Companies are tossing AI agents onto existing processes, but a transformative change — where AI is the boss — is still far away.
That was the view of IT leaders at this year’s Microsoft Ignite conference who’ve been putting AI agents to work, mostly with legacy processes. The IT leaders discussed their efforts during a conference panel at the event earlier this month.
“We’re probably living in some version of the horseless carriage — we haven’t got to the car yet,” said John Whittaker, director of AI platform and products at accounting and consulting firm EY.
Execs from EY, Pfizer and Lumen who sat on the panel said they have mostly used AI agents for knowledge management, content creation and research. That aligns with findings in an AI study released last month by McKinsey, which found heavy use of AI tools in those areas.
With IT leaders facing pressure to transform their companies into AI-first operations, decision-makers see AI agents as a way to uproot legacy processes for cost savings and productivity gains.
EY, the global tax and advisory firm, has 30 million documented processes internally and 41,000 agents in production. “Moving those processes faster through agentic assistance like Copilot are kind of the low-hanging fruit to get to an improved outcome,” Whittaker said.
The early benefits of AI agents are now visible at EY; the endgame is to abstract processes and applications where data sits, Whittaker said. “We’re beginning to see line of travel that really will allow us to completely transform the experiences,” he said.
One agent, called the EY tax assistant, can answer questions and provide up-to-date tax knowledge to personnel and customers. There are approximately 100 tax changes each day, and the agent functions as a research tool for people to stay updated on thoe changes.
The fine-tuned model incorporates information from 21 million research and domain documents and is further tuned for field offices to receive relevant updates. “A regular large language model [LLM] type deployment … can be very good, but nowhere near the quality of what you get out of a fine-tuned model,” Whittaker said.
Pharmaceutical company Pfizer is taking a phased approach to putting agents in production. The company first tries out models, gains confidence in the results, and then scales it up.
Pfizer’s call-center agents started in a few locations and ultimately spread to more locations. The agent answers queries and solves customer problems with real-time telemetry and information, said Tim Holt, vice president of colleague and consumer technology and engineering at Pfizer.
“Being able to start with a couple of them [and] make them more efficient, then gives us the opportunity to do it again and again and just make better and better efficiency gains as we go,” Holt said.
Pfizer is very process-centric, he said, stressing that the goal is not to reinvent processes right out of the gate. The company is analyzing how AI works for them, gaining confidence in the technology before reorganizing processes within the AI lens.
“Where we’re definitely heading … is thinking about, ‘I’ve solved this process, I’ve been following exactly the way it exists today. Now let’s blow it up and reimagine it…’ — and that’s exciting,” he said.
About half of Pfizer’s 75,000 employees use Microsoft’s Copilot tools.
Lumen CEO Kate Johnson is one such user, relying on Copilot for tasks that include research and executive briefings, said Sean Alexander, senior vice president at the telecom firm.
Using gaming terminology, Alexander outlined a multi-step process for autonomous agents driving company processes. Level one is human-to-agent, level two is human-to-multi-agent, and level three is “when you have full orchestration happening between the different agents,” Alexander said.
Lumen is now looking at where it wants the business to be in 36 months and linking it to AI agents and AI-native plans. “We’re … working back from that and ensuring that we have the right set of tools, the right set of training, and the right set of agents in order to enable that,” he said.
Every new Lumen employee in Alexander’s connected ecosystem group gets a Copilot license. The technology has helped speed up the process of understanding acronyms and historical trends within the company.
“It takes about six months for a new employee to become fully realized in terms of their potential. It is getting shaved down to three months,” Alexander said.
Lumen spent a week with Microsoft and Harvard looking at specifically where the puck is moving with AI agents. “We are absolutely in the early innings of agentic technology,” he said.
