Getty Images

Google Cloud customers disclose GenAI strengths, weaknesses

Troubling flaws remain a problem for enterprises, but Google Cloud customers, including Ford, Belk and Deutsche Bank, find the tech too compelling to pass up.

Listen to this article. This audio was generated by AI.

LAS VEGAS -- Enterprises are advancing their generative AI prototypes to production, having sufficiently diminished the technology's shortcomings.

At last week's Google Cloud Next conference, executives and technologists from a dozen companies shared what they learned from working with GenAI to improve business operations.

GenAI's troubling flaws remain unsolved. The models powering AI capable of producing text, imagery, video and audio still manufacture erroneous responses to questions, known as hallucinations. Prompt engineering, the process of ensuring the model's output is accurate and inoffensive, remains complex. And keeping sensitive data private continues to be a concern despite the safeguards offered by cloud providers.

Nevertheless, many Google Cloud customers say GenAI is good enough to roll out slowly. Waiting for perfection would delay reaping the technology's promise of higher revenues and significant cost reductions.

"Hallucinations will never go away," said Arindam Bhattacharya, a senior applied scientist at Uber. "If anybody tells you they're going to resolve hallucinations, don't believe them."

Enterprises must set their level of comfort with GenAI's risks, Bhattacharya said in an interview.

"If you're a healthcare provider, you might say it's not going to work for me," he said. "If you are a content creator, you're OK if it messes up one term."

GenAI in retail

Belk, the nation's largest privately held department store chain, uses Google's Gemini model and Vertex AI developer tools to create customer-enticing 150-word product descriptions from manufacturers' dry facts for its website. People review the AI-generated depictions before they're published.

Belk engineers spent three weeks figuring out the correct prompts to get an acceptable product brief. The retailer will have to continuously fine-tune the prompts to maintain the same level of quality.

Nevertheless, the potential benefits are significant. Belk currently generates descriptions for roughly 9,200 products, most targeting women. Eventually, it'll use the technology for its inventory of millions of goods, including clothes, shoes, accessories, cosmetics and home furnishings, in 300 stores across 16 states, mainly in the Southeast.

Belk expects to pay $40,000 a year to generate product descriptions using Google Cloud. Still, CIO Richard Spencer said he estimates the automated process will result in an overall 3% margin gain on products, which amounts to several millions of dollars a year.

Spencer is confident enough in the potential savings that he has requested an additional $250,000 in the next fiscal year for GenAI projects, including a chatbot that will help customers pick gifts for friends and family.

The service will come with warnings because Belk can't guarantee there won't be hallucinations.

"There's going to be plenty of little subtexts on the page that says, 'This is a generative AI experiment," Spencer said in an interview. "I think it draws interest, but it also protects us."

Retailer Belk at Google Cloud Next 2024
Belk IT Director Subramanian Seduraman explains how the retailer navigated the difficulties of prompt engineering when using Google Gemini. On the right, Google Cloud account manager Sam Doan looks on.

Data protection in healthcare, banking

Hackensack Meridian Health in New Jersey can't appease state and federal regulators with a warning label, so it's more cautious in its approach to GenAI. The medical center is exploring whether it can use the latest AI services within Google Workspace, its collection of cloud-based video collaboration and productivity software.

The medical center currently bans recording meetings other than training sessions to protect patient privacy.

"We're struggling with how we use the advancements appropriately and continue to protect ourselves, our patients and our caregivers," said Kathy Young, vice president of infrastructure and operations at Meridian Health, during a conference session. "It's our compliance officer that's probably having the hardest time with AI."

Deutsche Bank and KeyBank are two other highly regulated institutions approaching GenAI cautiously.

Deutsche Bank is doing development work on Google Cloud to help its 200 researchers write client reports faster by using AI to cull information from the bank's global financial and market data. More efficient report generation will allow the bank to provide information to clients quicker on the potential economic impact of world events, said Bernd Leukert, the bank's chief technology, data and innovation officer.

KeyBank is testing GenAI within back-office applications, such as summarizing regulatory policies and procedures in the 15 states where it has branches, CIO Dean Kontul said. The bank also wants to summarize some internal credit documents and is evaluating Microsoft 365 Copilot.

The initial use cases for the banks do not involve financial operations that would directly affect customers. The cautious approach is to advance AI use while avoiding mistakes that would raise concerns among overseers.

"The regulators are going to be very closely watching all banks as they move forward, without a doubt, and we're going to have to keep them in the loop," Kontul said in an interview. "From a risk perspective, there isn't a solution that we've deployed that hasn't led to a deep conversation about the benefits and risks."

AI at Ford Motor

While Meridian and the banks approach GenAI slowly, car maker Ford Motor has embraced GenAI-assisted coding. Roughly 80% of the company's developers use AI, increasing coding productivity and quality by 30% to 40%, said Bryan Goodman, director of Ford Motor's AI Advancement Center.

I cannot think of any good excuses at this point for developing software without using artificial intelligence. Bryan Goodman
Director, Ford Motor's AI Advancement Center

"I cannot think of any good excuses at this point for developing software without using artificial intelligence," Goodman said in an onstage interview. "[However], we have to have humans review all of the code."

KeyBank piloted GitLab's Google Cloud service for AI-assisted coding and found a 30% improvement in developer performance, Kontul said. So, the bank will roll out the service for its engineers.

Google became Ford's preferred cloud provider in 2021. Since then, the automaker has migrated nearly all its data onto the cloud, Goodman said in an interview. "We negotiated a really good rate for storage with Google, so storage on the cloud is, for us, a very good deal cost-wise and security-wise."

After the move, Ford remodeled two data centers that used to have web servers, databases and data storage to accommodate GPU servers for model training and running virtual simulations to test car designs. Having that high-performance computing in-house is less expensive than the cloud, Goodman said.

Ford uses Llama 2, Code Llama and Mistral open source models and proprietary models such as Google Gemini. Ford uses the open source vLLM library for model inferencing.

Ford fine-tunes and trains open source models internally for specific applications and plans to do more with the technology.

"Ford does contribute some of our software back to the open source community, but we think we have much more opportunities in that space and we're looking forward to really increasing the amount that we do there," Goodman said.

Future GenAI plans

KeyBank's Kontul expects small open source and proprietary models to have a future in many banking operations, such as managing the risk of fraud. Those models would run in the bank's data centers and be less expensive than using a fine-tuned large language model in the cloud.

"That's the future," Kontul said in an interview. "We're already talking to people about building industry-specific models."

Global advertiser WPP expects to see a change in customer demand as it uses AI to generate images, videos and text faster for ad campaigns. Rather than spend less on advertising, they'll demand more relevant content, CTO Stephan Pretorius said onstage. So, instead of ordering one ad for a new car, they'll want multiple ads for the same price to cover more markets globally with tailored content.

"Personalization and content for marketing channels is sort of an inexhaustible appetite," he said. "There's more personalized content that needs to be made than even today's systems can actually generate."

However, AI has reduced the number of photographers and filmmakers used by WPP. The subcontractors are no longer necessary for all video and image production, Pretorius said.

Antone Gonsalves is an editor at large for TechTarget Editorial, reporting on industry trends critical to enterprise tech buyers. He has worked in tech journalism for 25 years and is based in San Francisco. Have a news tip? Please drop him an email.

Dig Deeper on Artificial intelligence platforms

Business Analytics
Data Management