Technical AI accountancy
How controlled AI bookkeeping actually works: validation, audit trails, confidence scores and human accountability.
Prompt Injection and Financial Data: Documents Are Not Instructions
Why we treat invoices, emails and statements as untrusted input, and how validation plus human approval stops injected instructions reaching your ledger.
4 min read
Why Plain-Text Double-Entry Ledgers Suit AI and Audit
Why a plain-text double-entry ledger is the right substrate for AI bookkeeping: diffable changes, human-readable entries and deterministic validation.
4 min read
What Human-in-the-Loop Actually Means in Bookkeeping
What human-in-the-loop bookkeeping means in practice: AI drafts entries, AgentLedger validates them, and a named person approves every posting.
4 min read
How AgentLedger Validates: Double-Entry Invariants as a Safety Property
How AgentLedger enforces double-entry invariants — balanced postings, known accounts, sane dates — and why it rejects bad drafts instead of repairing them.
4 min read
Evaluating AI Bookkeeping Software: A Buyer's Checklist
A practical checklist for judging AI bookkeeping software: validation layers, human accountability, audit trails, access controls, exports and honest claims.
5 min read
Duplicate Detection by Design: Three Layers and a Human Decision
Inside our three-layer duplicate detection — file hash, draft-level checks and invoice content hash — and why suspected duplicates always wait for a person.
4 min read
Connecting AI Agents to Your Books with MCP and OAuth 2.1
How external AI agents connect to Elizabeth Bookkeeping over MCP with OAuth 2.1 scoped tokens, least-privilege tool scopes and a log of every call made.
4 min read
Confidence Scores in Practice: What They Tell You and What They Don't
What a confidence score on an AI-drafted ledger entry really tells you, why review queues run doubtful-first, and why thresholds are set per account.
4 min read
Automation Policies and Materiality: Deciding What May Skip the Queue
How per-account automation policies combine confidence and materiality thresholds, and why small transaction values do not mean no review by default.
4 min read
Audit Trails for AI Drafts: Who, What, When — and Why
What we record for every AI-drafted entry — source, score, edits, approver, policy — and how that trail answers questions from HMRC or an external auditor.
4 min read