Research operations · Experian · Ongoing
We have a saying on the team: users don't read. We usually mean the people we study, but stakeholders turned out to be users too, and the findings we delivered kept going unreferenced past the readout. Over three systems, we changed how research reaches people. The current one is a knowledge base stakeholders can question in their own words.
How do we make our research findings go further, and let stakeholders work with them directly when they have questions?
Put another way: could a stakeholder ask a question in their own words, the way they'd ask a colleague, and get a thorough, representative answer back, without waiting on a readout that might never get referenced again?
LLM wikis aren't really designed for multiple contributors, or for individual managers signing off on the addition of information. A research knowledge base needs both.
It took three systems to get here, each solving what the last one couldn't.
It's a pilot project I'm using to help build Experian's Applied AI process for future teams, working closely with the Applied AI team.
A lot of the work was defining the data pipeline: where we stop to get researcher sign-off, when a human operator steps in, and when a manager approves new information for release. Researcher sign-off and manager approval are fixed checkpoints. A human operator steps in only when the LLM-as-a-judge can't make a safe determination on its own or needs additional input.
A retrieval-augmented generation (RAG) knowledge base. It covers the full pipeline from the data lake to the warehouse, plus a front end that generates outputs from approved research artifacts.
Research now sits closer to the center of the room when business decisions get made. Not simply available somewhere, but present in the tools and moments where those decisions actually happen.
For stakeholders, the immediate benefit is a chatbot they can use to ask questions of approved research.
For the research team, a collective knowledge base means we can create comprehensive meta-analyses and extensive user journeys, with work from all of our research divisions automatically included.
The time it takes to review all the work done on a topic across teams went from weeks to hours.
The friction itself was the real cost. When accessing research took too much effort, most stakeholders didn't push through it. They simply moved forward and made the decision without an answer.
Given enough friction, stakeholders don't go looking for research. They just move forward without it.
The point was never speed for its own sake. It's making sure friction never gets to decide whether the people we study end up represented in a decision.