Every large company collects enormous amounts of information about its customers. Purchase patterns, call center transcripts, focus group recordings, social media comments, survey results from dozens of markets. Most of that information sits in separate folders, separate systems, and separate countries, touched by people who have no idea what their colleagues already know.
AI is now being used to fix that problem, and in some cases it works very well. The approach, broadly, is to feed all that stored knowledge into a system that can then answer questions in plain language. You type: "What do we know about how customers in Germany feel about our packaging?" and the system finds the relevant studies, summarizes them, and points you to the exact line in the relevant document. No librarian required, no three-week wait for a research team to pull a report.
Novartis built exactly this. Their system, called Sherlock, pulls from a single centralized pool of market research across the company. One team at Novartis wanted to know whether patients preferred blister packs or bottle packaging for a particular drug. One group commissioned a fresh survey: $100,000, three months, 50 patients. Another group used Sherlock and got answers drawn from 6,700 patient records in three weeks, at zero extra cost. The company saved over $29 million in primary market research costs in a single year by avoiding that kind of duplication at scale. Today, over 10,000 people across Novartis use the platform regularly.
PepsiCo went a step further. Their platform, called Ask Ada, collects results from advertising tests, consumer research, and market studies run across every country where PepsiCo operates. Before Ada, a team in Guatemala would run an ad test using a local agency and local tools. A team in Germany would do the same, separately. Neither team learned from the other. Now, every test feeds the same system. As PepsiCo's vice president of insights for Europe put it, Ada is their internal equivalent of Google, and every time the company runs a test, the whole organization gets smarter.
But the research that documents these cases, based on interviews with customer insights leaders at eight large consumer companies, is careful not to oversell the technology. The failures it describes are just as instructive as the wins.
One unnamed global consumer goods company bought an AI-powered knowledge tool, deployed it, and then watched it sit largely unused. The company operates in over 100 countries. Different divisions use different names for the same brands, the same customer segments, the same distribution channels. There is no shared vocabulary. The AI cannot reconcile what humans have not agreed to standardize. The tool is currently used by only a handful of regional teams, and the company is still trying to figure out how to build a strategy around something that requires a level of organizational discipline the company has not yet chosen to enforce.
That story repeats across industries. Research from knowledge management surveys shows that 55% of knowledge management programs fail because of poor change management or lack of leadership support, not because the technology did not work. The same finding showed up in the SharePoint era, the Lotus Notes era, and now the AI era. Better tools do not automatically produce better behavior.
The companies that made AI work for customer insights all share one trait: someone with real organizational authority decided to treat customer knowledge as a strategic asset and then built the governance before they built the platform. PepsiCo's Stephan Gans created a Global Insights Council of 15 leaders representing every region before the Ask Ada platform existed. He also made a firm rule: the company owns all research produced on its behalf by external agencies. Not the agency, not a shared arrangement. The client company owns it and can load it into the system.
That ownership question matters more than it might seem. Many companies that rely heavily on external marketing or research agencies discover that when they want to build an AI-powered knowledge base, they do not actually control the research they paid for. The knowledge lives in the agency's systems, in the agency's formats, and accessing it requires going back to the agency. That is not a technology gap; it is a contract and culture gap.
There is also a real limit on what these AI tools can do with data that has not yet been analyzed. Most of the current tools work best when good analysis has already happened and clean insights are stored in the system. They are very good at retrieving and summarizing. They are less capable of doing the original analytical work on raw customer data, particularly qualitative data like interview transcripts or focus group recordings. That kind of analysis still benefits significantly from human judgment.
The practical question for any organization is not whether to use AI for customer knowledge management. The question is whether the underlying conditions are in place: a single repository that teams are required to contribute to, standardized language across divisions, leadership that actively uses the outputs, and ownership of research that has been commissioned externally. Without those conditions, the tool will not fail because of the technology. It will fail because of the same reasons the last tool failed.