When revealed data brings AI rollouts to a screeching halt - and how to manage it

1 hour ago 5
gettyimages-2227611669
Gary Yeowell/DigitalVision via Getty

Follow ZDNET: Add america arsenic a preferred source on Google.


ZDNET's cardinal takeaways

  • AI tin boost productivity and amended information access.
  • Tech leaders person had to halt rollouts owed to information concerns.
  • Long-fogotten insights look with AI prompts.

Agentic and generative AI person opened up accusation and insights to professionals successful enterprises. However, grounds suggests that inclination could beryllium excessively overmuch of a bully thing. At a caller conference, veterans of endeavor AI rollouts issued cautionary words to professionals considering diving headfirst into AI.

The issues these professionals encountered adjacent led to impermanent halts successful AI rollouts meant to boost worker productivity, arsenic executives reassessed accusation that could beryllium exposed internally. At the aforesaid time, the executives, who spoke connected a sheet astatine the caller Veeam league successful New York City, emphasized that AI wasn't the root of the challenge. Both panelists' organizations had accumulated immense stores of data, and 1 required a caller governance structure.

Also: 96% of IT pros usage AI now: Their apical 7 agentic applications and biggest implementation roadblocks

Steve MacIntyre, elder vice president astatine Fidelity Investments, described however his 400,000-employee institution saw information agelong tucked distant successful the recesses of its enactment -- connected SharePoint sites oregon successful network-attached storage, for illustration -- abruptly aboveground via AI prompts. "It wasn't an AI problem," helium said. "It was the productivity and the quality of AI to find things quickly."  

Wim Geurden, main designer for endeavor tech astatine EY, described his company's situation arsenic pinning down information ownership crossed its planetary web of autarkic affiliates -- information that was besides surfacing done its AI engine. "When large endeavor hunt was launched, each kinds of worldly started to aboveground successful places that radical went," helium said. 

"EY Global doesn't ain immoderate of the data. Every subordinate steadfast owns its data. That is wherever the archetypal questions were raised. What's each this? How galore SharePoint sites? We had aggregate petabytes of data, and it was the Wild West. There was nary lifecycle absorption connected these SharePoint sites, and fractional of them had nary owners. We didn't cognize erstwhile they were past accessed."

Also: 51% of professionals accidental AI workslop lowers their productivity - halt it successful 2 steps

At Fidelity, accusation was emerging from a immense room of PowerPoint and PDF reports. "We person an full history, decades of probe notes astatine Fidelity, specified arsenic PDFs," said MacIntyre. "We gave retired a fewer licenses for Copilot, and immediately, 2 days in, ineligible came to maine and said we person an AI problem. One of my squad did a hunt to find thing and AI came backmost with each the PowerPoints that were connected SharePoint from years ago."

AI is simply a "tremendous hunt motor that runs astatine speed," MacIntyre continued. "Suddenly, it's searching everything that it has entree to, and surfacing that to america successful a meaningful way. Everybody thought we had an AI problem, but what it showed was a occupation of securing data. This contented deed location erstwhile we instantly realized that we had each of this information that we didn't deliberation we cared astir -- unstructured information -- and on came LLMs, and abruptly each of that information becomes gold."

Establishing guardrails

At EY, arsenic the gates of its immense information stores opened to AI, the precedence was to "find who owns the data," said Geurden. "The 2nd happening we did was we unopen everything off." Users could lone entree the Copilot instrumentality if they were licensed.

Also: Building an agentic AI strategy that pays disconnected - without risking concern failure

The information ownership verification process included identifying and labeling the information recovered crossed the EY enterprise, Geurden continued. For example, labels included "confidential" oregon "financial services."

AI itself offered a means to assistance statement the company's cognition repositories of unstructured data, Geurden explained, noting the situation of quality labeling with a 25% yearly turnover rate.

However, labeling needs to spell deeper than elemental high-level tags. "The archetypal happening is we person to cognize what was determination erstwhile the AI ran," said Geurden. "We request to person the humanities picture, the versions." Then, "we person to spell mode beyond the labeling of confidential information. We request to person geo-restrictions, geo-labeling, line-of-business labeling, linked to our contracts, due to the fact that we get an tremendous magnitude of lawsuit information specifying what we tin bash and what we cannot do."

Also: Over 80% of US authorities agencies already usage AI agents - and it's lone the beginning

All of this metadata has to beryllium codified into contracts, helium added: "That's the casual part. Then we person to codify it successful immoderate technological structure. That is, for now, inactive very, precise cumbersome."

Governance is the cardinal to success across each aspects of these AI implementations, the executives emphasized. "We person to cognize what's being used," said MacIntyre. 

"That brings into play the thought of shadiness AI, shadiness IT, each those kinds of things -- and it goes backmost to the endpoint data. We person to cognize that the plus inventory is accurate. Are they aligned with the usage cases that are registered and approved? That way, astatine slightest we cognize that if someone's moving connected something, they should beryllium utilizing Claude, due to the fact that it's tied to a peculiar task that was approved for that."

Also: These 4 captious AI vulnerabilities are being exploited faster than defenders tin respond 

Next, "we person to deliberation astir what's the harmless situation wherever we privation these agents to run?" MacIntyre continued: "How bash we privation them to interact with the foundational models? What architecture bash we enactment successful spot to funnel each that enactment into a spot that gives america the close visibility and telemetry truthful we tin spot that agents and applications utilizing AI are behaving successful the mode that was intended? Or misbehaving?"

An further situation -- possibly the astir vexing for each integer leaders astatine this clip -- is establishing cause identity, said MacIntyre: "How bash you springiness an cause identity? They past go an employee. But what if my cause lone lives for seconds? It's a truly absorbing problem, and I don't cognize if anybody's solved it truly good yet."

Read Entire Article