When Krishna C. Mukherjee observes brands trying to figure out whether to introduce AI assistants into their customer experience (CX), he suggests they ask themselves a fundamental question: does it respect the customer’s attention, or does it spend it?
Mukherjee has a career history that shows how well he’s applied that thinking. While working at Microsoft in the late 1980s, he was part of the team that developed Office features such as Autocorrect and Intelligent Filing Manager.
During that period he also witnessed the emergence of what might be considered one of the earliest – and most reviled – virtual assistants of all time: Clippy, an animated paperclip that would pop up and offer writing ups and assistance.
“I watched as one of (Microsoft’s) most technically ambitious efforts become a global punchline,” Mukherjee told 360 Magazine. “What went wrong was not a matter of ambition or capability. The product intruded, it presumed, and it was hard to dismiss. It spoke up unprompted, guessed at intentions it could not reliably infer, and made people work to get rid of it.”
The risk is that many brands could make the same mistake with chatbots and virtual assistants that are becoming part of their CX today. That was one of the catalysts for Mukherjee’s book, The Microsoft AI Insider’s Playbook: Your Guide to Cloud Strategy and Digital Transformation.
While Clippy failed to gain traction, Mukherjee contrasts it with the proofing tools he and his colleagues introduced later. These took a fundamentally more subtle approach to digital CX, he argues, such as a quiet, wavy underline when Word detected a spelling error, and a one-keystroke “undo” function.
“Hundreds of millions of people still use those features because they helped without demanding attention,” Mukherjee explained.
The reason AI pilots are increasingly failing, he continued, is likely because nobody measured the process AI is touching before automating it, or stakeholders are pursuing different goals under the same project name. Many pilots are also designed to demonstrate the technology rather than survive production, he adds, which means there is no plan for data readiness, governance, monitoring, or how the technology will deliver against real-world expectations.
“For AI agents in particular, I typically add three measures: how often the agent completes eligible work correctly on its own, how often it escalates it to a person, and how often a person has to reverse what the agent did,” he said. “An agent whose actions have routinely reversed has not earned autonomy.”
In the early days of office productivity software, failures were often highly visible compared with AI that runs quietly in the background, Mukherjee said.
“Its quality degrades without an error message as the surrounding data changes,” he said. “And because the software operates at scale, these mistakes are hard to notice and spread faster.”
Mukherjee suggested vendors and brands would do well to build AI deployments for the day after launch. In other words, thinking beyond getting the green light to deploy the technology in customer-facing scenarios. “Most AI today is built to demo, and very few systems are built to endure,” he said, adding that his own career has given him hope for the future. “As the technology matures, and it will take some time, people will have faith in this in this technology.”