George Jensen
Research operations, agentic analysis, and the infrastructure that keeps a conclusion tied to the evidence underneath it.
Qualitative analysis at scale is trusted only when every conclusion can be traced back to what a participant actually said. Agentic systems make that harder, not easier: probabilistic models produce fluent output whether or not the evidence supports it.
So I build the governance around them. Several AI workers each handle one narrow part of the analysis, and none of them sees more than it needs to. Around them sits ordinary software — not AI — that decides what happens in what order, what each worker is allowed to touch, and what gets written down. Every conclusion keeps a record of the evidence it came from, and a person approves the work before it moves forward.
Before research operations, I spent more than twenty years in deep technology — cloud infrastructure, distributed delivery, data platforms, and data science. The throughline since IHETS in 1999, and since my thesis in information science at Indiana University, has been the same question: how an institution holds onto what it knows.
What AI's history suggests about building agentic research systems. Why scaling AI in a research organization is a governance problem rather than a deployment problem, and why the machine is more useful as a sparring partner than an oracle.
Rules for turning interview quotations into verb-forward, participant-centered concept records that stay tied to exact transcript evidence. How the manifest is used →
An evolving ResearchOps literature resolved into reviewed concepts with clear definitions and provenance, published for people and agents in the same source of truth.
Drafts in editorial review. Each will be published here in full.
Two phrases that sound like relatives and are not. Where the human-to-human boundary of an interview has to hold, and where AI can do real work on either side of it — researcher preparation before, transcript analysis after, under a manifest a person wrote.
Regulated enterprises buy site-wide Work AI licenses, then lock the connectors down so hard the searches come back empty. An architecture that separates raw participant custody from governed, de-identified concept nodes — unlocking the investment without breaking a single compliance rule.
Five years of running cluster infrastructure, and then a move into qualitative research that looks like a career break and is not one. The object of care changed; the discipline did not. Where does an interpretation come from, and can you still inspect the lineage when AI participates?