Evaluating Ai Agents? Early Adopters Outline Practical Challenges

A sheet chat astatine Nvidia’s GTC Conference focused connected real-world proposal astir nan intricacies of adopting agentic AI.

Non-tech companies trying retired AI agents successful nan field said nan exertion still has a batch of  challenges to flooded earlier it tin beryllium utilized practically. 

Some of nan early uses of agentic AI are successful nan area of customer service, which was portion of a sheet chat astatine Nvidia’s GTC conference, held past week successful San Jose.

The main problem is humans still for illustration talking to humans because they don’t spot machines, said Cameron Davies, main information serviceman astatine Yum Brands, which includes brands specified arsenic Pizza Hut, Taco Bell and KFC.

The institution hopes to get to 100% integer ordering successful 5 years, pinch agentic AI playing a role, Davies said.

Specifically, Yum Brands wants to automate orders astatine drive-throughs, Davies said. One of nan astir challenging work jobs is moving drive-through windows — to take, capable and rate retired orders — and AI agents could lighten that load. 

But nan main rumor isn’t technological, Davies said. “What’s nan top situation successful putting this into place, is that cipher wants to talk to a instrumentality correct now. And you person to inquire yourself why is that nan case?” 

Yum Brands is besides eyeing AI agents to trim nan “cognitive” load connected quality servers and employees. Scripted agents tin do nan domiciled of “upselling and asking astir kindness donations,” Davies said.

“You do these things, past that personification tin now attraction connected being happier, making judge nan bid is right, getting nan alteration right, etc.,” Davies said, adding that Yum Brands has been testing agents for back-office and HR functions, pinch mixed results.

Beyond trusting AI systems, compliance and accuracy are concerns for Craig Daniels, nan caput of Mayo Clinic’s Smart Hospital and Unbound Project.

Healthcare, by its nature, tends to lag different industries successful adopting technology; Daniels is looking astatine nan advancement of AI agents astatine companies for illustration Yum Brands to spot what works.

For Mayo, nan situation is for doctors and patients to summation a precocious level of spot successful AI to assistance successful test and treatment. Then they tin see nan domiciled of agents successful helping doctors and patients.

Patients request to spot AI models conscionable arsenic they spot MRI machines without having to interest astir nan underlying technology, Daniels said.

Mayo Clinic is creating its ain information level pinch anonymous diligent accusation gathered from 61 different healthcare organizations successful 4 continents. For example, 1 Mayo AI strategy trained connected 7 cardinal electrocardiograms tin observe and diagnose bosom failure.

“There’s a constituent astatine which we person to spot nan exemplary works,” Daniels said. “It generates caller insights and we’ve researched and we spot that and we’re utilizing it pinch nan quality to make nan last decision. That’s a awesome advancement.”

The US Food and Drug Administration (FDA) will yet modulate AI successful healthcare done “medical-as-a-software device,” which will require proven research, Daniels said.

More than successful different industries, Mayo will require a assortment of guardrails beryllium successful spot to present trustworthy results arsenic there’s nary allowance for AI aliases agents to hallucinate. Many panelists besides mentioned they were still successful nan process of figuring retired those guardrails. “We want to beryllium safe,” Daniels said.

The panelists said that while customer work and chatbots aren’t new, they’re being revisited successful ray of nan presence of AI agents.

Agents are expected to turn much human-like successful guidance and voice, and they’re much flexible. Agents don’t travel scripts for illustration chatbots, and depending connected nan customer, they tin link to different agents to amended service customers.

“I tin make it talk for illustration a person. I tin alteration dialects. I tin do those things, and I want to power it,” Yum Brands’ Davies said.

Agents won’t afloat switch group immoderate clip soon, arsenic quality ingenuity is still required successful galore areas — to supervise AI agents, modify them to beryllium much effective, and beryllium capable to verify AI output and results.

U.S. Bank’s first AI agents are augmenting quality knowledge to work customers.

Human agents for banking, owe aliases investments are taxable matter experts that request accusation astatine their fingertips; this is wherever nan slope is testing AI agents, said Sumitri Kolavennu, caput of AI investigation and elder vice president astatine U.S. Bank.

“Keeping nan human-in-the-loop successful galore [regulated] industries is … really paramount,” Kolavennu said. “We emotion nan advantages and autonomy aspects of AI agents, but we do want to support nan quality successful nan loop.”

U.S. Bank is testing nan effectiveness of agents by seeing really quickly problems are resolved.