I’ve argued for some time that the future of AI for day-to-day work is less ultra-powerful models and more about well-constructed technical stacks - made up of vertical tools with user experiences and RAG processes built for very specific tasks.
At this stage, I’m willing to acknowledge I was wrong. It’s both.
In recent months, we’ve been crystallizing the concept of standardized AI pipelines for research tasks. Each pipeline is a series of prompts, models, data sources, vertical tools, and human actions with varying degrees of specificity, chained together for a specific type of research.
Most of our discussions on these topics - and across the ecosystem - have been focused on academic research. But increasingly, at Sourced Economics, we’re directing AI pipelines at areas like customer research, competitive intelligence, policy analysis, and other corporate research tasks.
With the appropriate mix of technologies, we can use AI pipelines to better analyze and solve complex strategic planning issues.
Background on Sourced Economics
Sourced Economics exists to produce formal research for strategic planning. Founding the company in 2022, my impression was that major executive decisions were made too often based on instincts or at least insufficient data. In many cases, we bring the tools and methods of academic researchers to address enterprise challenges.
Companies we’ve worked with include Coca-Cola, Kimberly-Clark, Novelis, and Smithfield Foods. Our projects focus on market entry, pricing & revenue management, supply chain management, policy analysis, macroeconomics, and many more topics.
In recent years, we’ve been confronted by (1) an explosion of new vertical AI tools supporting the research process (one of which we’ve developed ourselves), and (2) the emergence of multi-agent systems for tackling complex issues requiring significant research efforts.
What a pipeline actually looks like
A pipeline is a defined sequence of stages that data flows through from question to deliverable. Each stage has specific tooling, specific inputs and outputs, and rules governing when a human intervenes versus when the process runs autonomously.
I’ve previously detailed a pipeline for acquiring primary research interviewees here. The concept below is the same, applied to secondary and analytical research.
The core components:
Prompts - structured, reusable queries calibrated to the tools and models that receive them. Not one-off questions, but tested instructions that improve over time.
Models - general-purpose LLMs (Claude, ChatGPT) for scoping, synthesis, and analytical work. Different models have different strengths, and the pipeline routes tasks accordingly.
Vertical tools - purpose-built software for specific stages. Undermind for semantic literature search. Moara.io for reference management. Apify for web scraping. Fathom for documenting calls and feedback. These do things general-purpose models can’t.
Agent systems - multi-agent workflows where several models or tools work in parallel, each handling a sub-task, with a coordinator merging results. Most useful when the corpus is large enough that a single model’s context window can’t hold it all, or when parallel search across multiple channels is needed.
Data sources - academic literature, government databases, industry reports, web data, proprietary sources, and client-provided materials, often in the same engagement.
Human decision points - stages where the pipeline pauses for judgment. Scope decisions, source-quality assessments, interpretation of ambiguous results, or client-requested changes.
Output formatting - the same analysis packaged for different audiences. A leadership summary, a detailed appendix, a slide deck, or a working model.
We’re building and refining pipelines across five areas:
1. Literature & evidence review - search, and systematic synthesis of academic and grey literature against a defined set of questions.
2. Market & competitive intelligence - landscape mapping, competitor tracking, and pricing analysis from web sources, filings, and scraped data.
3. Policy & regulatory analysis - tracking regulatory changes across jurisdictions and assessing business impact.
4. Customer & demand research - processing survey data, interview transcripts, and published studies to identify preference patterns.
5. Financial & economic modeling - gathering inputs from multiple source types and feeding them into econometric models and valuation frameworks.
A pipeline in practice
I’ll make this concrete with a recent engagement - a large-scale economic valuation study for an automotive client. The project required synthesizing dozens of academic papers and industry sources to produce quantitative estimates across roughly a dozen variables (specifically, product attributes).
Stage 1: Scope decomposition and search design.
An anonymized statement of work was fed into Claude to generate keyword strings and search queries. This stage required several iterations with human approval - initial queries were over-engineered for search engines that don’t support complex syntax. The pipeline’s search design stage now includes constraints learned from this: keep queries short, calibrate to each tool’s actual capabilities.
Stage 2: Multi-channel discovery.
Parallel searches across keyword-based (Google Scholar) and semantic retrieval tools (Undermind). These use different strategies - a keyword search catches papers whose titles match your terms, while a semantic search catches papers whose content matches a described ideal, even with zero keyword overlap. One of the study’s most important sources surfaced only through semantic search.
All papers were managed through Moara.io - organizing sources, storing summaries, tracking enrichment fields as the corpus evolved over weeks.
Stage 3: Corpus diagnosis.
Before extracting a single data point, the full corpus was loaded into Claude and assessed against the deliverable structure. Rather than “summarize each paper,” instead “what’s missing?”
The result: the initial corpus was skewed. Five of the target variables had zero coverage. Queries were adjusted, searches re-run, and previously empty variables began to populate.
Stage 4: Structured extraction and anomaly detection.
Nearly 800 rows were extracted in Claude, normalized to current-year dollars, and flagged on eight quality dimensions. Anomalies surfaced automatically: sign-flipped coefficients in highly cited papers, broken standard errors, mislabelled units, duplicate entries. Each required a human judgment call.
Stage 5: Cross-domain evidence integration.
Several target variables had thin academic coverage. The pipeline expanded into industry and government data sources - engineering databases, cost studies, technical reports outside the academic literature. Finding these required domain knowledge about non-obvious sources that no search algorithm would surface independently.
Stage 6: Feedback integration.
Client feedback was regularly incorporated throughout the engagement, documented through Fathom and fed back into the pipeline. Priorities shifted, certain variables were elevated, others parked, the analytical tone adjusted. Each round restructured downstream work.
Stage 7: Multi-format output.
The same analysis was produced across multiple formats - internal working documents, client-facing reports with visualizations, presentation decks, and a reference appendix. Each required different tone and detail level, with human review at every stage.
Stage 8: State capture.
The project exceeded the context window of every model used. Handoff documents, workbook snapshots, and structured briefings let each new session pick up cleanly. Any pipeline for real research - multi-week or multi-month engagements - needs to be designed to survive this.
Where this goes
We’re investing heavily in building and refining these pipelines. Every engagement makes them sharper - new constraints get encoded, new tool combinations get tested, extraction logic gets tighter. The goal is a set of research pipelines that can be pointed at the hardest strategic planning questions and produce outputs that are more rigorous, more comprehensive, and dramatically faster than traditional methods.
If you’re working on similar problems - or have a strategic research question you’d like to pressure-test this approach against - I’d love to hear from you: john@sourcedeconomics.com.
-John