Finance

What We Build

We apply LLM technology to the structured, high-stakes parts of financial operations where accuracy is non-negotiable: report generation, variance commentary, budget-versus-actual analysis, and risk summaries. Every output is grounded in source data and signed with a full audit trail.

Core Capabilities

Financial Report Generation Automated narrative drafting from ERP or warehouse data — MD&A commentary, variance explanations, and period summaries. Output schemas validated against input figures before any text leaves the pipeline.

Number-Grounding Validation Every figure referenced in a generated narrative is cross-checked against the input payload. Discrepancies trigger structured retries, not silent errors. No inference, no estimation.

Regulatory Guardrails Prompt templates are reviewed against applicable disclosure conventions (IFRS, local GAAP). Forward-looking language is gated behind explicit data authorisation. Version-pinned models prevent style drift across reporting periods.

Variance & Attribution Analysis Structured decomposition of budget-versus-actual gaps, segment-level attribution, and cohort comparisons — delivered as validated JSON or formatted narrative depending on downstream requirements.

Finance-Specific Standards

• All generated figures trace back to a validated source payload — no model-inferred numbers
• Prompt templates are compliance-reviewed and change-controlled
• Model version is pinned in production; updates require explicit re-approval
• Audit log covers input hash, prompt version, model version, and output hash
• Forward-looking statements require explicit authorisation in the input schema

When this work is useful

The right starting point is a reporting workflow with stable source tables, repeated narrative effort, and a review step that can be made explicit. Typical triggers are period-end commentary, recurring variance explanations, or a need to reconstruct how a number reached a paragraph. This is not a replacement for accounting judgement or sign-off; it is a controlled drafting and checking layer around it.

Delivery model and deliverables

First, the source payload is profiled: units, periods, dimensions, null behaviour, and reconciliation rules. We then define a narrow output schema and a vocabulary for permitted explanations. The generation step receives validated data, not an unstructured database dump. A verifier extracts every referenced figure and compares it with the payload before a reviewer sees the draft.

Deliverables can include a source-to-field map, number-grounding rules, versioned templates, a review queue, audit-log specification, test cases for edge conditions, and a handover runbook. The automated reporting use case and the audit-trails insight show adjacent patterns.

Dependencies, risks, and acceptance

Source ownership, accounting definitions, disclosure policy, and a named reviewer are dependencies. Ambiguous mappings, late corrections, and unit conversions remain risks; the safe response is to stop or route to review, not to guess. Acceptance should cover reconciliation, forbidden language, traceability, and reviewer usability on representative periods. Discuss a bounded finance workflow.

Integration notes

The integration should preserve the source system as the authority. A report draft can carry references to the payload and validation results into the existing review process, while the generated text remains clearly identifiable as a draft until approved. This separation supports controlled adoption: the organisation can begin with one report section, prove the checks, and expand only when reviewers understand the failure modes.