AI for accounting: where it helps and where it must not be trusted

An AI model for accounting online is not an "accountant who never makes mistakes". It is an assistant that reads, matches and phrases quickly, but errs confidently in calculations and regulatory details. So the question is not which model is smarter but how to organize the work so that a model's error does not reach reporting. Below: a list of tasks where AI helps, a list of tasks where it must not be trusted, and an example with arithmetic checked by code.

This is not advice on accounting or tax. Rates, deadlines, reporting forms and rules change, so check any rules against official sources and a legal reference system.

AI for accounting: where it helps

Tasks where the source material is at hand and the result is easy to check:

• Parsing primary documents. From the text of an invoice, act or waybill, get details, line items and amounts into a table. Related topic: extracting data from documents↗.
• Letters and replies. A request for missing documents from a counterparty, a reminder to sign an act, a polite reply to a claim. The model writes a draft, you edit.
• Reconciliations. Find differences between your list of acts and the counterparty's list. Row matching is done by code, and the model helps where names are written differently ("Romashka LLC" and "Romashka, limited liability company").
• Explanations. Write an explanatory note for a discrepancy, turn a posting into plain text for a manager, explain to a trainee what a report shows.
• Searching internal policies. Answers to colleagues' questions on the company's accounting policy, if the policy is passed to the model as text. For large policies you will need retrieval by fragments rather than the whole text in every request.
• Working with exports. Normalize expense category names, sort payment comments into categories.

Where AI must not be trusted

• Calculations. Taxes, payroll, depreciation, exchange differences. The model explains a formula decently but can spoil arithmetic on long numbers. Calculate in code or in the accounting system.
• Rates, deadlines, limits and rules. The model may name an outdated value or invent an article. This is the main risk.
• Accounting treatment "from memory". Account correspondence depends on circumstances and accounting policy. The model suggests a typical variant but does not know your situation.
• Reporting forms. Check filling procedures and current formats against official documents.
• Decisions. What to accept for deduction and how to classify a transaction is your decision.

A practical rule: a model can be trusted to read and rephrase what you gave it, but not to recall what you did not give it.

Example: extract invoice data and check the arithmetic

The idea is that the model only reads and code does the counting. If the numbers do not add up, the document goes to a person.

import json
import os
from decimal import Decimal
from openai import OpenAI

client = OpenAI(base_url="https://api.zukko.pro/v1",
                api_key=os.environ["ZUKKO_API_KEY"])

SYSTEM = (
    "Extract data from the invoice text. Return JSON: number, date, seller, buyer, "
    "vat_rate (a number in percent, as stated in the document), "
    "items: [{name, qty, price, total_no_vat, vat, total}], "
    "grand_total. Copy numbers verbatim, do not recalculate or round. "
    "If a value is absent from the text, use null. JSON only."

Note that the code checks the internal consistency of the document (quantity times price, amount plus VAT, sum of lines), not its legitimacy. Whether the rate is right, whether VAT can be accepted for deduction and whether the document complies with the law is not decided by this code. If check returns remarks, the document goes to an accountant with a flag. A mirror check works too: recompute totals in code rather than trusting totals the model returned.

Scans and photos first need recognition or must be sent as an image to a model with image input; some fields in such documents are read with errors, so the share of documents for manual checking is higher for scans.

How to choose the "best" AI for accounting

There is no universal answer, and a ranking from general tests does not replace a check on your documents. A one-day method:

1. Collect 30-50 of your own (anonymized) documents with a manual reference for the fields.
2. Run them through two or three models with the same prompt.
3. Count the share of documents where all fields matched the reference.
4. Compare this with the price per document and choose the cheapest model that gives an acceptable result.

A template for such a test set is described in a simple eval harness in Python↗. The model strength index (the Artificial Analysis Intelligence Index) is a reference point, not a guarantee of accuracy on Russian primary documents.

Data and confidentiality

Accounting documents contain details, amounts and employees' personal data. Before sending, decide what may be passed to an external service, remove the extra from the text (individuals' tax IDs, addresses, passport data) and agree the policy with management. Read the data retention terms of model providers separately: we make no claims about them. Keep the API key in an environment variable and give it a daily dollar limit.

What it costs

Formula: (input × input price + output × output price) / 1,000,000. Assumptions: 300 documents a month, 1,200 input tokens each (document text and instruction) and 250 output tokens. In total 360,000 input and 75,000 output tokens. Zukko prices at October 18, 2026, $ per 1M tokens.

All values are divided by 1,000,000. The amounts are small, so choose the model by the share of correct fields on your set rather than by price, and keep the checking in any case. Current prices: GET https://api.zukko.pro/v1/models.

Connecting

The key is created in the dashboard↗; the request format is described in the chat completions reference↗. The balance is topped up from $1: SBP (the payment page adds 8.5% on top), cryptocurrency or a MasterCard.

FAQ

Can AI keep the books instead of an accountant?

No. It speeds up reading, reconciliation and text preparation. Responsibility for accounting and reporting stays with people.

Which AI is best for accounting?

There is no universal best. Run two or three models on 30-50 of your documents and choose by the share of correct fields and price.

Can I trust the model on tax rates and deadlines?

No. This data changes, and a model may err or name an outdated value. Check against an official source.

How do I protect data?

Pass only the necessary fields, remove personal data, keep the key in environment variables and limit the key.