It’s becoming more and more common in companies these days: the AI project performs great in pilot testing, gets the green light for wider adoption… and then stops working properly; Or it doesn’t deliver the expected business results.
There is blame, accusations and embarrassment.
The problem is not always the technology. In fact, the mistake often lies in the planning, processes and expectations that companies have established – or not established – around their AI projects, according to business leaders who spoke at a panel Fortune Brainstorm Tech this month.
First of all, not every AI project deserves widespread adoption, he said Amgen Chief Technology Officer Sean Bruich.
“With a pilot, it’s so easy to make a thousand flowers bloom,” he said. This isn’t a bad thing because it encourages experimentation. But, he said, “the key to successfully scaling pilots is actually having a large number of ideas but very tight governance where pilots actually get the green light.”
A key criterion before the next step, he said Salesforce Lashonda Anderson-Williams, Chief Customer and Commercial Officer, understands the intended outcome of the project. Too many companies focus on successfully implementing AI capabilities — the technical bells and whistles — rather than the business outcome, she says.
This mentality is a recipe for disappointment: the AI capabilities work great, but the new technology doesn’t drive meaningful business results.
Agents need a card
When it comes to agentic AI, Anderson-Williams says a detailed understanding of workflow – what people, groups or touchpoints are required to complete a task – is crucial. What many companies find is that workflow documentation is either nonexistent or poorly documented: “When you put AI on top of that, you expect to see some magic, and there is no magic.”
Access to data is a particularly common stumbling block that AI projects encounter as they move from pilot to full deployment. Because data is often scattered in different silos within an organization and all of that data has different access permissions and different privacy and security considerations, things can quickly get complicated. It is important to define the contours of the AI project and all potentially required data in advance, the panelists emphasized. “The sooner we can uncover this in discovery, the better equipped we are for success,” said Caitlin Halferty, chief data officer at Thomson Reuters.
This also means getting buy-in from the right groups and stakeholders within the organization. “Is there an element of PII (personally identifiable information) or sensitive data that will trigger data protection?” Halfery said. If the answer is yes, then the right people need to be part of the project. “Is there a cyber element? Let’s get security on board,” she said.
Amgen’s Bruich reiterated the importance of broad buy-in, noting that any AI project that is transformative for the company will necessarily involve leaders from finance, technology, human resources and other groups across the company. A truly effective AI project must do more than just make work processes more efficient for a small group of employees. It must “deliver an outcome that is meaningful to the company.”