Human-in-the-loop

High-touchDialogue over automationValidation-firstNuance preserved

AI handles ingestion, extraction, and structuring. Operators confirm, correct, and annotate — especially where reporting is pattern-shaped but exceptions matter (omitted ARR, restated quarters, issuer-specific definitions).

Schema-first, not generic extraction

Predetermined schemaQuarterly statementsContext engineeringDomain SMEsNo dumb extractor

We target portfolio company quarterly reporting — revenue, ARR, gross margin, retention, churn, headcount — not “any PDF.” A fixed schema unlocks lexicons, tuned prompts, and review UX for portfolio operators. New document classes are added deliberately, not bolted onto a one-size-fits-all parser.

Batch upload: quality over speed

Intensive by designContext compoundsOperator timeSpeed is false economy

Upload is the richest context moment. Batch notes, per-PDF review focus, and confirmed company×quarter mappings pay back in extraction accuracy and shorter review. Thin intake saves minutes upfront and costs hours downstream.

Lexicons & context

Shared vocabularyIntake hintsSource justificationsValidated = truth

Lexicons carry company names, metrics, periods, and review outcomes across batches. Extracted fields ship with quotes and page hints; validated corrections replace AI drafts as durable source of truth.

Automate mechanical, protect judgment

Agents generalizeLess retypingOptional ≠ skipQuality over speed

Automate folder validation, upload orchestration, parsing, and roll-up after validation. Reduce retyping via AI prepopulation — not by bypassing the high-touch upload step when context is unknown.

MDM mindset — AI is only as good as your data

Master data managementClean & mapValidateThorough by defaultSimplify on request

Across many firms, dedicated MDM (master data management) teams exist to clean, map, and validate data before it feeds analytics or AI. Portfolio reporting is no exception — company names, periods, metric definitions, and cross-company comparisons all depend on trusted reference data. What teams ultimately learn: AI is only as good as your data. Zoethales bakes MDM-style discipline into the workflow — schema-first fields, intake context, human validation, and lexicons — rather than treating extraction as a black box. The default design is thorough: confirmation steps, per-PDF context, and review gates are intentional. We can always remove clicks, skip optional steps, or streamline flows on request once operators trust the pipeline — but we start from quality, not speed.

In the product

Upload workflowReview workflowValidated roll-up

Upload — intensive intake before parsing. Review — onboarding is dialogue time with PDF side-by-side. Portfolio — aggregated metrics use validated extraction only.