← All workflows
Compliance

Regulatory filing review

10-K, prospectus, and fund-doc review using a long-context model to extract risk factors, covenants, and material changes.

10-K and fund-doc review on a Box + DealCloud stack: $0.84/run, 62% in embedding regeneration, cross-cycle context, and connector sync. Firedog joins the invoices into one P&L.

Model 35%Prompt 3%Orchestration 4%Retrieval 30%History 21%Connectors 7%
ComponentWhy it is unattributedcalls$ / run
Model · 35%Long-context inference (Gemini 2.5 Pro)Full document context (90k-200k tokens) is sent in one call. Gemini 2.5 Pro is preferred for the 1M-token flat-rate window and strong extraction on dense legal text.tokens in: 192,000 · tokens out: 5,500Visible on the invoice.2$0.29
Prompt · 3%Extraction schema and compliance checklistA 9,500-token prompt encodes the extraction schema (risk categories, covenant types, change-in-control triggers) and is sent twice per run without caching.tokens in: 19,000 · tokens out: 0Visible on the invoice.2$0.024
Orchestration · 4%Pre-processing and post-processing pipeline callsA chunking agent pre-processes the PDF into structured sections; a validation agent checks extraction completeness. Both are billed model calls.tokens in: 14,000 · tokens out: 1,800Pre- and post-processing calls are logged as 'pipeline infra' costs, not attributed to the filing review workflow.4$0.032
Retrieval · 30%Precedent-filing vector lookupPrior filings from the same issuer are retrieved from the vector store to enable change-detection. Embeddings are regenerated on each run rather than cached.tokens in: 24,000 · tokens out: 0Embedding regeneration on every run is a design default in LlamaIndex; the cost appears on the embedding-provider invoice, not in the LLM spend view.8$0.25
History · 21%Cross-document comparison contextSummaries from the prior cycle review are prepended to the current run to enable year-over-year change analysis without storing structured diffs.tokens in: 18,000 · tokens out: 0Cross-cycle context is framed as 'document memory' and handled by the application layer; it never appears as a token cost line in the model provider bill.2$0.18
Connectors · 7%Box, DealCloud, PitchBook, and Power BI connectorsThe workflow retrieves the filing pack from Box, enriches with PitchBook issuer data, writes extracted covenants and risk factors to the DealCloud record, and refreshes the Power BI compliance dashboard. Each external record is normalized by a small LLM mapping call.tokens in: 9,000 · tokens out: 700Box and DealCloud connector traffic bills to the document-platform and deal-CRM contracts; the PitchBook pull sits on the data-vendor invoice. None of it appears in the filing-review LLM spend view.6$0.060
Regulatory filing review costs ~$0.84/run but 62% is re-embedded precedent lookups, cross-cycle context, and Box/DealCloud connector sync hidden across several vendor invoices.
$0.84
cost per run
$13k
per year
62%
unattributed today
$4.20
per user, per day
Scenario: PE fund, ongoing portfolio and deal review. 12 active users, 5 runs/day, 250 working days. 2 documents/run. Avg doc: 95,000 tokens.

Default model: gemini-2-5-pro

Trigger: SEC filing events (10-K, 10-Q, prospectus, S-1) plus ad-hoc on fund subscription documents for new LP commitments.

Who runs it: Legal, compliance, and investment teams at companies, PE firms, and asset managers. Driven by fund doc review for new allocations and ongoing 10-K monitoring for covered names. Filings and fund docs are sourced from Box, enriched with PitchBook, written to DealCloud, and surfaced in Power BI, so each review round-trips through that stack.

What Firedog shows

  • Full per-run attribution split across model, retrieval, and history components with exact token counts.
  • Which retrieval calls are regenerating embeddings for documents already in the index.
  • Dollar impact of switching from per-run re-embedding to cached incremental updates.
  • The per-run cost of pulling filings from Box, enriching from PitchBook, and writing back to DealCloud that never joins the LLM spend view.

Decisions this informs

  • Cache prior-cycle summaries as structured JSON diffs rather than re-injecting full text each run.
  • Switch to incremental embedding updates so only changed filing sections are re-embedded.
  • Enable Gemini context caching for the extraction schema prefix to reduce prompt overhead by ~90%.
  • Sync only changed filing sections from Box and write a single structured update to DealCloud per review instead of per-section round trips.

Connected systems

  • Box · Document storage
    Fund docs, prospectuses, and prior filings live in Box; the connector pulls the review pack and stores extracted risk factors back as secured attachments.
  • DealCloud · Deal / CRM · Intapp (DealCloud)
    Extracted covenants and material changes are written to the relevant allocation or deal record in DealCloud.
  • PitchBook · Market data · Morningstar (PitchBook)
    Issuer profiles and prior-round data are pulled from PitchBook to contextualize filing changes.
  • Power BI · BI / dashboards · Microsoft
    Risk-factor and covenant summaries feed the compliance monitoring dashboards in Power BI.

Understand every AI dollar.

Book a demo to see how Firedog attributes every AI request across your organization, and what your leadership team can do with the data.

Book a demoBecome a design partner