How AI bookkeeping works — humans stay in control
The real pipeline behind AI bookkeeping — ingestion to audit trail — and the points where a human stays in command.
The Bookkeeper Agent is where raw financial activity becomes a clean, posted ledger. It ingests bank transactions and source documents, classifies each line, proposes the journal entry and links it back to the document it came from — so nothing enters the books as an unexplained guess.
Every posting carries a confidence score and an evidence link. High-confidence, routine entries flow straight through; anything ambiguous is held in an exception queue for a person to resolve before it touches the close.
What it does
- Import transactions from bank feeds (ПриватБанк, Монобанк) and a tolerant CSV statement parser.
- Classify each transaction and propose a balanced double-entry journal posting.
- Match payments to invoices and supporting documents; flag duplicates and missing documents.
- Reconcile ledger balances and surface anything that does not tie out.
- Learn recurring patterns as rules so similar transactions classify consistently.
How it works
Ingest
Bank transactions and documents arrive via API or upload — no manual export-and-rekey.
Classify with evidence
Each line gets a proposed account, a confidence score and a link to its source document.
Reconcile
Balances are matched against statements; exceptions route to a review queue rather than being force-posted.
The human boundaryThe Bookkeeper Agent posts routine, high-confidence entries automatically, but low-confidence classifications and the period close itself stop for human review.
Every agent operates under the same controls: confidence scores, evidence links and an immutable audit trail. See how they work together across the platform, or explore the full agent roster.