A leading multi-brand B2B media company had spent 20 years building proprietary research for the institutional investment market. Its subscribers increasingly expected to ask a question and get an answer, and competitors were already adding AI to their document libraries.
The publisher saw a larger opportunity than another subscription tier: turn its research into a licensable enterprise data product that subscribers could query directly and enterprise clients could bring into their own systems.
Actable delivered a working AI knowledge base proof of concept in about three months. Following the pilot, the publisher signed a second engagement to extend the architecture to another brand.
The Challenge: Valuable Research, Locked in Silos
The publisher’s institutional investment brand held some of the most valuable data in its portfolio: 20 years of proprietary survey results, a set of databases built from its own reporting, regulatory datasets, and roughly 17,000 pension fund and industry documents.
The value was there. Access was the problem.
Each source lived in its own system. Survey data sat in a survey platform. The reporting databases were fed by spreadsheet uploads. Documents sat in a separate document store, and articles lived in the content management system. Neither subscribers nor reporters could ask a question that spanned them. For a business that wanted to sell access to its data, that was the core constraint.
The workflows behind the data were largely manual. Research staff pulled data points from articles, press releases, and regulatory filings, standardized them in spreadsheets, and uploaded them one file at a time. The process worked, but it limited how much data the team could publish and how quickly subscribers could find what they needed.
The objective was to treat research as a data product, sold through enterprise licensing, and introduce it without putting existing research revenue at risk.
Earlier efforts each addressed a single source. Document search over the PDF library could help users find a report, but it could not answer a question that drew on survey data, reporting, and documents together. General-purpose AI tools tested internally had not outperformed the manual process for extracting structured data. And some third-party data contracts did not clearly permit generative AI use, so rights had to be resolved before any data was ingested.
The publisher did not need another point solution. It needed one governed foundation that could bring every type of content together.
The Solution: A Governed AI Knowledge Base Built for Licensing
As the publisher’s long-term data partner, Actable started a clear direction – to modernize how users interact with the data. Workflow automation was deferred to a later phase so the proof of concept could focus on what subscribers and licensees would actually use.
Actable designed a hybrid AI knowledge base that combines structured and unstructured content. Structured research data sits in governed warehouse tables. Articles and documents sit in a searchable index built for AI retrieval. This gives the publisher a foundation for making its research accessible through AI, without requiring users to navigate multiple disconnected systems.
Actable built the knowledge base in the publisher’s existing cloud environment, with automated ingestion, so the publisher’s own engineering team could build the subscriber-facing experience on top of it.
How It Worked
Step 1: Select a Representative Pilot Brand
The institutional investment brand holds all three content types found across the portfolio: structured research data, documents, and articles. A result there would apply to other brands, and starting with one brand kept integration risk and timelines contained.
Step 2: Map Data Sources and Ownership
Working sessions with the research team documented every source, its owner, the manual steps behind it, and its known quality issues before the build began. That work set the scope of what the proof of concept had to prove.
Step 3: Automate Ingestion
Actable built automated pipelines that refresh the knowledge base daily, so content stays current without manual uploads. Article ingestion started with over 1000 articles, with images and video deferred. The document pipeline processed an initial set of about 1,000 PDFs and was designed to scale to the full library.
Step 4: Benchmark Model Performance
Pension fund documents are long and dense, and processing cost grows with volume. Actable tested models against the publisher’s own documents. Anthropic’s Claude outperformed the alternative on both cost and extraction quality and was selected for document processing.
Step 5: Establish Governance and Hand Over
A licensable product depends on clear data rights and controlled access. Third-party data was reviewed for generative AI rights before ingestion, and one dataset was excluded until those rights were confirmed. Brand and paywall metadata were captured at ingestion so access can follow existing subscription entitlements. Client testers queried the knowledge base, the build was handed to the publisher’s engineering team, and a final demo closed the proof of concept on schedule.
The Results
Actable delivered a working AI knowledge base proof of concept on schedule, about three months from kickoff.
The proof of concept brought structured research data, documents, and articles into a unified AI knowledge base, establishing a foundation for querying information across previously disconnected systems.
The initial build ingested thousands of articles and tens of thousands of PDFs. Automated pipelines refresh the knowledge base daily, and the document pipeline was designed to scale to the full library.
User testing defined the next phase of work. The structured research data needs business definitions before AI can query it reliably, and a small number of core datasets account for most user questions. That gives the publisher a clear place to focus.
The next phase focuses on turning that foundation into commercial value through enterprise data licensing and more interactive subscriber experiences, including searchable, AI-visualized reports in place of static PDFs.
Following the pilot, the publisher signed a second engagement to extend the architecture to another brand, in a different industry, about four months after the proof of concept. The pilot was built to prove a pattern. The publisher is now investing to repeat it.
Turn your proprietary research into a governed, licensable data product. Schedule a call.
