AI is meant to produce.

AI agents need documented work

Anderson-Williams said agentic AI projects require a detailed view of how work gets done inside an organization. Fortune reported that she pointed to the need to identify the people, teams and contact points involved in completing a task before AI is placed into that process.

She said many companies discover that those workflows are missing or badly recorded. “When you put AI on top of that, the expectation is you’re going to see some magic, and there’s no magic there,” Anderson-Williams said, according to Fortune.

Data can create another barrier between a pilot and full rollout, the executives said. Fortune reported that company data is often spread across separate systems, with different access rules and privacy or security limits attached to it.

Caitlin Halferty, chief data officer at Thomson Reuters, said companies should identify those data needs early in project discovery. “The earlier we can uncover that in discovery, the better we’ll be set up for success,” she said, according to Fortune.

Halferty said privacy and security teams need to be included when an AI project touches personally identifiable information, confidential material or cyber risks. Fortune reported that she urged companies to bring in the right internal groups before those issues slow a rollout.

Bruich also said broad support across the company is needed for AI projects meant to change the business. According to Fortune, he said major efforts will involve leaders from finance, technology, human resources and other functions, and should aim for an enterprise-level result rather than a small efficiency gain for one team.

This story draws on original reporting from Fortune.