Business Situation
The client is a global investment research and market intelligence provider supporting buy-side and sell-side institutions across multiple sectors and geographies, publishing several hundred company profiles, sector notes, and thematic reports a month. As coverage expanded, the quality control process had not scaled with it and analysts were spot-checking citations and cross-referencing data points largely from memory of which sources mattered, which worked fine at lower volume but started slipping as output grew.
By the time a stale data point or an uncited claim got caught, it was often close to publication, which meant late-stage rewrites that ate into the next cycle’s timeline rather than a clean handoff.
SGA Approach
Deploy a layer of QA-focused AI agents ahead of analyst sign-off, so the checks that used to depend on memory and sampling became systematic and traceable, without adding a review step that slowed publication down.
Key Activities
- Built a source verification agent that cross-referenced extracted data points against the client’s approved source list and flagged discrepancies for analyst review, rather than letting mismatches surface downstream.
- Set up a citation and traceability agent that mapped every cited statement, statistic, or conclusion back to its originating source, replacing what had been an inconsistent, analyst-dependent citation habit.
- Built a consistency review agent that checked terminology, structure, and template adherence against the client’s house style guide before a report moved to final review.
- Added a risk-detection agent that flagged unsupported claims, missing references, and contradictory statements; the kind of issue that is easy to miss reading a document end-to-end under deadline.
- Kept a human-in-the-loop checkpoint throughout: agents flagged and surfaced issues, but every flagged item, and every report overall, still needed analyst sign-off before publication.
Execution
- Piloted the framework on the client’s technology-sector coverage for one full quarter before expanding to other sectors, to validate flag accuracy against editorial judgment.
- Retuned the consistency and risk-detection thresholds twice in the first six weeks, after early flag volume ran higher than analysts wanted on lower-risk reports; volume settled once thresholds matched what they considered worth flagging.
- Ran the new and old QA process in parallel for eight weeks before retiring manual spot-checks, to confirm nothing was falling through.
Key Takeaways
- Citation and source verification coverage moved from a 25-30% manual sample to 100% of citations checked on every report.
- Time analysts spent on manual validation and cross-referencing fell by an estimated 45%, freed up for interpretation and analysis.
- Average time from draft to publish-ready shortened by 20%, even with full citation coverage built in. The time saved from avoiding late-stage rewrites more than offset the new automated checks.
- During the parallel-run period, the traceability agent caught a stale valuation multiple carried over from an earlier draft and never updated; the kind of error that is easy to miss and hard to explain to a client after the fact.
