The numbers lined up on the conference room screen: "12 customer complaint categories. #1, shipping delays. #2, inconsistent product descriptions. #3, sizing complaints." This was what an insights team at a consumer goods company got after feeding three years' worth of voice-of-customer (VOC) data — thousands of records — into a RAG system. The team lead built next quarter's product roadmap around that summary. Six months later, a competitor built an entire new category around a single unmet need surfaced from just 15 qualitative interviews. The thousands of records the AI had combed through never contained that signal — because what customers hadn't yet said wasn't sitting in any database to begin with.
That boundary is exactly what MIT Sloan Management Review has been examining lately. A growing number of customer-facing companies are combining generative AI with large language models' reasoning ability to analyze their internal content. The technique at the center of this approach is RAG — Retrieval-Augmented Generation — which chains a retrieval step (pulling relevant documents out of a database) to a generation step (having the language model reason over those documents to answer a question). Customer surveys, interview transcripts, NPS comments, and internal analysis reports all become inputs to the system. What the research zeroed in on was exactly where that structure delivers value, and where it stops.
Where AI Actually Saves Time
RAG-based generative AI clearly earns its keep on problems of scale and speed. It can scan through hundreds of customer texts in a few hours — work that would take a human analyst weeks — and pull out recurring themes. It can also link internal reports scattered across departments, filter out redundant observations, and consolidate common complaint patterns in one place. It helps new team members get up to speed on a mountain of customer context quickly, and it measurably cuts the time it takes to draft a quarterly insights report.
At this stage, what AI is really doing is generating hypotheses. It pulls up whatever looks like a signal from a mass of documents and narrows down what an analyst should pay attention to. Instead of reading every document from scratch, analysts start from a set of AI-pre-screened leads. This isn't a technology that eliminates time spent on reporting — it's one that frees up more time for judgment, earlier in the process.
The same holds for high-volume text that arrives in short bursts, like customer inquiry logs or return-reason comments. When AI does a first pass over volumes too large for anyone to review in full and flags the anomalies, analysts can then divide the work by concentrating on exactly those flagged points.
Where AI's Reach Ends
There's also a gap built into this structure. What customers don't say — nonverbal reactions, facial expressions, hesitation, the brief silence that follows a particular question, the impulse to change the subject — never makes it into a text database. Those are signals a researcher picks up in a live interview, and they're simply not data a RAG system can reach.
The more fundamental weakness is directional. RAG works with whatever internal documents already exist. Needs customers haven't yet articulated, wants for which no vocabulary exists because the product doesn't exist yet, opportunities no competitor has built and that therefore have nothing to be compared against — none of that is sitting in any internal document. A tool built to extract patterns from data that has already accumulated cannot point to a signal that hasn't been spoken yet.
There's also the problem of bias in the input data. If dissatisfied customers' reviews make up most of a company's VOC data, the AI hands back insights skewed toward complaints. If the surveys it scans over-represent a particular age group or channel, that skew carries straight through into the analysis. Human analysts habitually ask, "is this sample representative of the whole?" In a system built for automated processing, that question is easy to skip.
The Skeptic's Case
As this approach spreads quickly, the objections raised in the field deserve an honest look too. Among the concerns researchers and practicing analysts raise most often is the erosion of analytical skill. When a team repeatedly consumes AI summaries instead of practicing reading raw data themselves, their instinct for catching errors atrophies slowly. Some researchers call this "the erosion of audit capacity." An organization needs to retain the ability to re-examine a tool's conclusions in order to catch the tool when it's wrong — and the observed pattern is that the longer a team relies on the tool, the more that capacity quietly declines.
There's also concern that direct customer contact itself shrinks. If RAG-based analysis makes qualitative interviews and field observation feel less necessary, the volume of what customers say may grow even as the depth of what customers actually want gets shallower. This isn't a flaw in the AI tool itself. Using the tool to accelerate analysis is reasonable; using it as a justification for less field contact produces a different outcome entirely.
What Smaller Teams Should Check First
This discussion may sound like it's about large-enterprise insights teams, but the structural trap in understanding customers actually shows up more starkly the smaller a team is. The fewer resources a team has, the stronger the pull to accept AI-generated conclusions without verifying them.
The first thing to check is the provenance of your input data. Keep a record of which channel the material fed into the AI came from, up to what date, and from which customer segment. Without that information, you lose the context needed to interpret whatever analysis the AI hands back. A biased sample produces a biased insight, and once that insight comes out formatted as an automated report, the error reads as fact.
Treating AI summaries as hypotheses, not conclusions, matters just as much. Read the report through the lens of "the AI proposed this as a hypothesis," not "the AI concluded this." Only once you've built in a final step that asks whether that hypothesis holds up in an actual conversation with a customer or a real use scenario does the AI's output become a genuine starting point.
It's also worth keeping qualitative and quantitative tracks operationally distinct. Quantitative text — surveys, reviews, comments — is well suited to AI processing. But asking why a particular customer made a particular choice, or under what circumstances that decision was made, has to start with face-to-face contact or direct observation. Blur the roles of the two tracks and processing speed goes up while the depth of understanding goes blurry.
Business design has long drawn a distinction between what users say they want and what users actually need but haven't yet found the words for. The former can be gathered from surveys and VOC data; the latter comes from observation and reading context. Most of the cases where companies carved out a market on the strength of experience rather than price started from the latter. RAG-based generative AI is a tool that processes the former far faster and far more broadly. The latter still belongs, today, to people doing the legwork themselves.
It's worth questioning the assumption that processing more customer data automatically means understanding customers better. If the pile of AI-generated insights keeps growing while your urge to ask customers something directly keeps shrinking, that's the moment to reconsider direction. Processing speed and depth of understanding don't always move in the same direction.
