How the AI accountant works: document queues, checks, exceptions, accountant control, posting to 1C, and action audit.
The AI accountant handles a work unit, not just chat replies
Input
Preparation
Workstation
Solution
Result
Agent checked
Fields were extracted from the UTD, and the counterparty and order were found in 1C. The document is ready for accountant review.
Suggested action
Create an incoming draft and attach the source UPD
After confirmation: 1C · receipt draft
Trace: email + UTD + 1C order + validation rule version
Parses PDFs, Excel files, and scans, identifies the document type, extracts key details, and prepares a review card.
Matches the document to the order, contract, payment, or counterparty data and highlights specific discrepancies.
Finds the payment, document, or period-close status based on the data and rules available to the employee.
Collects data from approved sources, prepares a draft, and shows deviations before submission.
Tracks the document queue and required actions, and passes tasks to a person with a clear reason.
Reads authorized data and creates drafts via API or MCP without hiding the source or the action performed.
These are real anonymized OSNO-VA screens, not an interface concept. The management P&L, org structure, and payroll give the accountant and manager a familiar working view; the AI agent receives the same entities through MCP and uses them for reconciliations, reports, and variance control.
The model, tools, and interface remain replaceable components
Sources
Agent
Control
Recording
Agent checked
The invoice amount exceeds the available contract limit. Payment draft creation is paused until the responsible person makes a decision.
Suggested action
Return the invoice to the requester or approve the limit overrun
After confirmation: approval queue · awaiting decision
Trace: invoice + contract + limit + rule version + user decision
Instructions, examples, and exceptions are versioned. The power user refines the working behavior, and developers are responsible for architecture and reliability.
Each agent tool has separate read and write permissions; critical actions require role approval.
Changes are tested on standard and exception scenarios before release to the production environment.
The log keeps the source, applied rule, suggested action, user decision, and posting result.
Cases
We choose a repeatable work unit with a clear owner, input, output, and exceptions.
We turn policies and examples into versioned skills, connect data, and grant only the minimum required permissions.
We run standard and exception scenarios, checking the record, log, and user actions.
The next processes plug into the same engineering stack: Git, reviews, CI/CD, observability, and a unified capability catalog.
Next step
Choose a primary document flow, reconciliation, report, or another repeatable process and see the scope, timeline, and first-run result.
