Title: AI Agents in Accounting — What Really Works in Kazakhstan
Text: In February 2026, the American startup Pilot announced the launch of the first fully autonomous AI accountant for small businesses: client onboarding, monthly closing, reporting — all without human intervention. According to Karbon, 98% of accountants worldwide already use AI in their work, with 46% using it daily. However, the same study shows that only 11% of companies have implemented AI agents in a full-fledged agent mode. The technology exists. Implementation — lags behind.
Since 2024, at West Star Ltd, we have been developing an AI-accountant product — a Telegram bot through which Kazakhstani accountants solve everyday tasks: clarifying tax code norms, checking counterparties, generating entries, and analyzing explanations from the State Revenue Committee (SRC). Over this time, thousands of requests from real users have passed through the bot, forming a fairly clear picture of what AI in accounting in Kazakhstan does well today and where it hits a wall. And importantly — why this wall exists and why it won't disappear in the foreseeable future.
This article is an attempt to provide an honest, non-marketing analysis. Six tasks where AI is already working now, four where it doesn't work and won't work in the coming years. Without promises of "AI will replace accountants" and without alarmism that "AI can't do anything."
WHAT AI REALLY COVERS ALREADY
First — consultations on the norms of the Tax Code and SRC explanations. Modern models like Claude Opus 4.7 or GPT-5.4 connected to the Kazakhstani legislation database via RAG answer questions like "what VAT rate applies in such a situation," "how to handle a transitional transaction from 2025 to 2026," "does my OKED fall under the restrictions of the simplified tax regime" — within seconds. According to our observations with the AI-accountant, about 60–70% of such requests are resolved with an answer that does not require clarification from a human expert. This is not magic — it's a properly configured RAG over adilet.zan.kz, kgd.gov.kz, and the Ministry of Finance's explanation database. Previously, such a consultation cost 5,000–15,000 tenge from a tax consultant and took from a day to a week. Now — free in half a minute.
Second — recognition and processing of primary documents. You upload a photo of an invoice, delivery note, or act of completed work into the bot — and receive structured data: the counterparty's TIN/BIN, details, nomenclature, amounts, VAT. The recognition accuracy of modern models reaches 98% on standard documents. This is a level where the economics of automation work: an accountant spends 30 seconds checking instead of 5 minutes entering data. With a volume of 200 documents per month, this is a difference of 15 hours of work. For an average Kazakhstani company with 500–1000 primary documents per month, the savings reach 50–80 working hours. Local services — our AI-accountant, FlowAi, and several others — cover this scenario for 10,000–50,000 tenge per month.
Third — counterparty verification. An AI agent quickly checks through kgd.gov.kz, statgov.kz, and open sources: whether the counterparty is active, whether it is on the risk category list, what its tax regime is (simplified or general), and what types of activities it engages in. With the new 2026 norm — expenses on transactions with simplified tax regime counterparties are not deductible for corporate income tax — this check has become critical. Previously, an accountant would go to the SRC website, search by TIN/BIN, and read the extract. Now — they send the BIN to the bot and receive a ready conclusion with a recommendation in a second.
Fourth — formation of template entries. AI knows the chart of accounts of Kazakhstan, standard operations, and accounting logic well. On a request like "how to record the purchase of a fixed asset worth 1.5 million tenge with VAT from a supplier on the general tax regime," AI provides a correct set of entries with explanations. This is especially valuable for micro and small businesses, where an accountant often handles several companies simultaneously and encounters atypical operations once a week. This is not a replacement for accounting in 1C, but it eliminates the intellectual load of making decisions about entries.
Fifth — preparation of responses to requests from tax authorities and counterparties. A notification of comparative control arrives — AI helps structure the response, select the norms to refer to, and formulate the justification. A requirement from the SRC for explanations on a specific transaction arrives — AI drafts a response, the accountant checks and sends it. According to our data, such requests take an average accountant 2–4 hours each, and with AI — 20–40 minutes. This does not mean that AI writes a perfect response. It means that 80% of the draft work — finding the norm, formulating the structure, main theses — is covered, leaving refinement.
Sixth — monitoring changes in legislation. AI tracks updates to the Tax Code, new SRC explanations, changes in reporting forms, updates to 1C configurations — and sends relevant notifications. Previously, an accountant either paid for a subscription to specialized services (ABT, BuhUchet.kz, and analogs) or risked missing an important change. An AI agent with a subscription to the necessary sources monitors this in the background and filters it according to the client's profile: one thing for small businesses on the simplified tax regime, another for a large company on the general tax regime with imports.
WHAT AI WON'T COVER IN THE FORESEEABLE FUTURE
Here begins the honest part. The marketing of agency AI systems often promises "full accounting automation." In reality, there are four classes of tasks where this doesn't work — not because the technology is immature, but because the task is structurally not reducible to AI automation.
First — making accounting decisions with ambiguous interpretation. When a transaction has several possible qualifications (is it a service or work, is it a financial lease or operational, are these materials or goods, are these investments or current expenses), the correct decision depends on dozens of factors: contract terms, actual circumstances, business intentions of the parties, relationship history. AI can suggest options, explain norms — but the final decision is made by a person because the person is also responsible for it, not to AI, but to the tax authorities and the auditor. This boundary is not technical. It's a boundary of responsibility.
Second — communication with complex counterparties and tax authorities in non-standard situations. When an SRC inspector calls and says "your line 12 in declaration 300.00 doesn't match your supplier's data," the accountant needs to negotiate, find a compromise, explain the context of the specific transaction, refer to previously reached agreements. AI can't do this — not because it's stupid, but because it has neither a history of relationships with a specific inspector nor an understanding of what can be said and what shouldn't be. This is working with people in conditions of information asymmetry. AI is out of place there.
Third — accounting for non-standard operations with complex economic substance. Debt restructuring with partial debt forgiveness, barter transactions between related parties, transfer pricing with justification for controlled transactions, operations with cryptocurrencies under unclear regulation. AI knows the general norms, but a specific operation requires court-quality reasoning: what would the auditor say, how would the tax authorities qualify it, how would the counterparty view it. Here, the cost of error is too high, and the nuances are too subtle. An accountant with 10 years of experience is more valuable here than ever. And AI is not approaching.
Fourth — financial strategy and tax planning. Questions like "how to optimize the tax burden of a group of companies while maintaining reputational risks," "how to structure a transaction with a foreign counterparty considering currency control and double taxation agreements," "what dividend policy to choose considering the long-term plans of the owners" — this is not about the correct application of the norm. It's about weighing interests, business vision, understanding people and their motivations. AI can provide a list of options with pros and cons. The decision is up to the chief accountant, financial director, owner. And this decision is determined not by the correct calculation, but by the correct assessment of the situation.
WHY THIS IS NORMAL
Sometimes such analyses lead to the conclusion "so AI is not yet suitable for accounting." This is the wrong conclusion. The correct conclusion is different: AI perfectly covers the routine part of the work — what takes time but doesn't require thinking. And it doesn't cover the part that is actually the profession of an accountant — judgment, responsibility, relationships.
This means a structural change in the profession, not its disappearance. An accountant who spent 70% of their time entering primary data, parsing details, and searching for norms in directories now spends 20% on it. The remaining 50% is freed up for what has always been valuable in the profession but was pushed to the background due to lack of time: analysis, advice to the owner, process optimization, training junior colleagues, dealing with non-standard situations.
In the long term, this will likely lead to a stratification of the accounting services market. The "do routine cheaply" segment will be fully automated — prices there will drop to the level of a service subscription. The "solve complex tasks and take responsibility" segment — on the contrary, will become more expensive because fewer accountants will be needed, but they will be paid more for their qualifications. Already now, according to our observations in the Kazakhstani market, a chief accountant with real experience in complex operations and tax audits costs 800,000–1,500,000 tenge per month — and this is the lower limit for medium-sized businesses. In two to three years, this figure will increase, and the salary of an ordinary accountant-executor will decrease.
BOUNDARIES AND RISKS — AN HONEST LOOK
The use of AI in accounting in Kazakhstan today has several real limitations that need to be remembered.
The accuracy of primary document recognition is 98%, not 100%. On a hundred documents, that's two errors. On a thousand — twenty. If AI uploads data directly into 1C without verification, these errors will enter the accounting and will only surface during reconciliation or a tax audit. Therefore, the architecture must be with mandatory human-in-the-loop: AI prepares, the accountant confirms, confirmed data goes into accounting.
Hallucinations in responses on norms. Even the most advanced models like Opus 4.7 sometimes confidently refer to articles of the Tax Code that don't exist or provide SRC explanations with made-up dates. RAG reduces the risk but doesn't eliminate it completely. For critical decisions, the source needs to be checked — which means AI should show links to specific documents, not just provide an answer. If the bot doesn't do this — using it as a source of truth is dangerous.
Data confidentiality. When an accountant uploads scans of documents with clients' financial information into an AI bot, this data goes to someone's servers. For large clients and state enterprises, this is a potential violation of commercial and personal secrecy protection requirements. Minimum hygiene: understand which provider processes the data, where it is stored, how protected it is, and whether the provider has a policy of refusing to use the data for training.
Dependence on external APIs. A modern AI-accountant is a bundle of a model (OpenAI, Anthropic, Google), a vector database, integrations with 1C via OData, SRC API. Any of these links can temporarily fail. This is not critical for a one-time consultation, but critical if AI is integrated into the constant workflow of document issuance or period closing.
Localization and Kazakh language. All leading models are stronger in Russian and English than in Kazakh. Documents in Kazakh are recognized worse, Tax Code norms in Kazakh are interpreted less accurately. This is not a blocker for most tasks, but for companies where a significant part of the document flow is in Kazakh, specific scenarios need to be tested, not trusted in advertising.
WHAT TO DO — PRACTICAL CONCLUSION
What makes sense for Kazakhstani businesses and accountants to do following the first wave of AI tools in the profession.
If you are an accountant on the client side or in a company, start with one tool for one task. Don't try to replace the entire workflow. Take an AI consultant on Tax Code norms (our AI-accountant in Telegram, similar local products, or Claude/ChatGPT with the right prompt), use it for two weeks for your requests. You'll understand where it really saves time and where it fails. Then add a second tool — primary document recognition. And so on, one by one.
If you are a business owner, don't buy "full accounting automation." Such a product does not exist in Kazakhstan today in a form ready for combat use without a person. Buy AI tools to enhance your current accountant — this will have a real effect in 1–2 months. The goal should not be staff reduction, but shifting the accountant's work to tasks that were previously not done due to lack of time: expense control, tax burden optimization, analytics for management decisions.
If you are the head of an accounting firm or outsourcing company, invest in AI infrastructure now. In a year or two, competitors who implemented automation earlier will be able to offer the same services for half your current price. Either you learn to do it cheaper — or you lose the price competition. The only alternative: shift the focus towards consulting and complex tasks, where competition is based on qualification, not price.
In our products — AI-accountant and OData Hub — we build the architecture precisely on the principle of human-in-the-loop. AI does the draft work, the accountant controls and makes decisions. And paradoxically, the value of the accountant in this model grows, not falls. Because AI takes away everything for which they didn't enter the profession 20 years ago. What remains is what they entered for — to solve problems, help the business, take responsibility for the result. This is the honest outcome of 2026: AI does not replace the accountant. AI returns accountants to their profession.