Researchers paste verbatims, transcripts, and client strategy decks into consumer AI tools every week, and almost none of it goes through IT. Deloitte's 2026 data shows worker AI access up 50%, with governance nowhere near catching up. Here's what shadow AI actually costs a research team, and what closes the gap.
Someone on your research team hit a deadline this month and pasted forty open-ended responses into their personal ChatGPT account to get a fast read on themes before a client call. Somewhere else, an account lead dropped a client's unreleased pricing strategy into a free consumer AI tool to draft a summary faster. Neither shows up on a security dashboard. Neither used sanctioned software. Both are shadow AI, and it's already inside your building. Deloitte's 2026 State of AI in the Enterprise report found worker access to AI tools rose 50% in 2025 alone, while only one in five companies has a governance model mature enough to track how that access actually gets used. Research teams handle exactly the kind of sensitive, unstructured text these tools process best. That's the whole problem in one sentence.
What Shadow AI Actually Looks Like in a Research Team
Shadow AI isn't a rogue engineer standing up an unapproved model in production. It's smaller than that, and much more common. An analyst under deadline pressure opens a browser tab instead of the enterprise tool IT vetted eighteen months ago, and pastes in a batch of verbatims for a quick thematic read. A moderator uploads a full interview transcript to summarize it for tomorrow's debrief deck. A project lead feeds a client's category strategy document into a chatbot to draft talking points for a readout. None of it feels like a security incident in the moment. It feels like getting the job done.
That's exactly why it spreads faster in research than almost anywhere else in a business. Analysts already treat unstructured text, interviews, open-ends, strategy decks, as raw material to move through quickly. The tools that process unstructured text fastest right now are consumer-grade AI products sitting outside any procurement review. Of course researchers reach for them first.
The Numbers Are Bigger Than Most Research Leads Assume
Ninety eight percent of organizations report unsanctioned AI use somewhere inside the business. Separate 2026 research puts the share of US employees using unsanctioned tools at roughly two thirds, with nearly a quarter doing it regularly rather than occasionally. Cyberhaven's analysis of enterprise prompts found sensitive corporate data inside 4% of prompts submitted to tools like ChatGPT, and in more than 20% of the files employees uploaded directly. Verbatims are files. Transcripts are files. Strategy decks are files. Research functions upload precisely the file types that turned up sensitive most often.
Worker access to AI rose 50% in 2025. Only one in five companies has a governance model mature enough to track how that access actually gets used. (Deloitte, 2026)
Why This Is Worse for Research Than for Most Functions
Most departments risk one kind of exposure when shadow AI creeps in. Research risks two, at the same time. Respondent data carries obligations, consent language, and in some markets formal privacy law, that no participant signed up expecting a language model to process their answers. Client data carries a separate risk entirely: an unlaunched product, a pricing test, a positioning shift nobody outside the account team has seen yet. Paste either into a consumer-tier tool and you're trusting that vendor's data handling terms, terms most researchers have never actually read, over the confidentiality commitment your business made in a signed contract.
There's a technical detail that makes this worse than it sounds. Researchers studying model memorization have found it can take a single document for a language model to memorize enough of it to reproduce identifiable fragments later. A transcript treated as disposable input isn't necessarily gone once the summary comes back. That should change how casually anyone treats "I'll just paste it in for a quick read."
What Actually Closes the Gap
Banning AI outright doesn't work. Most research leads already know that. It just pushes the behavior further underground, onto personal devices and personal accounts where nobody can even estimate the exposure. What actually closes the gap is smaller and more specific: a short list of sanctioned, enterprise-tier tools with real data retention and training opt-out guarantees, a policy written for respondent and client data specifically rather than a generic company-wide AI memo nobody in research reads twice, and onboarding that treats this the way confidentiality agreements already get treated. Not a once-a-year compliance email. A normal part of how the team works.
None of this requires slowing down the parts of the job AI genuinely speeds up. It requires knowing, with real confidence, where your team's sensitive data has already gone. So ask it plainly this week: if someone on your team pasted a client's unreleased concept boards into a personal AI account last month, would anyone actually know? See how Quali-Fi keeps AI-assisted research inside a governed environment ->