AI Transparency and Risk Management Notice
Effective and last updated: July 31, 2026.
GlacialBooks uses automation and limited machine-learning services to assist with bookkeeping. It does not replace professional judgment, and customer ledgers and documents are not used to train a general-purpose or foundation AI model.
1. What we mean by AI
GlacialBooks uses the term AI for a combination of deterministic accounting logic, organization-specific rules, statistical scoring, document extraction, anomaly detection, and agent orchestration. Arko sends the signed-in user's request and a compact, permission-scoped organization context to a GlacialBooks Azure OpenAI deployment when model reasoning is available. The model has no database credentials and cannot apply a change. GlacialBooks validates every proposed inspection or action against the current organization, role, supported fields, and accounting controls. Microsoft Azure Document Intelligence processes documents only when extraction is requested. Arko is software and is not a person, licensed CPA, tax professional, attorney, auditor, or fiduciary.
2. System purposes and controls
System
Purpose
Primary controls
Arko
Reads organization-scoped records and workflow state, explains findings, follows lineage, proposes actions, and performs authorized bookkeeping actions.
Permissions, entity grounding, audit history, confidence and close controls, dry runs or confirmation for consequential actions.
Categorization
Uses organization rules, account configuration, prior corrected treatment, and transaction attributes to recommend or apply categories.
Confidence thresholds, exceptions, correction history, immutable posted entries.
Reconciliation
Uses deterministic and scored matching to connect bank activity with ledger records.
Amount and date checks, duplicate protection, confidence thresholds, review queue.
Document extraction
Uses Microsoft Azure Document Intelligence to extract fields from uploaded receipts, bills, and invoices.
Field provenance, confidence scores, review-required state, user correction.
Month-end close
Runs configured checks, categorization, reconciliation, anomaly review, report preparation, and close controls.
Readiness blockers, period controls, run history, reversible adjustments where accounting rules permit.
Anomaly and profitability analysis
Uses financial records and configured business dimensions to identify unusual or decision-relevant patterns.
Traceable source records, user review, no external eligibility decision.
3. Data used and model training
Systems may use transaction descriptions, amounts, dates, account mappings, contacts, invoices, bills, receipts, jobs, payroll summaries, prior corrections, organization settings, and workflow history to produce organization-specific output. Azure OpenAI receives only the context selected for the current request. We do not use Customer Data to train a public, general-purpose, or foundation model. We do not permit a processor to use Customer Data for its own advertising or general model training. Aggregated operational metrics may be used to measure reliability when they do not reveal ledger content or identify a customer. We will provide conspicuous notice and obtain any consent required by law before materially changing these practices.
4. Automated decisions
GlacialBooks automates bookkeeping treatment and workflow decisions within a customer's ledger. It is not designed to make decisions about consumer credit, employment, housing, education, insurance, health care, legal rights, or another eligibility decision that produces legal or similarly significant effects on a person. Customers must not use GlacialBooks output as the sole basis for those decisions.
5. Known limitations
Automation can misunderstand merchant descriptions, business purpose, accounting policy, document text, timing, ownership, or unusual facts. Generated explanations can be incomplete or incorrect, including a confident but false statement sometimes called a hallucination. Confidence scores describe system evidence and do not guarantee correct accounting treatment. Provider delays, duplicate source records, missing statements, incorrect opening balances, and customer configuration can affect every downstream result.
6. Human control and correction
Authorized users can inspect source lineage, review exceptions, correct classifications, change mappings, and use supported reversal or adjustment workflows. Posted journal entries are not silently rewritten. Higher-risk actions may require confirmation, a dry run, an approval, or resolution of a close blocker. Organization owners control membership, and accountant access can be granted to a specific professional without sharing credentials.
7. Risk management program
Our AI risk process is organized around governance, use-case mapping, measurement, and ongoing management, consistent with the voluntary
NIST AI Risk Management Framework. Controls include scoped data access, provenance, test coverage, production monitoring, output review, audit history, incident handling, provider review, and correction paths. This statement does not claim NIST certification or an independent AI audit.
