IncQuery tells LDS the AI research failures that catch out experienced professionals

Singhal's explanation starts with scarcity moving rather than disappearing. "AI hasn't reduced the value of good research, it's raised the bar for it," he told Lets Data Science. "When anyone can generate a plausible-sounding answer in seconds, the answers that are actually verified and defensible become worth more, not less."

Two structural points follow. Proprietary data cannot be synthesized: "AI can synthesize public information beautifully, but it can't manufacture your own proprietary field research." And models are backward-looking by construction. They can report what has been written about a market, but not "how a target company's customers feel about renewing next quarter, or how a niche market is shifting right now."

He was explicit that this gap is not a temporary limitation. "That doesn't close as AI improves, it's a structural feature of how these models work."

Underneath sits governance rather than technology. These are capital deployment and deal decisions "where the cost of being wrong can be in the millions, if not billions," which is why investment committees still require independent primary validation. "It's a governance need, not a technology gap."

This is the section worth reading twice. Asked where AI research fails in ways that still surprise professionals, Singhal did not list the obvious errors, because those get caught.

"The failure modes that catch experienced professionals off guard aren't the obvious ones," he said. "An obviously wrong answer gets caught immediately. It's the subtle failures that are dangerous."

He named four:

• •Synthetic respondents. AI can produce something that looks exactly like a real survey response or persona "with nothing real behind it, and at a glance it's very hard to tell the difference"
• •False consensus. Models "tend to smooth toward the statistically likely answer, which can quietly erase the outliers and minority views that are often the most valuable signal in diligence work"
• •No accountability. A fabricated data point "has no professional reputation or stake attached to it, unlike a named respondent or a research firm standing behind its methodology"
• •Uniform confidence. Output "tends to sound equally confident whether it's on solid ground or making something up. That confident tone is exactly what makes experienced people lower their guard"

His summary is the sentence to keep: "What surprises experienced professionals isn't obviously fake output. It's plausible, well-written, confidently delivered output that happens to be wrong."

The false-consensus point deserves particular attention from anyone building analysis tooling. A model that regresses toward the modal answer is not merely averaging, it is deleting precisely the signal that diligence exists to find.

Since respondents prioritized defensibility over speed, we asked what the word means when a deal depends on it.

"Defensible means the research is built to survive pushback," Singhal said, "not just look convincing on first read, but hold up when a client or an investment committee member actively tries to poke holes in it."

In practice that runs from verifying respondents are real qualified professionals rather than fraudulent panelists, through a reproducible methodology "so how do you know this has a real answer instead of a black box," to weighted representative samples rather than "a handful of cherry-picked quotes dressed up as a finding," and triangulation so a conclusion does not rest on one question in isolation. Questions must avoid leading framing, and the data has to be current rather than "a stale dataset being recycled past its shelf life."

IncQuery's answer to synthetic responses is a two-layer quality control system. The first layer removes respondents showing clear fraud signals such as bot-like behavior or privacy-masking tools before they reach the dataset. The second scores everyone who survives that filter using behavioral signals plus survey-specific criteria, marking each High Confidence or Low Confidence.

Two design choices matter more than the mechanics. Nothing is hidden: every respondent gets a visible score and teams can see why anyone was flagged, producing documentation "that holds up when a client or an investment committee asks how we know the data is clean." And it is not fully automated. "It's not trust the algorithm and walk away," Singhal said. Panels independently review every automatic removal to catch edge cases, keeping a human check on the system itself.

He framed the whole effort in one line: "As AI makes it cheaper to fake a response, we've made verifying a real one a rigorous, auditable process rather than a judgment call."

IncQuery uses AI across its own workflow, which makes the boundary he draws more interesting than a blanket position. An AI drafter turns research objectives into a first draft in minutes, though the client always reviews before launch. Other tools add or refine questions, and a scope checker flags misalignment between stated objectives and the survey actually written.

The one he highlighted turns a static survey into a conversation, with AI probing deeper and tailoring follow-up questions in real time. "What's notable is that this doesn't just make things faster, it makes the human response itself richer, and the respondent and their answers are still entirely real."

That leads to his actual test: "AI touches the process, drafting, editing, probing, quality-checking, but never becomes the source of the data. Every response still traces back to a real, verified professional. AI-generated research breaks that chain; AI-assisted research strengthens it."

Singhal's closing answer was aimed squarely at teams building insights tooling inside their own firms, and it cuts against a common product instinct.

Quality, not speed, is the buying signal. Insight quality outranked speed, cost and automation in the survey, so "tooling optimized purely for faster may miss what actually drives adoption. Buyers are paying for confidence in the output, not just velocity."

Verification belongs in the product rather than bolted on afterwards. "If a tool touches decision-critical data, the trust layer is the product."

And the survey surfaced a split worth understanding regardless of what you are building: respondents were far more enthusiastic about AI for analysis and synthesis than about AI conducting interviews. "Buyers aren't rejecting AI, they're drawing a distinction between AI that helps make sense of data and AI that stands in for a human relationship."

His final point is the one that generalizes furthest. "The bar for adoption is trust under scrutiny, not feature count. The tools that win will be the ones designed so their output can survive a skeptical client or an investment committee asking hard questions."

More detail on the research is at IncQuery.