Separate copy or extract
Analysis runs on a prepared copy or on files produced inside the company infrastructure.
← Business control
Data diagnostics and control
Turn system data into a map of drivers and exceptions
The system receives an agreed data set from a database copy or controlled extract, validates its quality, and connects transactions into a broader business picture. The live database remains unchanged, and an expert forms the conclusions.
De-identified demonstration data
Controlled environment
Analysis without interfering with the live database
The data transfer method is agreed with the client. The analysis uses read-only access and makes no changes to the accounting system.
Only agreed data
Only necessary objects, fields, and periods are extracted.
Reproducible evidence
Algorithm versions, input files, and outputs are linked by checksums and can be independently repeated.
Ongoing monitoring can use a dedicated read-only route, but only after agreeing access, the exact data scope, and refresh rules.
How the work is structured
From data to an operating control layer
Data
A controlled copy or extract becomes a passport of scope, quality, and limitations.
Calculation
Reproducible indicators and comparisons form the factual report.
Drivers
Procurement, cost, sales, cash, and operations are connected into explainable chains.
Review
An expert assesses context, materiality, and alternative explanations.
Control map
Priority indicators receive sources, refresh frequency, and response rules.
Monitoring
Useful checks become a repeatable control layer with a history of signals and decisions.
First deliverable
What the data allows us to check now
The report shows both measured facts and analytical limits. A missing source is never replaced with an assumption.
AreaData obtainedCoverageStatus
Sales2,185 rows · 7 organisations100% of extractAvailable
Accounting entries115,317 rowsSelected periodAvailable
Manual operations215 entriesAccounting journalAvailable
User permissionsNo extract supplied—Not assessed
The figures illustrate the report format. Available objects, transaction volumes, and method coverage depend on the configuration and actual data quality.
Analysis examples
From a single metric to an explainable driver chain
Calculations show what changed and which transactions formed the change.
01 / MARGIN
Factor change in margin
Previous periodCurrent period
Revenue excl. taxRUB 10.00m → RUB 10.40m+4.0%
Cost of salesRUB 7.50m → RUB 8.94m+19.2%
Arithmetic marginRUB 2.50m → RUB 1.46m−RUB 1.04m
Margin rate25.0% → 14.0%−11 pp
Drivers of the cost change
Ingredient A · supplier BPrice +36%+RUB 1.12m
Other componentsPrice +2–8%+RUB 0.32m
Production mixNo changeRUB 0.00m
The reasons for changed purchasing terms and their business meaning are established through further analysis.
02 / PRICE
Purchasing comparison across organisations / one item’s price dynamics over time
| Item | Organisation | Supplier | Average price | Volume | Vs minimum |
|---|---|---|---|---|---|
| M16 bolt | Organisation A | Supplier 1 | RUB 101.50 | 18,000 | — |
| M16 bolt | Organisation B | Supplier 1 | RUB 128.00 | 12,500 | +26.1% |
| M16 bolt | Organisation C | Supplier 2 | RUB 119.40 | 8,200 | +17.6% |
A price difference is a fact that prompts review of specification, volume, logistics, payment terms, and contract date.
Data enrichment
1C is an important source, but not the whole business picture
Additional sources help establish whether an accounting entry is supported by independent operational events.
Accounting and cash
1C, banking transactions, payments, and accounting entries.
Sales and customers
CRM, registers, fiscal data, deals, returns, and relationship history.
Inventory and operations
Warehouse movements, balances, stock counts, and process events.
People and external context
Employees, access, event logs, registries, and counterparty data.
Each source is added independently, with a clear purpose and its own quality assessment.
Signal example
A sale is cancelled after a receipt is issued
One signal is built from several independent facts. The system shows which parts of the chain are supported and which still require data or review.
System output
Three facts agree and one source is absent. The event is passed to a human for review; the cause is not inferred automatically.
Controlled use of AI
The LLM helps navigate data but does not control calculations
The analytical core remains deterministic: Python, registered SQL templates, and rules produce the same indicators on every run.
Algorithms calculate
Completeness, amounts, trends, and document links are produced by reproducible code.
The LLM proposes
The model maps fields, groups names, and drafts hypotheses over prepared facts.
A human decides
An expert checks mappings and alternative explanations, determines significance, and forms conclusions.
The LLM has no connection to 1C, SQL, or accounting data: it receives only minimal de-identified context, and an expert checks every proposal. A local model can be used, or the LLM can be disabled entirely.
First step
Start with the data you already have
Describe your questions, and I will propose an approach.