AI might finally make years of customer research usable. Just don’t confuse finding an answer with doing the work that made the answer trustworthy.

Generative AI customer insights tools promise to turn years of forgotten research into answers people can actually use. That part is real. The harder question is what happens when the research underneath is weak.

There is a particular kind of frustration that shows up in big organisations.

Someone asks, “Do we know why customers are dropping out at this point?” Everyone is fairly sure the question has been researched before. There was a deck. Maybe an agency presented it. Someone remembers a chart. Nobody can find it.

So the team commissions another study.

This is where generative AI could be genuinely useful. Not futuristic-useful. Monday-morning useful — provided it does more than make work feel faster without making the organisation better.

New research from Thomas H. Davenport and Viktor Dörfler looks at how eight consumer-facing companies, including P&G, PepsiCo, and Novartis, are using GenAI to manage customer insight. The clearest lesson is also the least dramatic: AI is good at helping companies find, organise, and pull together things they already know.

That sounds modest. It isn’t.

Customer knowledge is usually scattered everywhere — research reports, survey results, interview transcripts, customer-service tickets, social posts, sales conversations, agency folders, and half-forgotten SharePoint sites. The problem is not always that insight is missing. Often, it is simply buried.

GenAI gives that archive a better front door.

But there is a catch, and it is a big one. Retrieving research is not the same as doing research. A confident summary is not automatically a customer insight.

What generative AI customer insights tools are good at

We have tried to solve this before.

Put everything in one place. Add tags. Create a sensible folder structure. Tell people to upload their final reports. Six months later, half the links are dead, two teams have commissioned the same study, and the person who understood the taxonomy has moved jobs.

GenAI improves the experience because people can ask a normal question instead of knowing exactly where to look. With retrieval-augmented generation — RAG, because AI needed another acronym — a model can search an organisation’s own material and build an answer from it.

P&G uses external software to store and access knowledge but has built its own GenAI layer for analysis and classification. Its insights leadership talks about getting “sharp, pointed answers” instead of a list of documents. It is a good example of generative AI customer insights technology being used as an access layer, not an artificial research department.

That is a proper improvement. If someone can ask what the company already knows about a customer problem and receive a useful answer with sources attached, hours of searching disappear. So does some duplicated research.

It is not magic. It is institutional memory with a search box people might actually use.

Novartis did the unglamorous work

The Novartis example stood out to me because the AI is not the most interesting part.

Its consumer business built a system called Sherlock. People can ask questions and receive answers tied back to a line of text or a timestamp in a video. It has curated topic areas, rules for what gets uploaded, and warnings when a finding only applies to a particular geography or group.

In other words, someone thought about context.

Research suppliers can upload work directly. Sources remain traceable. Old findings are easier to reuse without being presented as universal truths. The MIT Sloan article says the system helped Novartis avoid duplicated work and saved more than $29 million in primary market-research costs in one year.

That number will get the attention. Fair enough. But the saving did not come from asking a chatbot to imagine what customers think. It came from making previous research easier to find and safer to reuse.

The hard work was still there: governance, curation, ownership, and deciding what counted as good enough to enter the system.

A summary can still be wrong

This is the bit I would keep on a Post-it next to any customer-insight AI project.

GenAI can read hundreds of interview transcripts and produce a neat set of themes. It can compare studies, spot repeated topics, and turn a stack of reports into something a busy leader may actually read.

What it cannot tell you, on fluency alone, is whether the original research was any good.

Was the sample sensible? Was the question leading? Who was missing? Did three loud participants distort the conclusion? Is a finding from 2022 still relevant? Did the research measure behaviour, or just what people said they would do?

If those problems are in the source material, the model may carry them into the answer — only now the answer is shorter, cleaner, and easier to believe. It is the same verification bottleneck that appears in AI-assisted coding, moved into customer research.

That is the risk. Not an empty screen. A polished answer built on shaky ground.

A faster route to old knowledge is useful. A faster route to an unsupported conclusion is not.

This is a people problem wearing an AI badge

The MIT Sloan piece makes a useful distinction between storing knowledge and managing how knowledge moves.

Customer insight has a life before it reaches the archive. Someone chooses the question. Someone designs the research. Someone interprets it, adds caveats, decides where it belongs, and — ideally — checks later whether it is still true. Interpretation matters because the story around the facts can change the decision.

AI can help at several points. It cannot take responsibility for the whole chain.

PepsiCo’s experience makes that clear. The company created a Global Insights Council that brought together insight leaders from different regions and central teams. Yes, it needed systems. It also needed people with enough authority and interest to make research travel across organisational boundaries.

That part is easy to skip because buying a tool feels more decisive than fixing ownership.

Without it, we may end up rebuilding the old knowledge-management problem with a nicer chat interface. SharePoint, but it talks back.

What I would ask before buying anything

I would start with the archive, not the model.

What is actually in there? Who owns it? How much is duplicated, outdated, badly labelled, or missing? If nobody can describe the material, adding GenAI will not suddenly make it coherent.

Then I would ask whether context travels with the answer. A finding based on European patients should not quietly become a global consumer truth. Dates, sample, location, method, and limitations matter. They need to stay attached.

I would also insist that important answers link back to the source. Not because every user will open it, but because they must be able to. Traceability is what turns a fluent response into something you can check.

And someone needs the job of maintaining all this. Archives decay. Markets move. Products change. New research arrives, old research lingers, and duplicate material piles up. There is no “finished” version.

Finally, decide where humans still have to do the thinking. New research, methodological review, difficult qualitative interpretation, and high-stakes decisions should not drift into automation just because the interface is convenient.

The 1% takeaway

The best near-term use of generative AI customer insights technology may be less exciting than the sales pitch.

It is not an artificial customer. It is not a machine that understands your market. It is a way for the organisation to remember what it has already paid to learn.

That is valuable. It can stop teams repeating work, get evidence beyond the insight department, and give people a better chance of checking what customers have actually said before a decision is made.

But the answer still depends on everything that happened before the prompt: who was asked, what was measured, how it was interpreted, and whether anyone preserved the context.

GenAI can help a company remember its customers.

Understanding them is still work.

---

Primary source: Thomas H. Davenport and Viktor Dörfler, “How GenAI Can and Can’t Help Manage Customer Insights,” MIT Sloan Management Review, 13 July 2026.