1.5 months → 1.5 minutes — without losing depth.

Speed was the obvious problem. Depth was the actual one.

Every advisory organization in the world was running the same experiment in the same year: pointing a generic large language model at a stack of public-company documents and seeing if useful output came out the other side. It usually didn't, in the sense that the output sounded plausible and didn't quite hold up under scrutiny. Partners who had spent careers building strategic judgment could tell within a paragraph that the synthesis was shallow, even when they couldn't immediately name what was missing.

The actual constraint was not the LLM. It was the absence of a structured context the LLM could reason against. A model asked to summarize an earnings call produces a summary. A model asked to interpret an earnings call against a structured representation of the company's strategic position, its competitive set, and its KPI trajectory produces something closer to analysis. The difference is the mirror underneath.

A structured mirror of the relevant companies and KPIs, with a domain-tuned synthesis layer on top.

• MirrorWe built investment-grade structured representations of the companies in scope — KPIs, strategic drivers, competitive position, regulatory context. Automated ingestion pipelines kept the mirror current from earnings transcripts, filings, and internal analysis. The mirror became the single source of truth that downstream synthesis could reason against.
• CapabilitiesThe relevant constraint was the firm's ability to convert raw inputs into senior-decision-grade analysis at velocity. Generic GPT had improved speed at the cost of depth; the design challenge was preserving depth while collapsing time.
• AnalyzeWe engineered an applied AI synthesis layer specifically calibrated to the firm's strategic frameworks. Every output carried confidence qualification — a feature that turned out to be the deciding factor for partner adoption. Partners trusted output that told them where it was confident and where it wasn't.
• ExecuteQuarterly cycles tuned the synthesis against real partner feedback. Adoption spread organically once a handful of senior partners saw the depth held up. By the end of year one, the system was the firm's standard pre-call preparation engine.

The synthesis layer is now part of the practice.

The cycle-time number is the cleanest single read. The 11% commercial growth was a contributory outcome — the synthesis system supported commercial conversations the firm would not have been able to have at that quality and timing without it, but the growth had multiple drivers and we don't claim sole attribution. The adoption metric matters most for the methodological lineage: this is the engagement where applied AI synthesis became a repeatable practice, not just a one-off build.

Several patterns from this engagement — the structured mirror as substrate for synthesis, confidence qualification on every output, the calibration loop between domain experts and the synthesis layer — became the technical lineage for the Joyce intelligence layer in Rejoyce. We mention this here because it is materially true; we don't mention it on the other case pages because on those it would be marketing.