CASE 01 · AI AGENT
TELEGRAM · AI FINANCE AGENT
Change awaiting a decision
Category «Consumables» for 4 transactions totalling ₽18,450
Why: the payment notes and suppliers match earlier payments
AGENT SETTINGS · AUTO-APPLY
Level 85%: people accepted 47 of 50 changes
The change card and auto-apply settings as they look in the service. Sample data.
An agent for the owner and the finance staff of my client NagradyVsem, a company that manufactures awards and trophies. It answers questions about money, categorises transactions, reads receipts and bank statements, sends summaries and keeps an eye on the books. It writes nothing on its own: a person applies every change with one click — until the agent earns more.
THE TASK
The company runs its own management accounting service: one bank comes in via API, others as 1C statement files, and cash is entered by hand. Rules categorise most transactions, but the rest is still done by people:
What was needed: an agent that works in the books every day like one more member of the finance team — and changes nothing without a person.
HOW IT WORKS
In Telegram, in the chat inside the accounting service or right in the Bitrix24 work chat. Text, voice, a photo of a receipt, a statement file or a spreadsheet.
Transactions, reports, rules, the history of similar payments. It answers with numbers linked to transactions, and if something in the books needs fixing it prepares a proposal: what will change, why, and how confident it is.
«Apply» or «Reject» right in the card, in any channel. The service applies the change on behalf of whoever clicked and within their permissions. Once there's a track record, changes the agent usually gets right can be applied automatically.
CAPABILITIES
Income, balances, profit, categories, counterparties, payroll. With charts in the web chat and tables in Telegram. Every number comes from the books, not from the model's memory.
Categories, sales channels and counterparties — up to 200 transactions at once. Each group has its own confidence, and a person can untick individual transactions and apply the rest.
From a photo of a receipt it finds the payment in the books and categorises it. A 1C statement is previewed before import: what's new, what's already there, whether balances match. A cash or expense spreadsheet is imported without duplicates.
An employee writes «1500 petrol for the generator» — 15 seconds later a change with buttons appears in the message thread. An edited message gets a reworked change, a deleted one cancels it.
A checklist in one call: statement completeness, balances, uncategorised transactions, transfers, duplicates and pending changes. Code does the checks; the model explains and suggests fixes.
On weekdays at 9:00 — «Today's work» with prepared changes and questions; on Mondays — the «Money» summary. A watchdog without any AI watches for bank sync failures, large payments and low balances.
Accruals, documents, employees and salaries, cash payouts, importing the accountant's Excel sheet — all as changes a person approves.
Studies a year of the books, forms hypotheses about how the company works and checks them with multiple-choice questions. The profile, procedures and instructions live in the service, with a history for every version.
TRUST
A model can make a mistake or believe a hint slipped into a payment note. If protection rests on instructions alone, it's a matter of time. So the boundary is built into the service itself: even a deceived model can't apply a change — only a person can.
IN NUMBERS
ENGINEERING
Hermes Agent, unforked, in its own container on a separate server. Behaviour comes from instructions, 14 skills and a Python plugin with hooks before and after every model and tool call.
53 tools, 50 visible to the agent. The tools never touch the database directly: they call the same API and pass the same permission checks as the web interface.
A proposal is validated up front and stored as a ready list of steps with a checksum. Before applying, transaction fingerprints are compared: if they changed after the proposal, the change won't apply.
Who asked for a change is filled in by the plugin from the conversation context. The model doesn't see this field and can't forge it.
The agent server is separate from the accounting server. Traffic goes over HTTPS with a pinned certificate and a signature on every request; the port is open only to the accounting server.
Memory, changes, settings, the summary schedule and the log are stored in the service's database. The agent container can be recreated without losing anything.
FINDINGS
While the agent was learning to work with the books, things turned up that used to be noticed only at month-end — or not at all.
SEE ALSO
Answers from the price list, approval by a manager, an action log and failure protection. 77 automated tests.
Every bank account in one picture, 14 reports, an AI assistant and a mobile app.
Service pages with scope and pricing are available in Russian.
CONTACT
I'll explain how an agent would fit your processes: which systems to connect, where to start and what it costs. Message me on Telegram and we'll go through the task.
Message on Telegram dnovoselov@bk.ruI reply within 3 hours on weekdays
A project's budget and commercial terms belong to the client, and I don't publish them. Your project will be treated the same way. My own price ranges are in the Pricing section.